Build an AI Product Design Studio

Wire 25 AI agents with quality scoring to produce client-ready design work.

Introduction

30 Second Summary

On a construction site, nothing gets handed over to the client until every detail passes inspection. That same quality-control principle is what separates amateur creative work from professional output.

In this project, you will build a complete AI-augmented product design studio with built-in quality gates on your Windows PC. 25 specialized agents run through a scoring pipeline where nothing ships to a client below a 7 out of 10.

What You'll Build

A 7-component agentic pipeline that uses Claude Desktop, Paper, Obsidian, and Firecrawl to ingest client briefs, produce scored design artifacts, and compound learning from every approval decision.

By the end of this project, you'll have:

25 agent prompt templates with embedded design principles and an Artifact Approval Scorecard that gates every output on a 1-10 visual rating scale across 6 axes.
3 real Austin client briefs structured and routed through the full pipeline, with the first client's logo approved, produced in Paper, and exported with a value-based proposal.
A Continuous Intelligence Layer that captures every approval and rejection decision so the system improves its first-pass approval rate over time.
Secret Mission: Run all 3 clients through the full scored pipeline, achieve 8+ on every artifact, generate $9,000 in proposals, and publish a portfolio piece with visible quality scores.

What do I need before starting?

This project requires an active Claude Max 20x subscription ($200/mo) and a Paper Pro subscription ($16-20/mo) on a Windows 10 or 11 PC. You also need a company name chosen and 3 clients ready to work with.

Verify Subscriptions and Install Tools

Every tool in this pipeline needs to be live before you build anything on top of it. Two paid subscriptions power the production work. Everything else is free.

In this step, you will confirm both subscriptions are active, install two free tools (Obsidian for knowledge management and Git for version control), and create a Firecrawl account for competitive research.

In this step, get ready to:

Verify Claude Desktop Max 20x and Paper Desktop are active.
Install Obsidian and Git for Windows.
Set up Firecrawl for competitive research.

Verify your subscriptions are active

The two paid tools are Claude Desktop Max 20x ($200/mo) and Paper Desktop Pro ($16-20/mo). Together they cost ~$216-220/mo. This is the entire paid toolchain.

Press the Windows key to open your search bar.
Type Claude and press Enter to open Claude Desktop.
Click the hamburger menu (three horizontal lines) in the top left.
Click Settings.
Click Billing.
Confirm your plan shows Max 20x.

Why Max 20x?

Claude Max 20x gives you 20 times the capacity of Claude Pro per session. This means you can run long, multi-agent conversations without hitting usage limits mid-Council-pass. For a pipeline with 25 agents, that capacity is essential.

Now open Paper Desktop. Press the Windows key, type Paper, and press Enter.
Create or open any file in Paper so the MCP server starts.
Confirm Paper is running by checking for its icon in your Windows taskbar.

What is MCP?

MCP (Model Context Protocol) lets Claude Desktop talk directly to other apps on your machine. Paper's MCP server starts automatically when Paper has a file open. It listens at http://127.0.0.1:29979/mcp on your local machine.

In the next step, you will wire Claude Desktop to this address so Claude can create designs directly inside Paper.

Paper not showing in your taskbar?

Make sure you downloaded Paper Desktop from paper.design (not just the web app). The desktop app is required for the local MCP server to run. After installing, open it and create a new file to trigger the server.

help me verify Paper Desktop is installed and running its MCP server

Install Obsidian and Git for Windows

Obsidian is a free note-taking app that stores everything as plain text files in a folder called a vault. Your vault becomes the central nervous system of the entire pipeline.

Open your browser and go to obsidian.md.
Click the download button for Windows.
Run the installer and follow the default prompts.
Once installed, press the Windows key, type Obsidian, and press Enter to open it.

When Obsidian opens for the first time, it asks you to create or open a vault.

Click Create new vault.
Enter Studio as the vault name.
For the location, browse to C:\Users\your Windows username\Documents.
Click Create.

Why this location?

The vault lives at C:\Users\YourName\Documents\Studio. Claude Desktop will connect to this folder through the filesystem MCP server in the next step. Keeping it in Documents makes the path predictable across Windows machines.

Now enable community plugins so you can extend Obsidian later.

In Obsidian, click the gear icon in the bottom-left corner to open Settings.
In the left sidebar, click Community plugins.
Click Turn on community plugins.
Close Settings.

You do not need any plugins yet

Enabling community plugins now just removes the restriction for later. You will not install any specific plugins in this project.

Next, install Git for Windows. Git tracks every change to your vault so you never lose work.

Open your browser and go to git-scm.com.
Click the download for Windows (the latest version will download automatically).
Run the installer. Accept all the default settings and click through to the end.

What does the Git installer do?

The installer adds the git command to your system PATH. This means you can run Git from any Command Prompt or PowerShell window without specifying where it lives on disk.

Verify Git installed correctly by opening Command Prompt and running this command:

git --version

✔️ I see a version number

Git is installed and ready. You should see output like git version 2.x.x. The exact number does not matter as long as you see a version.

ⓧ Command not found

The installer may not have added Git to your PATH correctly.

Close your Command Prompt window completely.
Open a new Command Prompt window (press the Windows key, type cmd, press Enter).
Try git --version again.
If it still fails, re-run the Git installer from git-scm.com and ensure you leave the "Add Git to PATH" option checked during setup.

help me troubleshoot Git not being found in my Command Prompt on Windows

Now initialize your Studio vault as a Git repository. This lets you track every change to your vault over time.

Navigate to your Studio vault and initialize it by running these commands:

cd C:\Users\[[WINUSER="your Windows username"]]\Documents\Studio
git init

What does git init do?

This command creates a hidden .git folder inside your Studio vault. That folder is where Git stores the history of every file change. You will commit your vault at the end of the project to save a snapshot of the entire system.

You should see a message confirming an empty Git repository was initialized in your Studio folder.

Seeing 'The system cannot find the path specified'?

Double-check that you replaced your Windows username with your actual Windows username in the cd command. You can find your username by looking at the folder name inside C:\Users\ in File Explorer.

help me find my Windows username for the vault path

Set up Firecrawl for competitive research

Firecrawl is a web scraping API that turns any webpage into clean, structured data. You will use it to research competitors for each client.

Open your browser and go to firecrawl.dev.
Click Sign Up and create a free account. No credit card is required.
After signing in, navigate to your API keys page.
Copy your API key and save it somewhere safe (a temporary note, a password manager, or a text file on your Desktop). You will paste it into Claude Desktop's configuration in the next step.

What does the free tier include?

Firecrawl's free tier gives you 1,000 credits per month with no credit card. Each page scrape costs 1 credit. That covers approximately 300 pages of competitive research per month across all your clients.

Can't find your API key?

After signing in, look for an API Keys section in your Firecrawl dashboard. The key is a long string of characters. Copy it immediately after creation, as some providers only show it once.

help me find my Firecrawl API key in the dashboard

Every tool is live. Claude Desktop Max 20x is confirmed. Paper Desktop is running its MCP server. Obsidian has a vault. Git is tracking it. Firecrawl is ready to scrape. Next up, you will wire all three MCP connections together so Claude can talk to your vault, your design tool, and the web in a single conversation.

Wire MCP Connections and Open Parallel Windows

Your tools are installed. Now they need to talk to each other.

Claude Desktop connects to external tools through MCP (Model Context Protocol) servers. Each server gives Claude a specific capability: reading your vault files, creating designs in Paper, or scraping the web with Firecrawl. In this step, you will wire all three connections and open parallel windows so you can run multiple pipelines simultaneously.

In this step, get ready to:

Configure Claude Desktop with 3 MCP server connections.
Test all 3 connections in a single conversation.
Open 3 parallel Claude Desktop windows for simultaneous pipeline work.

Edit the MCP configuration file

Claude Desktop reads a JSON configuration file on startup to know which MCP servers to launch. You will add three server entries: one for your Studio vault filesystem, one for Paper's design surface, and one for Firecrawl's web scraping.

Open Claude Desktop.
Click Settings (gear icon in the bottom-left corner).
Select the Developer tab in the left sidebar.
Click Edit Config. This opens claude_desktop_config.json in your default text editor.

Where is this file?

On Windows, the config lives at %APPDATA%\Claude\claude_desktop_config.json. The Edit Config button opens it automatically so you do not need to navigate there manually.

Replace the entire contents of the file with this configuration:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "C:\\Users\\[[USERNAME=\"your Windows username\"]]\\Documents\\Studio"
      ]
    },
    "paper": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "http://127.0.0.1:29979/mcp"
      ]
    },
    "mcp-server-firecrawl": {
      "command": "npx",
      "args": ["-y", "firecrawl-mcp"],
      "env": {
        "FIRECRAWL_API_KEY": "your-firecrawl-api-key-here"
      }
    }
  }
}

What does this configuration do?

filesystem launches the @modelcontextprotocol/server-filesystem package via npx. The path at the end tells it which directory Claude can read and write. This is your Studio vault.
paper connects to Paper's local MCP server at http://127.0.0.1:29979/mcp. Paper starts this server automatically when the app is open with a file loaded.
mcp-server-firecrawl launches the firecrawl-mcp package. Your API key is passed as an environment variable so Firecrawl can authenticate your requests.

Replace your Windows username with your actual Windows username (the folder name in C:\Users\).
Replace your-firecrawl-api-key-here with the Firecrawl API key you saved in the previous step.
Save the file and close the editor.
Fully quit Claude Desktop (right-click the system tray icon and choose Quit).
Reopen Claude Desktop.

Double-backslashes in JSON

JSON requires backslashes to be escaped. That is why the path uses \\ instead of a single \. If you see connection errors, check that every backslash in your path is doubled.

Config not loading?

Confirm the JSON is valid. A missing comma or bracket breaks the entire file. Paste it into a JSON validator online if unsure.
Make sure you fully quit and reopened Claude Desktop. Closing the window is not the same as quitting from the system tray.
Verify that Node.js is installed and npx is available in your PATH. Open Command Prompt and run npx --version to check.

Why won't my MCP servers connect in Claude Desktop?

Test all 3 MCP connections

A single conversation prompt can exercise all three servers at once. If all three respond, your wiring is complete.

In a new Claude Desktop conversation, type this prompt:

List the contents of my Studio vault. Create a red rectangle in Paper. Scrape the homepage of paper.design and show me the title.

What should I see?

Claude will make three separate tool calls. You should see:

A listing of folders/files in your Studio vault (even if it is mostly empty, you should see the .git folder).
A confirmation that a red rectangle was created on your Paper canvas.
The scraped page title from paper.design.

If all three respond successfully, your MCP connections are live.

Paper server not connecting?

Confirm Paper Desktop is open with a file loaded. The MCP server only starts when Paper has an active document.
Check that no firewall is blocking 127.0.0.1:29979.

Help me troubleshoot Paper's MCP connection

Filesystem server showing an error?

Verify the path in your config matches your actual vault location. Open File Explorer and confirm C:\Users\YourName\Documents\Studio exists.
Check that the backslashes are doubled in the JSON (\\ not \).

Why is my filesystem MCP server failing to connect?

Open parallel Claude Desktop windows

You will run multiple agent pipelines simultaneously. Each Claude Desktop window maintains its own conversation context while sharing the same MCP connections to your vault, Paper, and Firecrawl.

With Claude Desktop open, press Ctrl+N to open a second window.
Press Ctrl+N again to open a third window.
You now have 3 windows. Assign each a mental role:

Window 1 = Client Pipeline (research, brand, production).
Window 2 = Growth Pipeline (identity, business plan, pricing).
Window 3 = Approval and Intelligence (scoring, learning, quality review).

