Build an AI-Assisted UX Prototype

Introduction

30 Second Summary

Messy feedback rarely points in one direction. Similar comments can hide different needs while confident interpretations can drift beyond what people actually said.

In this project, you will use Claude Free to turn ten fictional feedback fragments into evidence-backed themes and a human-owned product direction. You will build Trust Finder as an interactive Claude Artifact that helps people find datasets and inspect the evidence behind their trust status.

What You'll Build

You will demo Trust Finder by searching for customer data, filtering certified assets, and opening a dataset to inspect its trust evidence.

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

• An evidence map that connects every accepted theme to the original feedback IDs.
• A human-owned design decision that compares three concepts and records why Trust Finder was selected.
• An interactive Trust Finder prototype with search, filters, dataset details, transparent status evidence, and similar-asset guidance.
• Secret Mission: Add a neutral side-by-side comparison that helps people distinguish two similar datasets without declaring a winner.

Are there any prerequisites?

You need a Mac with Safari and a personal Claude Free account. Product or UX design experience is helpful.

You do not need coding experience, an API key, or external research.

Before We Start

Before any hands-on work, lock in what you are building and which design decision stays yours. This commitment keeps Claude in the role of a fast collaborator while you remain accountable for the final choices.

Set Up Claude Artifacts

You already defined the boundary between AI assistance and human design judgment. The next risk is whether your Claude workspace can support the interactive prototype you plan to evaluate.

This step checks whether Claude can render an interactive Artifact in Safari. A tiny counter proves that Artifact creation works before you invest time in synthesis.

In this step, get ready to:

• Start one dedicated Claude conversation for the entire project.
• Enable Cloud code execution and file creation.
• Verify Artifact creation with a working interaction counter.

Start a dedicated Claude conversation

One conversation keeps your fictional research packet beside every human decision. It also gives Claude the context needed for later prototype iterations.

• Visit https://claude.ai in Safari.

You should see the official Claude web experience with an option to sign in.

• Sign in with your personal Claude Free account.
• Start a fresh conversation from the left sidebar.

Your blank conversation is now ready to hold the complete project context.

• Use this conversation for every later research prompt.
• Use the same conversation for every later prototype prompt.

Enable Artifact creation

Claude needs permission to run cloud code before it can create interactive Artifacts. The required capability lives in your account settings.

• Open Settings from your Claude account menu.
• Select Capabilities.

You should see the Cloud code execution and file creation setting.

• Turn on Cloud code execution and file creation.

The enabled setting confirms that Claude can create the interactive preview used in this project.

• Return to the dedicated conversation you started moments ago.
• Create the readiness check by pasting this prompt into the message box:

Create a tiny self-contained interactive Artifact called "Artifact Readiness Check." Show a button labeled "Test interaction," a visible count starting at 0, and increment the count on every click. Do not use external assets, packages, APIs, storage, or network requests.

What does this prompt test?

• The visible count creates a clear baseline before you interact with the Artifact.
• The button proves that the preview can respond to your input.
• The restrictions keep this readiness check independent from external services.

• Submit the prompt to Claude.
• Wait for Claude to finish creating the Artifact.

You should see an Artifact titled Artifact Readiness Check beside the conversation. It should contain a visible count of 0 plus a Test interaction button.

Artifact not rendering?

• Select Try fixing with Claude if the Artifact displays an error.
• Download Claude Desktop for Mac from the official Claude download page if Safari cannot render the preview.
• Install Claude Desktop for Mac using the downloaded installer.
• Sign in to Claude Desktop with the same Claude account.
• Return to the same project conversation in Claude Desktop.
• Help me troubleshoot an Artifact that does not render in Safari or Claude Desktop.

Verify the interaction

The rendered preview proves that Claude created an Artifact. The changing count proves that the Artifact can handle interaction.

Before you click, what count do you expect to see after two clicks?

• Confirm that the visible count starts at 0.

The initial 0 gives you a clear baseline for the interaction test.

• Click Test interaction once.
• Click Test interaction a second time.

You should now see a visible count of 2. Claude should not ask for an API key or a local development environment.

That clears the main setup risk: your Artifact responds to input inside the dedicated project conversation.

Your Claude conversation now has a working interactive baseline. Next, you will use the same context to synthesize the fictional feedback packet.

Synthesize Feedback into Themes

Your readiness check proved that Claude can create an interactive Artifact in your dedicated project conversation. That conversation can now hold the evidence trail behind Trust Finder.

Raw feedback is noisy. Some comments qualify others. Claude can spot patterns quickly. You remain responsible for tracing every pattern to the fictional source fragments.

In this step, get ready to:

• Submit the complete fictional feedback packet.
• Audit Claude's themes against the source fragments.
• Record the approved synthesis as a human decision.

Submit the fictional research packet

The research packet contains ten perspectives on finding trustworthy enterprise data. The prompt asks Claude to separate direct observations from interpretations before grouping the evidence.

• Return to the dedicated Claude conversation from earlier.
• Prepare the complete research request by copying this prompt:

Act as a senior UX research synthesis partner. This is fictional discovery material for an enterprise data catalog. Use only the evidence below. Do not invent users, quotes, metrics, causes, or certainty.

Scenario:
Employees struggle to find trustworthy datasets, understand metadata, and distinguish between similarly named data assets.