Why 3 parallel windows?

Each window runs an independent conversation. You can launch a Council run in Window 1, switch to Window 2 to build your brand identity, then come back to review results. All three windows read and write to the same Studio vault through the filesystem MCP connection.

Think of it like a construction site. You managed multiple subcontractors working in parallel on different parts of the build. Same coordination skill. Different tools.

Your MCP connections are live and your three-window workspace is ready. Next up, you will build the vault structure, generate 25 agent templates, and create the Artifact Approval Scorecard that gates every piece of work.

Build Vault, 25 Agents, and Approval Scorecard System

Your MCP connections are live. Claude can read and write files, create shapes in Paper, and scrape the web. Now you need to give it a brain.

The vault is the nervous system of your entire pipeline. Every agent template, every client brief, every approval score, and every lesson learned lives in this structure. In this step, you will ask Claude to build the full vault, generate 25 specialized agent prompts with design canon embedded, and create the Artifact Approval System that gates every piece of work before it ships.

In this step, get ready to:

Create the full PARA vault structure with client folders and intelligence layer via filesystem MCP.
Generate 25 specialized agent templates with embedded design principles and approval criteria.
Build the governance layer: Council Lead, Standards Gate, Intelligence Agent, and Artifact Approval Protocol.

Create the vault structure via filesystem MCP

The PARA method organizes information by actionability: Projects (active client work), Areas (ongoing responsibilities), Resources (reference material), and Archive (completed items). Your vault adds two specialized layers on top: a Council folder for the 25 agents and an Intelligence folder for the learning engine.

Switch to Window 1 (Client Pipeline).
Paste the following prompt to ask Claude to create the full vault structure via the filesystem MCP:

Using the filesystem MCP, create the following folder and file structure inside the Studio vault. Create all folders and placeholder .md files listed below.

# Top-level PARA structure
Projects/
Areas/
  Brand/
  Operations/
  Portfolio/
  Intelligence/
Resources/
  Canon/
  Templates/
Archive/

# Client folders (inside Projects/)
Projects/01_casa_verde_taqueria/
  00-Discovery/
  01-Research/
  02-Strategy/
  03-Design/
  04-Production/
  05-Delivery/
  06-Feedback/
  07-Approval-Scores/

Projects/02_graceful_home_care/
  00-Discovery/
  01-Research/
  02-Strategy/
  03-Design/
  04-Production/
  05-Delivery/
  06-Feedback/
  07-Approval-Scores/

Projects/03_eastside_capital/
  00-Discovery/
  01-Research/
  02-Strategy/
  03-Design/
  04-Production/
  05-Delivery/
  06-Feedback/
  07-Approval-Scores/

# Council folder (inside Areas/)
Areas/Council/
  agents/
  council-lead.md
  standards-gate.md
  intelligence-agent.md
  approval-scorecard.md
  artifact-approval-protocol.md

# Intelligence files (inside Areas/Intelligence/)
Areas/Intelligence/
  lessons-learned.md
  design-decisions-log.md
  client-feedback-registry.md
  quality-scores.yaml
  pattern-library.md
  market-intelligence-austin.md
  approval-history.yaml

Create each folder. For .md files, add a YAML frontmatter block with title, created date (today), and status: active. For .yaml files, add a comment header describing their purpose.

What does this create?

Projects/ holds client work. Each client gets numbered subfolders from Discovery through Feedback, plus an 07-Approval-Scores/ folder where every rated artifact lives.
Areas/ holds ongoing operations: your brand, the Council agents, the Intelligence layer, and your portfolio.
Resources/ holds the design canon and reusable templates.
Archive/ holds completed projects you want to reference later.

Wait for Claude to confirm it created all folders and files.
Switch to Obsidian. In the left sidebar, you should see the full folder tree: Projects/, Areas/, Resources/, and Archive/ with all subfolders expanded.

Don't see folders in Obsidian?

Obsidian sometimes needs a moment to detect new files. Click anywhere in the sidebar or press Ctrl+R to reload the vault. If folders still do not appear, confirm Claude's response mentioned the filesystem MCP and did not say the tool was unavailable.

Help me figure out why Claude's filesystem MCP isn't creating folders in my vault.

Generate 25 agent templates and the Approval Scorecard

Each agent in your Council is a specialized prompt template that Claude follows when executing a specific task. The templates embed design canon directly into the instructions so Claude applies professional principles automatically.

Still in Window 1, paste the following prompt to generate all 25 agent templates:

Using the filesystem MCP, generate 25 agent prompt templates as .md files inside Areas/Council/agents/. Organize them by team:

Team A - Brand (5 agents):
  A1-brand-strategist.md - Positioning and competitive differentiation
  A2-logo-designer.md - Logo concepts with Golden Ratio (1.618) proportions
  A3-typographer.md - Typography selection with hierarchy (3 sizes max, 1.5x minimum scale ratio)
  A4-color-specialist.md - Palette creation using 60-30-10 color rule, WCAG 4.5:1 contrast
  A5-brand-voice.md - Tone, messaging, taglines

Team B - Web/Digital (5 agents):
  B1-information-architect.md - Site structure, F-pattern and Z-pattern layouts
  B2-wireframe-designer.md - Layout with Gestalt principles (proximity, similarity, closure)
  B3-ui-component-designer.md - Component design with Material 3 and Apple HIG references
  B4-accessibility-specialist.md - WCAG 2.2 AA compliance, 4.5:1 contrast, 18px minimum for elderly
  B5-responsive-designer.md - Mobile-first with breakpoint strategy

Team C - Research (3 agents):
  C1-competitive-analyst.md - Market landscape, competitor teardowns via Firecrawl
  C2-market-researcher.md - Audience demographics, trends, Austin-specific data
  C3-trend-spotter.md - Design trend tracking, emerging patterns

Team D - Discovery (3 agents):
  D1-client-interviewer.md - Discovery questions, client goals, brand values extraction
  D2-brief-generator.md - Structured brief from interview data
  D3-scope-definer.md - Deliverables, timeline, revision boundaries

Team E - Production (4 agents):
  E1-paper-integrator.md - Paper MCP artboard creation, shape composition
  E2-export-specialist.md - SVG, PNG, PDF export with mockup placement
  E3-mockup-designer.md - Context mockups (signage, business cards, screens)
  E4-quality-reviewer.md - The Squint Test, visual hierarchy check

Team F - Growth (5 agents):
  F1-pricing-strategist.md - Value-based pricing, anchoring psychology
  F2-portfolio-curator.md - Case study structure (Challenge, Insight, Process, Outcome, Impact)
  F3-proposal-writer.md - Persuasion scoring, ROI calculation, 50/50 payment
  F4-acquisition-planner.md - Referrals, cold outreach, LinkedIn, Behance
  F5-retention-manager.md - Follow-up cadence, upsell triggers

For EACH agent file, use this template structure:
---
title: [Agent Name]
team: [A-F]
role_id: [A1-F5]
created: 2026-05-17
status: active
---

# Role
[One paragraph defining what this agent does and when it is called]

# Canon
[Design principles this agent MUST apply. Include specific ratios, rules, and references:
- Golden Ratio (1.618) for proportions where applicable
- Gestalt Principles (proximity, similarity, closure, continuity) for composition
- 60-30-10 color rule (dominant, secondary, accent)
- Typography hierarchy (3 sizes maximum, minimum 1.5x scale ratio)
- White space as intentional design (not empty)
- The Squint Test (hierarchy must survive squinting)
- WCAG 2.2 AA (4.5:1 normal text, 3:1 large text)
Include ONLY the principles relevant to this specific agent's work.]

# Intelligence Query
[What this agent should ask the Intelligence Layer before starting. Example: "Any lessons from previous [logo/web/pitch] projects? Any Austin-specific patterns for [industry]?"]

# Input Schema
[What this agent expects to receive: client brief, research output, previous agent output, etc.]

# Output Format
[Exact structure of what this agent produces]

# Constraints
[Hard limits: timeline, budget awareness, brand guidelines compliance]

# Quality Bar
[Minimum standard for this agent's output to be considered complete]

# Approval Criteria
[The specific 1-10 scoring axes this agent's output will be judged on. Example for A2: Originality, Scalability, Brand Fit, Trademark Safety. Each axis gets a one-line description of what 7+ looks like.]

Generate all 25 files with relevant, specific content for each agent. Do not use generic placeholders.

Why embed canon directly in each agent?

When Claude reads an agent template to execute a task, it applies whatever principles are written in that template. By embedding the Golden Ratio, Gestalt principles, and WCAG standards directly into each agent's Canon section, every output automatically inherits professional design knowledge. The agent does not need to look these up separately.

The Approval Criteria section at the bottom of each template tells the scoring system exactly which axes matter for that agent's output. This creates accountability. The scoring is not generic. It is specific to what each agent produces.

Wait for Claude to finish generating all 25 files. This may take a minute.
In the same Window 1 conversation, paste this follow-up prompt to generate the Approval Scorecard template:

Now create Areas/Council/approval-scorecard.md with the 6-axis rating system used to score ALL artifacts:

---
title: Artifact Approval Scorecard
created: 2026-05-17
status: active
type: scoring-template
---

# Artifact Approval Scorecard

Every Council output and every Paper design gets scored 1-10 on relevant axes before moving forward.

## Universal Scoring Axes (apply to ALL artifacts)

1. **Craft Quality** (1-10): Technical execution. Clean lines, proper spacing, no amateur tells.
2. **Client Alignment** (1-10): Does this solve the client's stated problem? Does it match their brand values?
3. **Market Differentiation** (1-10): Does this stand apart from competitors identified in research?
4. **Canon Compliance** (1-10): Does this follow the embedded design principles (Golden Ratio, Gestalt, 60-30-10, WCAG, typography hierarchy)?
5. **Scalability** (1-10): Will this work at 16px favicon AND on a billboard? Across all media?
6. **Emotional Impact** (1-10): Does this evoke the intended feeling? Does it pass the 3-Second Test?

## Scoring Thresholds

- **Below 7 on ANY axis** = REJECT. Specific feedback to responsible agent. Intelligence Layer captures the failure.
- **7-8 on all axes** = APPROVE with notes. Minor improvements suggested but not blocking.
- **9-10 on any axis** = APPROVE + flag for portfolio. Intelligence Layer captures what made it excellent.

## Aggregate Score

Average of all 6 axes. Aggregate below 7.5 = operator review before Paper production. Aggregate above 8.5 = portfolio candidate.

## How to Score

For each axis, ask: "If a professional designer in Austin charging $86/hr produced this, would they be proud?" Score honestly. The Intelligence Layer only works if scores are truthful.

Write this file via filesystem MCP.

In Obsidian, navigate to Areas/Council/agents/. You should see 25 .md files named A1-brand-strategist.md through F5-retention-manager.md.
Open Areas/Council/approval-scorecard.md and confirm it contains the 6-axis scoring system with thresholds.

Missing some agent files?

If Claude produced fewer than 25 files, it may have hit a response length limit. Ask it to continue: "Please generate the remaining agent files starting from [last file created]." Check the agents/ folder count in Obsidian's sidebar.

Claude stopped generating agent files mid-way. How do I get it to finish?

Build the governance layer and Artifact Approval Protocol

The governance layer sits above the 25 task agents. The Council Lead routes work. The Standards Gate checks compliance. The Intelligence Agent captures lessons. The Artifact Approval Protocol defines the rules that bind them all together.

Switch to Window 2 (Growth Pipeline). You are running this in parallel while Window 1 finishes the agent generation.
Paste the following prompt to create the governance agents and the Approval Protocol:

Using the filesystem MCP, update these 3 existing files and create 1 new file in Areas/Council/:

## 1. Update Areas/Council/council-lead.md

---
title: Council Lead
created: 2026-05-17
status: active
type: governance
---

# Council Lead

The Council Lead receives a client brief and determines which agents run, in what order, and with what dependencies.

## Responsibilities
- Classify brief complexity (Simple / Standard / Complex)
- Route to appropriate teams based on deliverable type
- Set execution order and dependencies between agents
- Monitor aggregate approval scores and escalate to operator if below 7.5

## Routing Logic
- Logo/Brand Identity: D1 → D2 → C1 → C2 → A1 → A2 → A3 → A4 → A5 → Standards Gate
- Website: D1 → D2 → C1 → C2 → A1 → B1 → B2 → B3 → B4 → B5 → Standards Gate
- Pitch Deck: D1 → D2 → C1 → C2 → A1 → A4 → B1 → B2 → Standards Gate
- All routes end at Standards Gate before Paper production

## Scoring This Agent's Output
Route Plan scored on: Efficiency (no redundant steps), Coverage (all needed agents included), Intelligence Alignment (queries relevant lessons)

---

## 2. Update Areas/Council/standards-gate.md

---
title: Standards Gate
created: 2026-05-17
status: active
type: governance
---

# Standards Gate

Final quality checkpoint before any artifact enters Paper production. Reviews the AGGREGATE output of all preceding agents.

## Checks
- Nielsen 10 heuristics compliance (where applicable)
- WCAG 2.2 AA contrast and accessibility
- Brand consistency across all deliverables
- Gestalt principle application in compositions
- Typography hierarchy (max 3 sizes, proper scale)
- Color rule adherence (60-30-10)
- Trademark conflict risk (visual similarity to known marks)
- Golden Ratio application in proportional elements

## Output
A Gate Report with PASS/FAIL per check, specific failure descriptions, and recommended fixes.

## Scoring
Gate Report scored on: Thoroughness (all checks run), Specificity (failures described concretely), Actionability (fixes are implementable)

---

## 3. Update Areas/Council/intelligence-agent.md

---
title: Intelligence Agent
created: 2026-05-17
status: active
type: governance
---

# Intelligence Agent

Captures structured learning from every approval/rejection decision. Maintains the Intelligence Layer files.

## Triggers
- Any artifact scores below 7 (capture: what failed, why, what would fix it)
- Any artifact scores 9+ (capture: what made it excellent, patterns to replicate)
- Client feedback received (capture: alignment between scores and client reaction)
- Standards Gate produces a FAIL (capture: which check, recurring pattern?)

## Output Files Updated
- Areas/Intelligence/lessons-learned.md (new entries appended)
- Areas/Intelligence/design-decisions-log.md (decision + rationale)
- Areas/Intelligence/quality-scores.yaml (running scores)
- Areas/Intelligence/approval-history.yaml (pass/fail record)
- Areas/Intelligence/pattern-library.md (promoted patterns at 3+ occurrences)

## Promotion Rule
Any lesson that appears 3+ times OR improves scores by 2+ points gets promoted into the relevant agent template's Canon section permanently.

---

## 4. Create Areas/Council/artifact-approval-protocol.md

---
title: Artifact Approval Protocol
created: 2026-05-17
status: active
type: protocol
---

# Artifact Approval Protocol

This protocol governs ALL artifacts produced by the pipeline. No exceptions.

## Rule 1: Score Before Pass
Every agent output gets scored 1-10 on its relevant Approval Criteria axes BEFORE passing to the next agent in the chain.

## Rule 2: Visual Quality Rating
Every Paper design gets scored on the 6-axis Approval Scorecard (Craft Quality, Client Alignment, Market Differentiation, Canon Compliance, Scalability, Emotional Impact) BEFORE export.

## Rule 3: Proposal Gating
Every proposal gets scored on: Value Communication, Anchoring Effectiveness, Urgency, Professionalism. Must score 8+ to send to client.

## Rule 4: Rejection Protocol
Score below 7 on ANY critical axis = REJECT.
- Specific feedback returned to responsible agent
- Intelligence Agent captures the failure
- Agent re-runs with feedback incorporated
- Maximum 3 rejection cycles before operator escalation

## Rule 5: Approval with Notes
Score 7-8 on all axes = APPROVE with notes.
- Minor improvements suggested but not blocking
- Notes logged in Intelligence Layer for future reference

## Rule 6: Excellence Capture
Score 9-10 on any axis = APPROVE + portfolio flag.
- Intelligence Agent captures what made it excellent
- Pattern added to pattern-library.md
- Artifact flagged for case study potential

## Rule 7: Aggregate Threshold
If aggregate score (average of all axes) falls below 7.5, the artifact requires operator review before proceeding to Paper production. Above 8.5 = automatic portfolio candidate.

## Rule 8: Client Delivery Threshold
Nothing ships to a client below 7 on all axes. Nothing ships to portfolio below 8 on all axes.

Write all 4 files via filesystem MCP.

Why run this in a separate window?

Window 1 may still be generating the 25 agent files. By switching to Window 2, you run the governance layer creation in parallel. Both windows share the same filesystem MCP connection and write to the same vault. This is the same principle as managing subcontractors on a construction site. Multiple crews work simultaneously on different parts of the same structure.

Wait for Claude in Window 2 to confirm all 4 files are written.
In Obsidian, open Areas/Council/artifact-approval-protocol.md. Confirm it contains all 8 rules with the scoring thresholds: below 7 = REJECT, 7-8 = APPROVE with notes, 9-10 = APPROVE + portfolio flag.
Open Areas/Council/council-lead.md and confirm the routing logic shows agent chains for logo, website, and pitch deck deliverables.

Claude says a file already exists?

The vault structure prompt created placeholder files. If Claude in Window 2 reports the file already exists, tell it: "Overwrite the existing file with the full content above." The filesystem MCP supports writing to existing paths.

Claude won't overwrite my existing vault files. How do I fix this?

✔️ Awesome, I've got everything!

Great. Double check you see 25 agent files in Areas/Council/agents/, plus council-lead.md, standards-gate.md, intelligence-agent.md, approval-scorecard.md, and artifact-approval-protocol.md in the Areas/Council/ folder.

ⓧ I'd like to double check the full structure

Your vault should now contain this structure. Verify in Obsidian's sidebar:

Studio/
├── Projects/
│   ├── 01_casa_verde_taqueria/
│   │   ├── 00-Discovery/
│   │   ├── 01-Research/
│   │   ├── 02-Strategy/
│   │   ├── 03-Design/
│   │   ├── 04-Production/
│   │   ├── 05-Delivery/
│   │   ├── 06-Feedback/
│   │   └── 07-Approval-Scores/
│   ├── 02_graceful_home_care/
│   │   ├── 00-Discovery/ ... 07-Approval-Scores/
│   └── 03_eastside_capital/
│       ├── 00-Discovery/ ... 07-Approval-Scores/
├── Areas/
│   ├── Brand/
│   ├── Operations/
│   ├── Portfolio/
│   ├── Intelligence/
│   │   ├── lessons-learned.md
│   │   ├── design-decisions-log.md
│   │   ├── client-feedback-registry.md
│   │   ├── quality-scores.yaml
│   │   ├── pattern-library.md
│   │   ├── market-intelligence-austin.md
│   │   └── approval-history.yaml
│   └── Council/
│       ├── agents/
│       │   ├── A1-brand-strategist.md
│       │   ├── A2-logo-designer.md
│       │   ├── A3-typographer.md
│       │   ├── A4-color-specialist.md
│       │   ├── A5-brand-voice.md
│       │   ├── B1-information-architect.md
│       │   ├── B2-wireframe-designer.md
│       │   ├── B3-ui-component-designer.md
│       │   ├── B4-accessibility-specialist.md
│       │   ├── B5-responsive-designer.md
│       │   ├── C1-competitive-analyst.md
│       │   ├── C2-market-researcher.md
│       │   ├── C3-trend-spotter.md
│       │   ├── D1-client-interviewer.md
│       │   ├── D2-brief-generator.md
│       │   ├── D3-scope-definer.md
│       │   ├── E1-paper-integrator.md
│       │   ├── E2-export-specialist.md
│       │   ├── E3-mockup-designer.md
│       │   ├── E4-quality-reviewer.md
│       │   ├── F1-pricing-strategist.md
│       │   ├── F2-portfolio-curator.md
│       │   ├── F3-proposal-writer.md
│       │   ├── F4-acquisition-planner.md
│       │   └── F5-retention-manager.md
│       ├── council-lead.md
│       ├── standards-gate.md
│       ├── intelligence-agent.md
│       ├── approval-scorecard.md
│       └── artifact-approval-protocol.md
├── Resources/
│   ├── Canon/
│   └── Templates/
└── Archive/

Your vault now has structure, 25 specialized agents, and a scoring system that gates everything. Next up, you will seed your identity, business plan, and the design canon library that powers every agent's work.

Seed Identity, Business Plan, and Design Canon Library

Your vault is structured, your 25 agents are loaded with approval criteria, and the Scorecard system is ready to gate every artifact.

Now the system needs to know WHO is using it. This step fills the vault with your brand identity, a design canon reference library that every agent reads, business weapons for handling clients, and a pricing structure calibrated to the Austin market.

In this step, get ready to:

Generate your freelance brand identity and the Design Canon Quick-Reference.
Create your business plan and Client Management Weapons pack.
Build a pricing ladder and Proposal Persuasion Template with psychological anchoring.

Generate brand identity and the Design Canon Quick-Reference

Your 25 agents reference design principles every time they produce work. The Design Canon Quick-Reference is the single document that contains all those principles in one place.

In your Window 2 (Growth Pipeline), paste this prompt to generate your brand identity files and the Design Canon:

I'm building my freelance design brand called Construction Paper. Here's my story:

[[YOURSTORY="Your background: construction worker in Austin TX, 5 years building other people's structures, $2,000 in debt, 3 clients waiting, opened Paper to build your own designs"]]

Generate the following files in my vault using the filesystem MCP:

1. Areas/Brand/freelance-brand-identity.yaml - My freelance brand identity including:
   - Brand name and tagline
   - Origin story (construction to design)
   - Core values
   - Voice and tone guidelines
   - Visual style direction

2. Areas/Brand/client-identity.yaml - How I present to clients:
   - Professional name and title
   - Services offered
   - Market positioning
   - Key differentiators
   - Target client profile (Austin small businesses)

3. Resources/Canon/design-canon-quickref.md - Design Canon Quick-Reference containing ALL core principles my agents reference:
   - Golden Ratio (1.618) with proportion examples for logos and layouts
   - Gestalt Principles (proximity, similarity, closure, continuity) with visual descriptions
   - 60-30-10 color rule (dominant, secondary, accent)
   - Typography scale options (1.25x, 1.333x, 1.5x, 1.618x) with use cases
   - White space ratios and intentional negative space
   - F-pattern and Z-pattern for web layouts based on eye-tracking
   - The Squint Test: if you squint and the hierarchy disappears, the design fails
   - WCAG contrast requirements (4.5:1 normal text, 7:1 enhanced)
   - Rule of Thirds for composition

Make the canon reference concise but complete. Every agent in my Council reads this document.

What does this prompt generate?

This single prompt creates three foundational files. The brand identity files tell your agents who you are and how you present to the world.

The Design Canon Quick-Reference is the educational backbone of your entire system. It embeds professional design knowledge directly into the pipeline so your agents apply the Golden Ratio, Gestalt Principles, and WCAG standards without you needing to remember them on every project.

Wait for Claude to finish writing all three files via the filesystem MCP.
Open Obsidian and expand Areas/Brand/ in the sidebar.
Confirm you see freelance-brand-identity.yaml and client-identity.yaml.
Expand Resources/Canon/ and open design-canon-quickref.md.
Verify it contains all 9 principle sections (Golden Ratio through Rule of Thirds).

Files not appearing in Obsidian?

If the files do not show up, Claude may have created them in the wrong path. Check that your filesystem MCP is pointed at C:\Users\YourName\Documents\Studio with double backslashes in the JSON config.

If Claude says it cannot write files, confirm the MCP connection is active by typing "List my vault contents" in the same window. Help me troubleshoot filesystem MCP not writing files in Claude Desktop

Why embed design education in a reference file?

Most AI design workflows give the AI zero context about design principles. The output looks generic because the AI has no quality standard to aim for.

Your Design Canon Quick-Reference gives every agent a shared vocabulary. When Agent A2 creates a logo, it checks Golden Ratio proportions (1.618). When Agent A4 picks colors, it applies the 60-30-10 rule. The canon is the education you do not have to carry in your head.

Create your business plan and Client Management Weapons

Your brand identity tells the system who you are. Now you need a business plan that tells it where you are going, and a weapons pack for the hard conversations every freelancer faces.

In the same Window 2, paste this prompt to start the business plan interview:

Interview me to generate my business plan. Ask me 5-7 questions one at a time about:
- What I want to build (services, clients, revenue target)
- My constraints (time, budget, current debt, location: Austin TX)
- My advantages (skills from construction, story, local connections)
- My timeline (30-day, 60-day, 90-day goals)
- My twelve favorite problems (recurring questions and curiosities about design, business, and creativity)

After the interview, generate:
1. Areas/Operations/business-plan-brief.md
2. Areas/Operations/twelve-favorite-problems.yaml

Ask me the first question now.

What are twelve favorite problems?

This concept comes from Tiago Forte's productivity framework. Your twelve favorite problems are the questions you keep coming back to throughout your career.

Examples: "How do I make something feel premium on a tight budget?" or "How do I explain design value to someone who has never hired a designer?" These questions guide your growth. The Intelligence Layer uses them to filter which lessons matter most to you.

Answer each of Claude's interview questions honestly.
After the interview, wait for Claude to generate business-plan-brief.md and twelve-favorite-problems.yaml in Areas/Operations/.
Confirm both files appear in Obsidian's sidebar under Areas/Operations/.

Now generate the Client Management Weapons pack. These are pre-written frameworks for the hardest conversations in freelancing.

In the same window, paste this prompt:

Generate my Client Management Weapons pack. Create these files using filesystem MCP:

1. Areas/Operations/client-weapons/scope-creep-defense.md
   - Clear scope boundaries template
   - Change order process (new scope = new quote)
   - Professional language for saying no
   - Red flag detection (scope creep warning signs)

2. Areas/Operations/client-weapons/revision-boundary-agreement.md
   - 2 revision rounds included in every project
   - Additional revisions at $150/round
   - Definition of what counts as a revision vs a new request
   - Language to include in proposals and contracts

3. Areas/Operations/client-weapons/difficult-client-responses.md
   - "My cousin knows Canva" reframe: "Your cousin can make something. I make something that WORKS for 5 years."
   - "Can you just make it pop?" translation framework
   - "I'll know it when I see it" redirect to concrete criteria
   - Late payment follow-up sequence (friendly > firm > final)

4. Areas/Operations/client-weapons/value-pricing-conversation-script.md
   - How to anchor high and offer tiered options
   - Handling "that's too expensive" with ROI demonstration
   - Asking the right discovery questions to justify value pricing
   - Never quoting hourly (always project-based)

Why prepare these before you have difficult clients?

Scope creep, revision spirals, and price objections happen to every freelancer. Having pre-written frameworks means you respond professionally in the moment instead of caving or getting defensive.

The revision boundary ($150 per additional round) is not aggressive. It teaches clients to consolidate feedback rather than sending 12 separate rounds of one-off changes.

Wait for Claude to generate all four files.
In Obsidian, expand Areas/Operations/client-weapons/ and confirm you see all four .md files.
Open difficult-client-responses.md and verify it includes the "my cousin knows Canva" reframe.

Claude skipped one of the files?

Long prompts with multiple file outputs can sometimes get partially completed. If a file is missing, ask Claude: "You missed [filename]. Please generate it now."

If the client-weapons/ subfolder was not created, ask Claude to create the directory first, then the files inside it. Help me get Claude to create all the client weapons files

Build your pricing ladder and Proposal Persuasion Template

Pricing is the difference between a hobby and a business. Your pricing ladder positions you at mid-market in Austin, credible enough to win trust but not so high that first-time clients hesitate.

In Window 2, paste this prompt to generate your pricing ladder and proposal template:

Generate my pricing structure and proposal template. Use this Austin 2026 market data as calibration:

- Austin UI/UX freelance median: $86/hr
- Austin graphic design freelance median: $66/hr
- Logo packages in Austin: $450-$1,800
- Website packages in Austin: $2,000-$7,000 (basic informational)
- Pitch deck packages nationally: $1,500-$3,500 (mid-level)

Create:

1. Areas/Operations/pricing-ladder.yaml
   - Tiered pricing for: logos, websites, pitch decks, brand identity packages
   - Position me at mid-market (credible for a new studio with a quality-scored system)
   - Each service has 3 tiers: Budget, Recommended, Premium
   - Include what is included/excluded at each tier

2. Resources/Templates/proposal-template.md - Proposal Persuasion Template using:
   - Premium option listed FIRST (anchoring psychology: the first number sets expectations)
   - Recommended option as the obvious middle (compromise effect)
   - Budget option exists but feels incomplete (loss aversion)
   - ROI calculation section ("This logo will represent your business for 5+ years")
   - "Cost of doing nothing" frame ("Every week without a professional brand costs you...")
   - Payment structure: 50% upfront / 50% on delivery
   - Professional scope statement referencing revision-boundary-agreement.md

Why list Premium first?

This is anchoring psychology. When a client sees $3,500 first, the $1,500 Recommended option feels reasonable by comparison. If you listed $800 first, the $1,500 feels expensive.

The Budget option uses loss aversion. It exists but is deliberately incomplete (fewer revisions, no brand guide, no mockups). Most clients upgrade to Recommended because losing those features feels worse than paying more.

Wait for Claude to generate both files.
In Obsidian, open Areas/Operations/pricing-ladder.yaml and confirm it has 3 tiers for each service type.
Open Resources/Templates/proposal-template.md and verify the Premium option appears before Recommended.
Confirm the template includes the ROI calculation section and 50%/50% payment structure.

Pricing feels too high or too low?

The market data positions you at mid-level. If Claude generated prices that feel unrealistic for your first 3 clients, tell Claude: "Adjust the Recommended tier to [your target] and scale the others proportionally."

Your Artifact Approval System justifies mid-market pricing. You can tell clients: "Every piece goes through 6-axis quality review. Nothing ships below a 7/10." Help me adjust my pricing ladder for Austin clients

✔️ Awesome, I've got everything!

Great. Double check you have saved all files in Obsidian and that nothing is missing from the sidebar.

ⓧ I'd like to double check the full file list

After completing this step, your vault should contain these new files:

Areas/
  Brand/
    freelance-brand-identity.yaml
    client-identity.yaml
  Operations/
    business-plan-brief.md
    twelve-favorite-problems.yaml
    pricing-ladder.yaml
    client-weapons/
      scope-creep-defense.md
      revision-boundary-agreement.md
      difficult-client-responses.md
      value-pricing-conversation-script.md
Resources/
  Canon/
    design-canon-quickref.md
  Templates/
    proposal-template.md

If any file is missing, ask Claude in Window 2: "Please generate [missing filename] now using the filesystem MCP."

Your studio now has an identity, a canon, a business plan, and weapons for every hard conversation. Next up, you will ingest 3 real Austin client briefs and score each one before production starts.

Ingest 3 Austin Client Briefs and Score Each

Your vault is built, your agents are loaded, and the Approval System is ready to gate every piece of work that moves through the pipeline.

Now the system meets reality. You have 3 people in Austin who trusted you with their money before you had a single portfolio piece. In this step, you will structure their needs into scored briefs and catch any gaps BEFORE production begins. The Artifact Approval System does not just gate outputs. It gates inputs too.

In this step, get ready to:

Create and score Client 1 Casa Verde Taqueria's logo brief.
Create and score Client 2 Graceful Home Care and Client 3 Eastside Capital briefs in parallel.
Run Council Lead routing classification on all 3 briefs and score the routing plans.

Create and score the Casa Verde Taqueria brief

Every project starts with a discovery brief. But a brief is only useful if it is complete enough to build from. Scoring the brief before production catches missing information, vague scope, and impossible timelines.

Switch to Window 1 (your Client Pipeline window).
Send the following prompt to Claude. This runs the D1 (Discovery Interview) and D2 (Brief Generation) agents on your first client:

# Client Brief: Casa Verde Taqueria

Use the filesystem MCP to create a client brief at:
Projects/01_casa_verde_taqueria/00-Discovery/brief.md

## Client Details
- Business: Casa Verde Taqueria
- Location: East Riverside, Austin TX
- Context: Food truck going brick-and-mortar. Opening first physical location.
- Owner personality: Proud, family-oriented, wants to honor grandmother's recipes while looking modern.

## Project Scope
- Deliverable: Logo design (primary mark + alternate + icon)
- Budget range: ~$1,500 (reference pricing-ladder.yaml)
- Timeline: 3 weeks
- Use cases: Signage, menus, social media, merch (t-shirts, cups)

## Discovery Questions Answered
- Target audience: East Austin locals, 25-45, values authenticity and quality
- Competitors: Veracruz All Natural, Pueblo Viejo, Suerte
- Must-haves: Green color family, modern but warm, no clip-art cactus/sombrero stereotypes
- Avoid: Generic Tex-Mex visual language, overly corporate feel
- Inspiration: Clean lines of modern Mexican restaurants in CDMX, hand-painted murals

Generate a complete brief following the Discovery agent template structure.
Include all fields from the D1 and D2 agent schemas.

What does this prompt do?

This prompt feeds Claude everything it needs to run the Discovery agents (D1 and D2) and produce a structured brief. The brief captures scope, budget, timeline, audience, visual direction, and constraints.

Claude uses the filesystem MCP to write the brief directly to 01_casa_verde_taqueria/00-Discovery/brief.md in your vault.

Wait for Claude to generate the brief and confirm it was saved via filesystem MCP.
Open Obsidian and navigate to Projects/01_casa_verde_taqueria/00-Discovery/brief.md to confirm the file exists and contains a complete brief.

Now score the brief itself. The Approval System does not wait until a design is finished to catch problems. It catches incomplete inputs before you spend hours on production.

In the same Window 1 conversation, send this scoring prompt:

# Score This Brief: Artifact Approval

Read the brief you just saved at:
Projects/01_casa_verde_taqueria/00-Discovery/brief.md

Score it on these 3 axes (1-10 each):

1. **Completeness** - Are all fields filled? Target audience, competitors, deliverables, budget, timeline, visual direction, constraints?
2. **Clarity** - Is the scope unambiguous? Could another designer pick this up and know exactly what to deliver?
3. **Feasibility** - Is the deliverable realistic within the stated timeline and budget?

## Rules
- If ANY axis scores below 7: generate specific follow-up questions I should ask the client.
- If all axes score 7+: mark as APPROVED for production.

## Output
Save the scores as YAML to:
Projects/01_casa_verde_taqueria/07-Approval-Scores/brief-score.yaml

Format:
```yaml
artifact: brief
client: casa_verde_taqueria
scored_at: [current date]
axes:
  completeness: [score]
  clarity: [score]
  feasibility: [score]
status: APPROVED | NEEDS_FOLLOWUP
notes: [any observations]
follow_up_questions: [if needed]
```

Why score the brief before production?

Starting production on a vague brief is how scope creep happens. If Completeness is low, you are missing information that will force expensive revisions later. If Clarity is low, you and the client have different expectations. If Feasibility is low, you are about to overcommit.

The Intelligence Layer captures these scores. Over time, you learn which types of briefs need more discovery upfront.

In Obsidian, navigate to Projects/01_casa_verde_taqueria/07-Approval-Scores/brief-score.yaml and confirm the scores are saved.
Verify all 3 axes show 7 or above. If any axis scored below 7, Claude will have generated follow-up questions. Note them for your next client conversation.

Brief scored below 7 on an axis?

This is the system working correctly. Review the follow-up questions Claude generated. If you have the answers (because you know the client), reply in the same conversation with the missing information and ask Claude to regenerate the brief and re-score. If you genuinely need to ask the client, note the questions and move on. You can return to this brief after the call.

If the filesystem MCP failed to save the file, confirm your vault path is correct and that Obsidian is not locking the file. Help me troubleshoot my brief scoring in Claude Desktop

Create and score Client 2 and Client 3 briefs in parallel

While Client 1's brief is scored and ready, you will use Window 2 to generate the remaining two briefs simultaneously. This is the parallel architecture from Step 2 in action.

Switch to Window 2 (your Growth Pipeline window).
Send the following prompt to generate both briefs and score them in one pass:

# Client Briefs: Graceful Home Care + Eastside Capital

Generate 2 client briefs using the Discovery agent templates. Save each via filesystem MCP.

---

## Client 2: Graceful Home Care
Save to: Projects/02_graceful_home_care/00-Discovery/brief.md

- Business: Graceful Home Care
- Location: Austin TX (serving Travis County)
- Context: New home care startup targeting Austin's 91,000+ residents aged 65+. Owner is a former RN who left hospital work to provide personalized in-home care.
- Deliverable: Website (5-7 pages: Home, Services, About, Testimonials, Contact, Careers, Resources)
- Budget range: ~$4,000 (reference pricing-ladder.yaml)
- Timeline: 4 weeks
- Target audience: Adult children (35-55) researching care for aging parents. Secondary: seniors themselves.
- Competitors: Home Instead (Austin), Comfort Keepers, BrightStar Care
- Must-haves: Warm but professional, large text (elderly visitors), clear CTAs for scheduling consultations, WCAG AA compliance minimum, trust signals (certifications, testimonials)
- Avoid: Clinical/hospital feel, stock photos of generic smiling nurses, tiny text
- Inspiration: Modern healthcare startups (One Medical, Devoted Health) but warmer

---

## Client 3: Eastside Capital
Save to: Projects/03_eastside_capital/00-Discovery/brief.md

- Business: Eastside Capital (solo real estate investor)
- Location: East Austin / Mueller area
- Context: Solo RE investor developing a fourplex near Mueller. Raising $400K from 3-5 passive investors. Pitch deck is the gating artifact for the raise.
- Deliverable: Investor pitch deck (12-15 slides)
- Budget range: ~$3,500 (reference pricing-ladder.yaml)
- Timeline: 2 weeks (URGENT - investor meeting scheduled)
- Target audience: High-net-worth individuals in Austin ($500K+ liquid), 40-60, conservative risk tolerance
- Competitors: Other Austin RE syndicators (Wildhorn Capital, Sumrok Capital)
- Must-haves: Professional data visualization, clear financial projections, property photos/renders, market data (Austin growth stats), team credentials, risk mitigation section
- Avoid: Hype language, unverified financial projections, cluttered slides
- Inspiration: Clean institutional pitch decks (a]0, Sequoia style) but adapted for real estate

---

## Score Both Briefs

After generating each brief, score them on:
1. **Completeness** (all fields filled?)
2. **Clarity** (scope unambiguous?)
3. **Feasibility** (deliverable in timeline and budget?)

Flag Client 3's 2-week deadline on the Feasibility axis. Is a 12-15 slide pitch deck with custom data visualization feasible in 2 weeks with the Council pipeline?

Save scores to:
- Projects/02_graceful_home_care/07-Approval-Scores/brief-score.yaml
- Projects/03_eastside_capital/07-Approval-Scores/brief-score.yaml

Use same YAML format as the Casa Verde scoring.

Why flag Client 3's deadline?

The Feasibility axis exists to catch overcommitment. A 12-15 slide pitch deck with custom data visualization in 2 weeks is tight. The Approval System forces you to acknowledge this BEFORE you promise delivery. Maybe it scores a 6 on Feasibility and Claude suggests reducing to 10 slides, or front-loading the financial slides that matter most for the investor meeting.

This is the same instinct as checking load-bearing walls before you start framing. Know the constraints. Plan around them.

Wait for Claude to generate both briefs and both score files.
In Obsidian, verify these 4 files now exist:

02_graceful_home_care/00-Discovery/brief.md
02_graceful_home_care/07-Approval-Scores/brief-score.yaml
03_eastside_capital/00-Discovery/brief.md
03_eastside_capital/07-Approval-Scores/brief-score.yaml

Open the Eastside Capital score file. Confirm that the Feasibility axis includes a note about the 2-week deadline.

Claude did not flag the 2-week deadline?

If the Feasibility score came back 7+ with no mention of the tight timeline, reply in the same conversation: "The 2-week deadline for a 12-15 slide pitch deck with custom data viz. Re-score Feasibility with this constraint in mind. Is this realistic with the Council pipeline?" The system should acknowledge the risk even if it still approves.

Help me get Claude to properly flag timeline risks in brief scoring

Run Council Lead routing on all 3 briefs

With all 3 briefs scored and approved, the Council Lead now classifies each project and determines which agent teams need to run. Scoring the routing plan ensures the Council is not wasting steps on irrelevant agents or missing critical checks.

Switch back to Window 1.
Send this prompt to run the Council Lead agent across all 3 briefs:

# Council Lead: Route All 3 Client Briefs

Read the Council Lead template at: Council/council-lead.md
Read all 3 approved briefs:
- Projects/01_casa_verde_taqueria/00-Discovery/brief.md
- Projects/02_graceful_home_care/00-Discovery/brief.md
- Projects/03_eastside_capital/00-Discovery/brief.md

For each client, produce a routing plan:
1. Which agent teams (A-F) are needed?
2. In what order should they run?
3. Which agents can be skipped for this project type?
4. Any special considerations (accessibility for Client 2, data integrity for Client 3)?

## Score Each Routing Plan

After generating routing plans, score each on:
1. **Efficiency** - Does the routing avoid unnecessary agents? Are skipped agents justified?
2. **Coverage** - Does the routing include all critical checks for this project type?
3. **Intelligence Alignment** - Does the routing incorporate any existing Intelligence Layer data?

All scores 1-10. If any axis <7, revise the routing plan.

## Save Output

Save all 3 routing plans + scores to:
Council/routing-scores.yaml

Format:
```yaml
routing_scored_at: [current date]
clients:
  casa_verde_taqueria:
    project_type: logo
    teams_routed: [list]
    teams_skipped: [list with justification]
    special_notes: [any]
    scores:
      efficiency: [score]
      coverage: [score]
      intelligence_alignment: [score]
    status: APPROVED | REVISE
  graceful_home_care:
    project_type: website
    teams_routed: [list]
    teams_skipped: [list with justification]
    special_notes: [accessibility priority, WCAG AA minimum]
    scores:
      efficiency: [score]
      coverage: [score]
      intelligence_alignment: [score]
    status: APPROVED | REVISE
  eastside_capital:
    project_type: pitch_deck
    teams_routed: [list]
    teams_skipped: [list with justification]
    special_notes: [data integrity critical, 2-week constraint]
    scores:
      efficiency: [score]
      coverage: [score]
      intelligence_alignment: [score]
    status: APPROVED | REVISE
```

What does routing classification do?

Not every client needs every agent. A logo project does not need the B4 Accessibility agent (that is critical for the website). A pitch deck does not need the A2 Logo Concept agent. The Council Lead reads each brief and routes only the relevant teams, saving time and keeping outputs focused.

Scoring the routing plan catches two problems: under-coverage (missing a critical check) and inefficiency (running agents that add no value for this project type).

In Obsidian, open Council/routing-scores.yaml and verify all 3 clients have routing plans with scores.
Confirm that Client 2 (Graceful Home Care) routes through the B4 Accessibility agent with a note about WCAG AA compliance.
Confirm that Client 3 (Eastside Capital) has a special note about data integrity on financial slides.

Routing plan scored below 7?

If Efficiency scored low, Claude may have routed unnecessary agents. Reply: "Why is [agent] needed for a [project type]? Justify or remove." If Coverage scored low, a critical check is missing. Ask Claude which canon principle or quality axis is unguarded by the current routing.

Help me understand why my routing plan scored low

✔️ Awesome, I've got everything!

Great. Confirm you have saved all files in Obsidian and your vault shows the complete brief and scoring structure.

ⓧ I'd like to double check the full file list

At this point, your vault should contain these new files from this step:

01_casa_verde_taqueria/00-Discovery/brief.md
01_casa_verde_taqueria/07-Approval-Scores/brief-score.yaml
02_graceful_home_care/00-Discovery/brief.md
02_graceful_home_care/07-Approval-Scores/brief-score.yaml
03_eastside_capital/00-Discovery/brief.md
03_eastside_capital/07-Approval-Scores/brief-score.yaml
Council/routing-scores.yaml

All brief scores should show values on Completeness, Clarity, and Feasibility. Routing scores should show values on Efficiency, Coverage, and Intelligence Alignment for each client.

Three briefs scored. Three routing plans approved. The Approval System already caught the tight deadline on Client 3 and ensured accessibility is prioritized for Client 2. Next up, you will run the first full Council pass on Casa Verde and score every single output the agents produce.

Run First Council Pass and Approve Every Output

Your briefs are scored and your routing plans are locked in. Now it is time to run the first production pass through the Council on Client 1.

Casa Verde Taqueria is your proving ground. In this step, you will push their logo brief through Team C (Research) and Team A (Brand), scoring every single output against the Artifact Approval Protocol before it moves forward. Nothing passes without a 7.

In this step, get ready to:

Run competitive research via Firecrawl and score the matrix on Market Coverage, Insight Quality, and Data Freshness.
Run all 5 Brand agents and score each output against its approval criteria.
Run Standards Gate and produce the aggregate scorecard for the full Council pass.

Run Team C (Research) and score the competitive matrix

Before the Brand team can produce anything meaningful, they need competitive intelligence. Team C uses Firecrawl to scrape competitor websites and produce a competitive matrix that maps the visual landscape Casa Verde will enter.

In Window 1 (Client Pipeline), paste the following prompt to run Team C's research agents on Casa Verde:

You are running Team C (Research) for Casa Verde Taqueria.

Read the client brief at Projects/01_casa_verde_taqueria/00-Discovery/brief.md
Read your agent templates at Council/agents/ (C1 Market Researcher + C2 Competitor Analyst)
Read the Design Canon at Resources/Canon/design-canon-quickref.md

Execute:
1. C1: Use Firecrawl to scrape 5-8 taqueria/Mexican restaurant websites in Austin. Focus on their logos, brand colors, typography choices, and overall visual identity. Save raw research to Projects/01_casa_verde_taqueria/02-Research/

2. C2: Analyze the scraped data into a competitive matrix. For each competitor, document: business name, logo style (wordmark/icon/combo), primary colors, typography, visual strengths, visual weaknesses, differentiation opportunity for Casa Verde.

Save the final competitive matrix to: Projects/01_casa_verde_taqueria/02-Research/competitive-matrix.md

After completing the matrix, SCORE your own output on these 3 axes (1-10 each):
- Market Coverage: Did you capture enough competitors to identify real patterns?
- Insight Quality: Are the differentiation opportunities specific and actionable?
- Data Freshness: Is the scraped data current and from live websites?

If any axis scores below 7, explain what is missing and use Firecrawl to scrape additional competitors until the gap is filled. Then re-score.

Save scores to: Projects/01_casa_verde_taqueria/07-Approval-Scores/research-scores.yaml

What does this prompt do?

This prompt activates two research agents in sequence. C1 scrapes real Austin taqueria websites using Firecrawl's MCP connection. C2 transforms that raw data into a structured competitive matrix.

The self-scoring requirement at the end forces the agents to evaluate their own work against specific criteria. If the matrix only covers 3 competitors, Market Coverage will score low and trigger additional scrapes automatically.

Wait for Claude to complete the research pass. This may take 1-2 minutes as Firecrawl scrapes multiple pages.
Check that Claude reports scores of 7 or higher on all 3 axes.
Verify the file exists by asking Claude: "List the contents of Projects/01_casa_verde_taqueria/02-Research/"

You should see competitive-matrix.md in the folder listing. If Claude reports any axis below 7, it will automatically scrape additional competitors and re-score. Let it finish this loop before moving on.

Firecrawl not returning results?

If Claude says it cannot reach Firecrawl or gets empty responses, confirm your Firecrawl API key is valid and you have credits remaining. The free tier provides 1,000 credits per month at 1 credit per page.

Also confirm Paper Desktop is not blocking the connection. Close any Paper files temporarily if Firecrawl times out, then retry.

help me troubleshoot why Firecrawl MCP is not returning scrape results in Claude Desktop

Why score research before moving to brand?

Every agent downstream depends on the quality of research inputs. If Team A builds a brand identity based on a matrix that only captured 2 competitors, the differentiation strategy is built on incomplete data. Scoring the matrix first prevents cascading quality failures across the entire Council pass.

The Intelligence Layer captures how many competitors were needed before the matrix became useful. Over time, this tells you the minimum research depth for each client type.

Run Team A (Brand) and score each agent output

With approved research in hand, Team A can produce brand artifacts. Five agents run in sequence, each building on the previous output. The critical difference from a typical AI workflow: every single agent output gets scored immediately. Nothing accumulates unchecked.

In the same Window 1 conversation, paste the following prompt to run the full Brand team with inline scoring:

You are running Team A (Brand) for Casa Verde Taqueria.

Read the approved competitive matrix at Projects/01_casa_verde_taqueria/02-Research/competitive-matrix.md
Read each agent template at Council/agents/ (A1 through A5)
Read the Design Canon at Resources/Canon/design-canon-quickref.md
Read the Artifact Approval Protocol at Council/artifact-approval-protocol.md

Execute each agent in order. After EACH agent output, score it on its specific axes (1-10 each). If ANY axis scores below 7, REWORK that output with specific feedback about what failed before moving to the next agent.

Agent A1 - Brand Positioning:
Score on: Positioning Clarity, Competitive Differentiation, Client Alignment
Save output to: Projects/01_casa_verde_taqueria/03-Brand/agent-outputs/A1-positioning.md

Agent A2 - Name/Logo Concepts:
Score on: Originality, Scalability, Brand Fit, Trademark Safety
Save output to: Projects/01_casa_verde_taqueria/03-Brand/agent-outputs/A2-name-concepts.md

Agent A3 - Typography:
Score on: Readability, Personality Match, Pairing Harmony
Save output to: Projects/01_casa_verde_taqueria/03-Brand/agent-outputs/A3-typography.md

Agent A4 - Color Palette:
Score on: WCAG Compliance (4.5:1 minimum), Emotional Resonance, Versatility, 60-30-10 Adherence
Save output to: Projects/01_casa_verde_taqueria/03-Brand/agent-outputs/A4-color-palette.md

Agent A5 - Brand Voice:
Score on: Consistency, Memorability, Client Tone Match
Save output to: Projects/01_casa_verde_taqueria/03-Brand/agent-outputs/A5-brand-voice.md

For every rework: note which axis failed, what score it received, and what specific feedback you gave to fix it. Save ALL scores (including rework attempts) to: Projects/01_casa_verde_taqueria/07-Approval-Scores/brand-team-scores.yaml

Format the scores file as:
agent: A1
first_pass_scores:
  positioning_clarity: X
  competitive_differentiation: X
  client_alignment: X
rework_count: 0
final_scores:
  positioning_clarity: X
  competitive_differentiation: X
  client_alignment: X
(repeat for each agent)

What does this prompt do?

This single prompt runs all 5 Brand agents in sequence with built-in quality gates. Each agent must score 7 or higher on every relevant axis before the next agent begins.

The WCAG compliance check on A4 (Color Palette) ensures your color combinations meet the 4.5:1 contrast ratio minimum for normal text. This is not aesthetic preference. It is an accessibility standard that determines whether elderly users or people with low vision can read the taqueria's menu board.

The 60-30-10 rule on A4 means 60% dominant color, 30% secondary, 10% accent. This prevents the palette from becoming visually chaotic when applied to real materials like signage, menus, and packaging.

Wait for Claude to work through all 5 agents. This is the longest single prompt in the project and may take 3-5 minutes.
Watch the conversation as Claude runs each agent. You will see it score, potentially rework, and then move to the next.
When complete, verify the outputs exist by asking Claude: "List the contents of Projects/01_casa_verde_taqueria/03-Brand/agent-outputs/"

You should see 5 files: A1-positioning.md, A2-name-concepts.md, A3-typography.md, A4-color-palette.md, and A5-brand-voice.md. You should also see the scores file at 07-Approval-Scores/brand-team-scores.yaml.

Reworks are progress, not failure

If Claude reworks an agent 2-3 times, that is the system working as designed. First-pass approval rates for beginners typically land around 60%. Each rework teaches the system what "good" looks like for YOUR client. The Intelligence Layer captures those rework patterns so future passes improve.

Claude stopped mid-way through the agents?

If Claude's response cuts off before completing all 5 agents, it likely hit a response length limit. Ask it to continue from where it stopped. For example: "Continue from Agent A3. You completed A1 and A2 successfully."

If scores are consistently below 7 and reworks are not improving, the competitive matrix may be too thin. Go back to Window 1 and ask Claude to expand the matrix with 2-3 more competitors before re-running the failing agent.

help me figure out why my Brand agent keeps scoring below 7 even after reworks

Run Standards Gate and produce the aggregate scorecard

The Standards Gate is your final quality checkpoint before anything moves to Paper for visual production. It reviews all agent outputs together, checking for internal consistency and canon compliance across the full brand system.

In the same Window 1 conversation, paste the following prompt to run the Standards Gate and produce the aggregate scorecard:

You are running Standards Gate for the Casa Verde Taqueria Council pass.

Read:
- Standards Gate template: Council/standards-gate.md
- Artifact Approval Protocol: Council/artifact-approval-protocol.md
- All Team A outputs: Projects/01_casa_verde_taqueria/03-Brand/agent-outputs/ (all 5 files)
- Research matrix: Projects/01_casa_verde_taqueria/02-Research/competitive-matrix.md
- Research scores: Projects/01_casa_verde_taqueria/07-Approval-Scores/research-scores.yaml
- Brand scores: Projects/01_casa_verde_taqueria/07-Approval-Scores/brand-team-scores.yaml
- Design Canon: Resources/Canon/design-canon-quickref.md

Execute Standards Gate checks:
1. Internal Consistency: Do A1-A5 outputs align with each other? (Colors match voice, typography matches positioning, etc.)
2. Canon Compliance: Does every output respect Golden Ratio, Gestalt principles, 60-30-10, WCAG 4.5:1, typography hierarchy?
3. Client Brief Alignment: Does the final brand system answer the original brief's requirements?
4. Competitive Differentiation: Does the brand system stand apart from the competitive matrix?

Score the Gate Report on: Thoroughness (1-10)

Then produce the AGGREGATE SCORECARD:
- Average ALL individual scores across Team C and Team A outputs
- Calculate the aggregate average
- If aggregate < 7.5: flag "OPERATOR REVIEW REQUIRED BEFORE PAPER"
- If aggregate > 8.5: flag "PORTFOLIO CANDIDATE"

Save the aggregate scorecard to: Projects/01_casa_verde_taqueria/07-Approval-Scores/council-pass-scorecard.yaml

Format:
client: Casa Verde Taqueria
council_pass: 1
date: [today]
research_scores:
  market_coverage: X
  insight_quality: X
  data_freshness: X
brand_scores:
  A1_average: X
  A2_average: X
  A3_average: X
  A4_average: X
  A5_average: X
aggregate_average: X
status: APPROVED / OPERATOR_REVIEW / PORTFOLIO_CANDIDATE
gate_thoroughness: X
notes: [any consistency issues found]

What does the aggregate scorecard tell you?

The aggregate average is your single number for the entire Council pass. Below 7.5 means the work is not ready for Paper production. You would need to review the weakest outputs and decide whether to rework specific agents or adjust the brief itself.

Above 8.5 means the work is strong enough to eventually appear in your portfolio. This flag matters because portfolio pieces need to represent your best capabilities, not just "good enough for the client."

Wait for Claude to complete the Standards Gate and produce the scorecard.
Check the aggregate average. If it is below 7.5, Claude will flag it for your review. Read the notes to understand which outputs need attention.
Verify the scorecard file by asking Claude: "Read Projects/01_casa_verde_taqueria/07-Approval-Scores/council-pass-scorecard.yaml"

You should see a complete YAML file with all scores filled in and a status of either APPROVED, OPERATOR_REVIEW, or PORTFOLIO_CANDIDATE.

Now switch to Window 3 (Approval and Intelligence) and paste this prompt to capture what the system learned:

You are the Intelligence Agent for the Casa Verde Taqueria Council pass.

Read:
- Intelligence Agent template: Council/intelligence-agent.md
- All approval scores: Projects/01_casa_verde_taqueria/07-Approval-Scores/ (all yaml files)
- All agent outputs: Projects/01_casa_verde_taqueria/03-Brand/agent-outputs/

Capture the following intelligence from this Council pass:

1. Rework Analysis:
   - Which agents required reworks?
   - Which axes failed most often?
   - What specific feedback improved scores?

2. Pattern Recognition:
   - What worked well on first pass? (What does this client type respond to?)
   - What consistently struggled? (What do Austin food businesses need extra attention on?)

3. Baseline Metrics:
   - Total rework count across all agents
   - Average first-pass score vs final score
   - Time indicator (this is pass 1, future passes should be faster)

Save to:
- Areas/Intelligence/lessons-learned.md (append new section for Casa Verde pass 1)
- Areas/Intelligence/approval-history.yaml (structured log of all approval decisions)

Format approval-history.yaml as:
- client: Casa Verde Taqueria
  pass: 1
  decisions:
    - agent: C1
      first_pass: APPROVED/REJECTED
      final: APPROVED
      rework_count: X
    - agent: A1
      first_pass: APPROVED/REJECTED
      final: APPROVED
      rework_count: X
    (etc for all agents)
  aggregate: X
  key_lesson: "one sentence summary of biggest learning"

Why capture intelligence separately?

The Intelligence Layer runs in a separate window because it needs to analyze the Council pass as a whole, not as individual agent outputs. It identifies patterns that span multiple agents.

This data compounds over time. After 3 Council passes, the Intelligence Agent can tell you that "Austin food businesses consistently need 2 reworks on typography because readability at menu-board distance is a non-obvious constraint." That pattern then gets embedded into the A3 agent template permanently.

Wait for Claude to finish the intelligence capture.
Verify the files exist by asking Claude in Window 3: "List the contents of Areas/Intelligence/"

You should see lessons-learned.md and approval-history.yaml both populated with data from this Council pass.

Intelligence files are empty or missing?

If Claude says it cannot read the approval score files, confirm they were saved correctly in the earlier prompts. Ask Claude in Window 1: "List all yaml files in Projects/01_casa_verde_taqueria/07-Approval-Scores/" to verify they exist.

If the files exist but Window 3 cannot read them, the filesystem MCP connection may have dropped. Restart Claude Desktop and reopen Window 3.

help me troubleshoot why my Intelligence Agent cannot read the approval score files from another Claude Desktop window

Your first Council pass is complete and scored. Every artifact has been rated, failures have been captured, and the Intelligence Layer is seeded with its first real data. Next up, you will take these approved brand artifacts into Paper and produce the actual logo design.

Produce in Paper, Rate Visually, and Deliver Proposal

Your Council has done its job. Casa Verde has approved brand outputs: positioning, name concepts, typography, color palette, brand voice. All scored 7+ and sitting in your vault.

Now those decisions become a real logo. You will load the E1 Paper Integration agent, have Claude create artboards directly in Paper via MCP, then score what you see. If it does not hit 7 on every visual axis, it goes back. This is where the Approval System earns its keep.

In this step, get ready to:

Create the Casa Verde logo in Paper using approved Council outputs.
Rate the visual output on 6 design quality axes and iterate until all hit 7+.
Export the final logo and generate a scored value-based proposal.

Load E1 Paper Integration and create artboards

The E1 Paper Integration agent translates approved brand decisions into visual elements on the Paper canvas. Claude reads the approved outputs from your vault and uses the Paper MCP connection to create artboards, shapes, and text directly in the app.

Switch to Window 1 (Client Pipeline).
Make sure Paper Desktop is open with a file (the MCP server only runs when Paper has a file open).
Paste the following prompt into Claude to load E1 and begin production:

Load the E1 Paper Integration agent from Council/agents/.

Read the approved brand outputs from:
- 01_casa_verde_taqueria/03-Brand/agent-outputs/

Using the Paper MCP connection, create artboards for the Casa Verde Taqueria logo:
1. Primary logo artboard (1024x1024)
2. Horizontal lockup artboard (2048x512)
3. Icon-only artboard (512x512)

Apply the approved color palette, typography choices, and positioning from the Team A outputs.
Use the Golden Ratio (1.618) for proportional relationships between logo elements.
Apply Gestalt principles (proximity, similarity) to group related elements.

Show me what you created.

What does this prompt do?

You are asking Claude to read the brand decisions your Council already approved (colors, fonts, positioning) and translate them into actual shapes on the Paper canvas.

The three artboard sizes cover different use cases. 1024x1024 is the primary square logo. 2048x512 is for horizontal placements like website headers. 512x512 is the simplified icon version for social media and favicons.

Claude will use Paper's MCP tools to create nodes, apply styles, and compose the logo elements. You should see new artboards appear in your Paper canvas within a few seconds.

Nothing appearing in Paper?

Confirm Paper Desktop is open with a file. The MCP server at http://127.0.0.1:29979/mcp only starts when Paper has an active document.

If Claude says it cannot reach Paper, restart Claude Desktop. MCP connections initialize at launch.

help me troubleshoot why Claude cannot connect to Paper via MCP

Rate the visual output on 6 axes

The logo is on your canvas. Now you score it. The Design Quality Scorecard measures 6 specific axes. Each axis gets a 1-10 rating. If any axis scores below 7, the logo does not move forward.

Look at the logo Claude created in Paper.
Score each of the following 6 axes from 1 to 10:

Composition - Does it use the Golden Ratio (1.618) for proportions? Is the visual weight balanced?
Color Harmony - Does it follow the 60-30-10 rule (dominant, secondary, accent)? Do colors meet WCAG 4.5:1 contrast?
Typography - Is there clear hierarchy? Is readability maintained at small sizes?
Negative Space - Is white space intentional and purposeful, not just empty?
Scalability - Does the logo hold up from 16px favicon to billboard size?
Emotional Impact - Does it feel like the brand? Does it evoke the right response?

How do I score if I am new to design?

Use the Squint Test. Squint at the logo until it blurs. If the hierarchy disappears (nothing stands out as primary), score Composition lower. If the colors feel muddy or clash, score Color Harmony lower. If you cannot tell what the brand name says, score Typography lower.

Trust your gut on Emotional Impact. You know what a taqueria feels like. Warm. Inviting. A little bold. If the logo does not feel like that, score it honestly.

If ANY axis scores below 7, tell Claude exactly what scored low and why. Paste this iteration prompt:

The logo scored below 7 on these axes:

- [AXIS NAME]: [SCORE]/10
  Why: [YOUR SPECIFIC REASON - e.g. "the icon and text feel disconnected, no visual grouping"]

Iterate on the Paper design to fix ONLY these issues.
Do not change elements that scored 7+.
Show me the updated version.

Why specific feedback matters

Telling Claude "make it better" produces random changes. Telling Claude "Composition scored 5 because the icon and wordmark have no proportional relationship" produces targeted fixes.

Each iteration also teaches you design vocabulary. By Week 4, you will articulate design problems faster than most junior designers because you have practiced scoring and explaining dozens of artifacts.

Repeat scoring and iteration until all 6 axes hit 7 or higher.
Once all axes pass, save the scores to your vault by pasting this prompt:

Save the following Design Quality Scorecard to
01_casa_verde_taqueria/07-Approval-Scores/paper-design-scorecard.yaml
via the filesystem MCP:

artifact: Casa Verde Taqueria Primary Logo
date: 2026-05-17
iterations: [NUMBER OF ITERATIONS]
final_scores:
  composition: [SCORE]
  color_harmony: [SCORE]
  typography: [SCORE]
  negative_space: [SCORE]
  scalability: [SCORE]
  emotional_impact: [SCORE]
aggregate: [AVERAGE]
status: APPROVED
notes: [WHAT FEEDBACK IMPROVED SCORES MOST]

Also append to Areas/Intelligence/lessons-learned.md:
- What axes scored lowest on first pass
- What specific feedback improved those axes
- How many iterations were needed

What does the Intelligence Layer learn here?

The Intelligence Agent captures which axes fail most often and what feedback fixes them. After 10 logo passes, a pattern emerges. Maybe your Composition scores always start low because Claude defaults to centered layouts with no proportional system.

That pattern gets promoted into the E1 agent template. Future logos start with Golden Ratio grids applied automatically. The system stops making the same mistake.

Stuck in iteration loops?

If you have iterated 4+ times and an axis will not climb above 6, the problem may be in the brand inputs (wrong typography pairing, conflicting palette). Go back to the Team A outputs and check whether the approved decisions work together visually.

help me figure out why my Paper design keeps scoring low on a specific axis after multiple iterations

Export the logo and generate a scored proposal

The logo passed all 6 axes at 7+. Now you export it in production formats and generate a value-based proposal that also goes through the Approval System.

Ask Claude to export the final logo and generate the proposal by pasting this prompt:

Export the approved Casa Verde Taqueria logo from Paper:
1. SVG (vector, scalable)
2. PNG (1024x1024, transparent background)
3. PDF (with brand mockup showing logo on a taco wrapper and storefront sign)

Save exports to 01_casa_verde_taqueria/04-Production/logo-exports/ via filesystem MCP.

Then load agents F3 (Proposal Writer) and D2 (Client Manager).
Read:
- 01_casa_verde_taqueria/00-Discovery/brief.md
- Areas/Operations/pricing-ladder.yaml
- Resources/Templates/proposal-template.md
- Areas/Operations/client-weapons/value-pricing-conversation-script.md

Generate a value-based proposal for Casa Verde Taqueria:
- Premium option first (anchoring)
- Recommended middle option (compromise effect)
- Budget option (feels incomplete by comparison)
- Include ROI calculation
- Include "cost of doing nothing" frame
- 50% upfront / 50% on delivery payment structure

Save proposal to 01_casa_verde_taqueria/05-Delivery/proposal.md

Why three pricing tiers in one proposal?

Psychological anchoring. The Premium option ($2,500+) makes the Recommended option ($1,500) feel reasonable by comparison. The Budget option ($800) feels incomplete, pushing clients toward the middle.

This is not manipulation. It is honest framing. Each tier delivers real value at its price point. The structure helps the client make a decision instead of freezing.

Once Claude saves the proposal, score it on 4 axes. Each must be 8+ (proposals have a higher bar than production work because they represent your business directly):

Value Communication - Does the client understand what they get and why it is worth the price?
Anchoring Effectiveness - Does the Premium option make the Recommended option feel like the obvious choice?
Urgency - Is there a reason to act now rather than next month?
Professionalism - Would you trust this document if someone sent it to you?

If any axis scores below 8, tell Claude what scored low and iterate (same pattern as the design scoring).
Once all 4 axes hit 8+, save the proposal scores and update Intelligence:

Save proposal scorecard to
01_casa_verde_taqueria/07-Approval-Scores/proposal-scorecard.yaml
via filesystem MCP:

artifact: Casa Verde Taqueria Proposal
date: 2026-05-17
scores:
  value_communication: [SCORE]
  anchoring_effectiveness: [SCORE]
  urgency: [SCORE]
  professionalism: [SCORE]
aggregate: [AVERAGE]
status: APPROVED

Append to Areas/Intelligence/approval-history.yaml:
- Paper design approval decision (scores, iterations)
- Proposal approval decision (scores)

Append to Areas/Intelligence/lessons-learned.md:
- What made the proposal score well (or what needed fixing)
- Patterns for future proposals

What does your delivery folder now contain?

The 01_casa_verde_taqueria/05-Delivery/ folder holds your approved proposal. The 01_casa_verde_taqueria/04-Production/logo-exports/ folder holds SVG, PNG, and PDF files ready to attach when you send the proposal.

Every artifact behind this proposal has been scored. The brief was scored. The Council outputs were scored. The Paper design was scored. The proposal itself was scored. When a client asks why you charge $1,500 for a logo, you have a documented quality process that justifies every dollar.

Verify your vault now contains these new files by asking Claude:

List the contents of:
- 01_casa_verde_taqueria/04-Production/logo-exports/
- 01_casa_verde_taqueria/05-Delivery/
- 01_casa_verde_taqueria/07-Approval-Scores/

You should see logo files (SVG, PNG, PDF) in 04-Production/logo-exports/, your proposal in 05-Delivery/, and three score files in 07-Approval-Scores/ (brief, council pass, paper design scorecard).

Missing files in the vault?

If Claude says directories do not exist, it may not have created them yet. Ask Claude to create the missing folders first, then save the files.

Check that your filesystem MCP path points to C:\Users\YourName\Documents\Studio with double backslashes in the JSON config.

help me verify my vault structure and find missing files

Your first approved deliverable is ready to send. A scored logo. A scored proposal. Every decision documented. Next up, you will capture everything the Intelligence Layer learned from this entire pass and set up the growth cadence that keeps the system improving every week.

Capture Intelligence and Set Up Growth Cadence

Casa Verde is done. The logo is exported, the proposal is scored 8+, and the client folder is airtight.

But the real power of this system is not what it produced. It is what it learned from producing it. Every score, every rework, every feedback loop contains intelligence that makes the next client pass faster and better. In this step, you will run the Intelligence Agent to capture that knowledge, build a recurring cadence that keeps the system improving, and create a dashboard to track your trajectory from 60% first-pass approval to 85%.

In this step, get ready to:

Run the Intelligence Agent on your entire Casa Verde pass to establish baseline metrics.
Set up Growth Pipeline cadence with scoring integrated and an acquisition plan.
Create the Improvement Velocity Dashboard and commit your vault to git.

Run the Intelligence Agent on Casa Verde

The Intelligence Agent is the 7th component of your pipeline. Its job is to read every approval decision, every score, and every piece of feedback from a completed client pass and distill it into structured intelligence. This intelligence feeds future Council runs so agents do not repeat the same mistakes.

Switch to Window 3 (Approval and Intelligence).
Run the Intelligence Agent on the entire Casa Verde pass by sending this prompt:

Run the Intelligence Agent on the full Casa Verde Taqueria pass.

Capture the following to Areas/Intelligence/lessons-learned.md:
- Hardest scoring axes (which axes scored lowest across all artifacts)
- Total rework count (how many times an artifact was sent back)
- Feedback that improved scores (what specific notes led to score increases)
- Standards Gate catches (what the gate flagged before approval)
- Aggregate score (average across all artifact scores)
- Time spent (estimate based on conversation length)

Also update:
- Areas/Intelligence/quality-scores.yaml with all Casa Verde artifact scores
- Areas/Intelligence/design-decisions-log.md with key design choices made
- Areas/Intelligence/approval-history.yaml with the complete decision trail

This establishes baseline. Format everything as structured data I can compare against future passes.

What does this capture do?

The Intelligence Agent reads the approval scores from 07-Approval-Scores/ and the existing entries in Areas/Intelligence/. It organizes everything into a structured baseline you can measure future passes against.

The key insight: future improvements equal faster time at equal or higher scores. If your next client pass takes less time but scores the same or better, the system is compounding.

Review what Claude generates. Confirm it created entries in lessons-learned.md, quality-scores.yaml, design-decisions-log.md, and updated approval-history.yaml.
Verify that the hardest scoring axes and rework count are clearly identified. These are the areas you improve first.

Claude not finding the score files?

Make sure Claude can access the filesystem MCP. If it says it cannot read files, confirm the Studio vault path in your MCP config uses double backslashes. Check that Window 3 has filesystem access by asking Claude to list the contents of 01_casa_verde_taqueria/07-Approval-Scores/.

Help me troubleshoot why Claude cannot read my vault files through MCP

Set up Growth Pipeline cadence and acquisition plan

A pipeline without a cadence decays. The Growth Pipeline ensures you review scores weekly, spot trends monthly, and promote winning patterns quarterly. It also defines how you find new clients beyond the 3 you already have.

Stay in Window 3. Generate the Growth Pipeline cadence document by sending this prompt:

Create Areas/Operations/growth-cadence.md with the following sections:

## Sunday Council Cadence
- Review all active client scores from the week
- Run Brand Council on score trends (are averages rising or flat?)
- Identify one axis to deliberately improve next week
- Update Intelligence Layer with weekly observations

## Monthly Brief
- Average scores across all clients (compare to previous month)
- Score trajectory visualization (text-based chart)
- Revenue per artifact this month
- Rework rate trend (number of rejects before approval)
- One promoted pattern (a lesson that consistently improved scores)

## Quarterly Promotion Process
- Any lesson that improved scores 2+ points across multiple passes gets permanently baked into the relevant agent template
- Update the agent's Canon section with the promoted insight
- Document the promotion in Areas/Intelligence/pattern-library.md
- This is how agents get smarter over time without manual prompt editing

## Acquisition Plan
- Referrals: Ask each satisfied client for 1 introduction
- Behance: Publish 1 scored portfolio piece per month
- Austin local groups: Austin Creative Alliance, AIGA Austin, East Austin Studio Tour
- Cold outreach: 5 per week to Austin small businesses without professional branding
- LinkedIn: Weekly story content showing process (scores, before/after, lessons learned)
- Target: 2 new clients per month by Month 3

Why bake lessons into agent templates?

The quarterly promotion process is the compounding engine. When a lesson improves scores by 2+ points consistently, it stops being a note in the Intelligence Layer and becomes part of the agent's permanent instructions. By Month 6, your agents operate on canon plus 6 months of Austin-specific experience no competitor can copy.

Confirm Claude created Areas/Operations/growth-cadence.md with all four sections: Sunday Council, Monthly Brief, Quarterly Promotion, and Acquisition Plan.

Set a real Sunday reminder

Open Windows Task Scheduler and create a weekly reminder for Sunday morning. This keeps the cadence alive even when client work gets busy.

Create the Improvement Velocity Dashboard and commit

The dashboard is a single markdown file that tracks 6 metrics over time. It gives you a snapshot every week of whether the system is improving or stalling.

In Window 3, generate the dashboard by sending this prompt:

Create Areas/Intelligence/dashboard.md as the Improvement Velocity Dashboard.

Track these 6 metrics (seed with Casa Verde baseline data):

1. Average score per week — current baseline from Casa Verde aggregate. Target: trending up.
2. First-pass approval rate — percentage of artifacts that score 7+ on first submission. Baseline: ~60%. Target: 85% by Month 3.
3. Rework rate — average number of rejects per artifact before approval. Target: trending down.
4. Time per Council pass — estimated time from brief to approved output. Target: trending down.
5. Revenue per artifact — dollar value delivered per scored piece. Current: $1,500 (Casa Verde logo).
6. Client satisfaction correlation — track which score axes correlate with positive client feedback.

Format as a markdown table with columns: Metric | Baseline (Week 1) | Current | Target | Trend
Include a section for weekly entries where I log new data points each Sunday.
Add a "Milestones" section: Day 30 = first invoice paid. Day 60 = debt-free. Day 90 = 85% first-pass rate.

What do these metrics tell you?

The first-pass approval rate is the most important number. It measures how often your Council produces work that passes the quality gate on the first attempt. Starting at 60% means 4 out of 10 artifacts need rework. Reaching 85% means the system learned what "good" looks like in your market.

Revenue per artifact and time per pass together give you your effective hourly rate. As time drops and revenue holds steady, your rate climbs from the baseline toward that $150/hr target.

Verify Claude created Areas/Intelligence/dashboard.md with all 6 metrics seeded with Casa Verde baseline data.
Now commit your entire vault to Git. Open your terminal (Git Bash or Command Prompt) and run these commands:

cd C:\Users\%USERNAME%\Documents\Studio
git add -A
git commit -m "Construction Paper: Foundation complete"

What does this commit do?

The git add -A command stages every new, modified, and deleted file in your vault. The commit saves a snapshot of your entire studio at this moment. You can always return to this state if something breaks later.

After running the commit, you should see a summary listing all the files added. Look for entries from Areas/Intelligence/, Areas/Operations/, and all 3 client folders.

Git says 'not a git repository'?

You may not be in the right folder. Make sure you ran cd C:\Users\%USERNAME%\Documents\Studio first. If you skipped the git init step from Step 1, run it now before the git add command.

Help me fix my git repository setup in the Studio vault

Your studio now has a memory. Every future client pass feeds the Intelligence Layer, raises your scores, and proves to clients that your process is accountable. Next up, the Secret Mission takes all 3 clients through the full scored pipeline and aims for $9,000 in proposals.

Secret mission

Score 8+ on Every Artifact and Ship $9,000 in Proposals

You scored one client through the pipeline. Now run all three. Achieve 8+ on every artifact across all 3 Austin clients, generate $9,000 in proposals, publish your first scored portfolio piece with a full case study, seed the Intelligence Layer with 15+ approval decisions, and bring in Gemini to challenge your grading.

Clean Up Your Resources

This project uses recurring paid subscriptions that charge monthly. Decide whether to keep your resources running, pause them to come back later, or delete them entirely.

Cost warning

Claude Max 20x costs $200/mo and Paper Pro costs $16-20/mo. Together that is $216-220/mo in recurring charges.

If you are not actively sending proposals and building your client pipeline, consider pausing or canceling these subscriptions to avoid unnecessary costs.

Resources you used:

Claude Max 20x subscription ($200/mo recurring).
Paper Pro subscription ($16-20/mo recurring).
Firecrawl free account (1,000 credits/mo).
Obsidian vault at C:\Users\YourName\Documents\Studio (local files and Git repository).
Installed applications: Obsidian, Git for Windows.
MCP configuration in Claude Desktop.

Keep everything running

No action needed. Choose this if you are actively sending proposals and building your client pipeline.

Your proposals go out Monday. Day-30 target: first invoice. Day-60 target: debt-free.
Claude Max 20x and Paper Pro remain active at $216-220/mo combined.
Your Obsidian vault, MCP connections, and all agent templates remain ready.
The Intelligence Layer continues to compound with every new client interaction.

Pause - I'll come back to this later

Reduce monthly costs while keeping your vault, templates, and system structure intact.

Downgrade Paper Pro to the Free tier through your account settings at paper.design. The Free tier gives you 100 MCP tool calls/week at no cost.
Keep Claude Max 20x active. This is the core engine that powers the entire Council. Without it, the 25-agent pipeline cannot run.
Your Obsidian vault, Git repository, and all local files remain untouched on your machine.
When your first client payment arrives, upgrade Paper back to Pro to resume full production capacity.

Emergency budget mode

If you need to cut costs further, you can temporarily downgrade Claude to Pro ($20/mo) instead of Max 20x. You lose the 20x capacity per session, which means longer Council runs and more manual splitting of prompts. Upgrade back to Max 20x as soon as your first invoice is paid.

Delete - I don't want to use this again

Remove all project resources and cancel recurring charges. This is permanent.

Cancel your subscriptions:

In Claude Desktop, click your initials in the lower left corner, then select Settings. Navigate to Billing and click Cancel. Your Max 20x access continues until the end of your current billing period.
Cancel your Paper Pro subscription through your account settings at paper.design. You retain access until the end of your billing cycle.
Delete your Firecrawl account through your account settings at firecrawl.dev if you no longer need web scraping capabilities.

Delete local files:

Delete your Obsidian vault and Git repository by running this command in Command Prompt:

rmdir /s /q "C:\Users\[[WINUSER="your Windows username"]]\Documents\Studio"

Reset your MCP configuration in Claude Desktop by going to Settings > Developer > Edit Config and removing the server entries you added.

Uninstall applications:

Press the Windows key and search for Add or remove programs. Press Enter to open it.
Find Obsidian in the list, click it, and select Uninstall.
Find Git in the list, click it, and select Uninstall.

That's a Wrap!

Construction Paper is live. You built a quality-controlled product design studio from a Windows desktop in Austin, powered by 25 agents that score every artifact before it ships.

You've learned how to:

Build a 25-agent Council with embedded design canon (Golden Ratio, Gestalt, 60-30-10, WCAG) and a 6-axis Artifact Approval System that scores every output before it moves forward. Nothing ships below a 7.
Wire Paper, Firecrawl, and filesystem MCP into a parallel production loop across multiple Claude Desktop windows, producing scored designs, competitive research, and value-based proposals.
Structure 3 real Austin client briefs, run the first scored Council pass with all artifacts approved at 7+, seed the Continuous Intelligence Layer, and prepare a $9,000 proposal pipeline with client management weapons built in.
Secret Mission: Run all 3 clients through the full scored pipeline with 8+ on every artifact, generate $9,000 in proposals, publish a portfolio piece with visible scores, seed the Intelligence Layer with 15+ approval decisions, and run Gemini adversarial review to validate your scoring calibration.

Ready to quiz yourself?