Feedback:
F1, Senior analyst: "I searched for customer revenue and got six tables with almost the same name. I picked the newest-looking one, then Finance told me it was a legacy extract."
F2, Governance lead: "We keep telling people to use certified data, but the certification badge is easy to miss and nobody knows what checks sit behind it."
F3, Data steward: "Ownership goes stale. Two popular assets still list someone who left the company months ago, so questions land in a shared Slack channel."
F4, Marketing analyst: "Descriptions read like pipeline documentation. I need to know what business question a dataset can answer, not which job writes the table."
F5, Analytics engineer: "The catalog shows a last-updated timestamp, but users confuse that with the expected refresh schedule and assume a healthy monthly table is stale."
F6, Risk stakeholder: "Certification does not mean appropriate for every use. A governed finance table can still be wrong for campaign targeting."
F7, Product analyst: "Search works if I know the exact table name. It does not help when the business says customer value and engineering says lifetime_value_agg."
F8, Product manager: "Please do not solve this with another dashboard. The decision support needs to appear in the search flow where people choose an asset."
F9, Support lead: "Most catalog questions are variations of: Is this current, who owns it, and can I trust it? People ask us because those answers are scattered."
F10, Platform administrator: "Users request access before noticing restrictions or intended use. Then they wait for approval and discover the dataset cannot support their task."

Produce four sections:
1. Evidence map: one row per feedback item with speaker, explicit signal, likely underlying need, and a confidence note.
2. Themes: 3 to 5 themes. For each theme, cite the relevant feedback IDs, explain the pattern, include contradictory or qualifying evidence, and state what is still unknown.
3. Risk check: list any tempting conclusion that the evidence does not support.
4. Synthesis headline: one sentence that captures the central tension without proposing a solution.

Keep observations separate from interpretations. Do not recommend a feature yet. Do not ask follow-up questions.

What does this prompt control?

• The evidence map keeps each interpretation attached to one original feedback fragment.
• The theme rules require traceability to feedback IDs. They also preserve qualifying evidence.
• The risk check exposes conclusions that the fictional packet cannot support.
• The final constraints prevent feature recommendations from entering the synthesis too early.

• Paste the copied prompt into the message field in the same conversation.
• Send the prompt to Claude.
• Check that the evidence map contains one row for each ID from F1 through F10.
• Count the themes to confirm that Claude produced 3 to 5.
• Find the risk check beneath the themes.
• Find the one-sentence synthesis headline.

You should see an evidence map plus a table of 3 to 5 themes. A risk check and one-sentence synthesis headline should follow.

Response missing a section?

• Ask Claude to regenerate only the missing section from the original packet.
• Remove any output that introduces evidence beyond F1 through F10.
• Help me identify unsupported claims in Claude's synthesis.

Audit Claude's interpretation

An evidence-backed theme is useful only when you can trace it to the source fragments. This audit also protects the qualification in F6 that certification is contextual.

• Read F1 through F10 again in your original prompt.
• Compare the first theme with its cited source fragments.
• Repeat the same comparison for every remaining theme.
• Confirm that every theme cites at least two feedback IDs.
• Rename any theme that sounds like a feature.
• Remove every claim that lacks support from a cited ID.
• Preserve the qualification that certification is contextual instead of universal.
• Keep feature recommendations out of the approved synthesis.

Your reviewed synthesis should now separate direct evidence from interpretation. It should also make uncertainty visible.

Record and verify the approved synthesis

A decision record tells Claude which parts of the synthesis you approve. It also leaves a visible trail of the judgment that remained yours.

• Prepare your human decision record with this template:

Human synthesis decision: I accept [[ACCEPTED_THEMES="theme names"]], I revised [[REVISED_THEME_AND_REASON="theme and reason"]], and I reject or defer [[REJECTED_OR_DEFERRED_CLAIM="claim and reason"]]. Treat this as the approved synthesis for the rest of the project.

Why record the decision?

This sentence establishes your reviewed synthesis as the conversation's working source. Future concept work can now build from choices you approved.

• Complete the accepted themes field with the theme names you approved.
• Complete the revision field with the theme you changed plus your reason.
• Complete the rejected or deferred field with the claim you excluded plus your reason.
• Paste the completed record into the same Claude conversation.
• Send the decision record to Claude.

Before the final audit, do you think every accepted theme still meets the two-source rule after your edits?

• Scan the approved theme table to confirm that every theme cites at least two feedback IDs.
• Check that no invented users or quotes remain in the approved synthesis.
• Check that no invented metrics or causes remain in the approved synthesis.
• Check that no unsupported certainty or feature recommendation remains in the approved synthesis.
• Confirm that certification is described as contextual instead of universally appropriate.
• Locate the human synthesis decision record beneath Claude's response.

You should be able to trace every accepted theme to multiple source fragments. The conversation should also show what you accepted plus what you revised or excluded.

You have turned ten messy comments into an approved synthesis without letting AI inference pose as research evidence.

Theme still lacks traceability?

• Return to the cited feedback fragments to identify the exact statements supporting the theme.
• Remove the theme if fewer than two source fragments support it.
• Help me test whether this UX theme is supported by multiple feedback fragments.

Your evidence trail is approved. Next, you can turn these themes into prioritized needs without giving up control of the product decision.

Choose the Trust Finder Direction

Your approved synthesis now connects every accepted theme to the fictional feedback. You have also preserved the boundary that certification depends on context.

Themes still leave priority, scope, and tradeoffs unresolved. In this step, Claude will generate structured options while you choose the Trust Finder direction.

In this step, get ready to:

• Turn the approved synthesis into a prioritized opportunity frame.
• Compare three distinct concept directions through explicit tradeoffs.
• Record a human-owned decision that selects Trust Finder.

Generate needs, opportunities, and concepts

Strong user needs focus on the outcome people seek. The prompt below anchors each ranking to the approved evidence.

• Return to the dedicated Claude project conversation from the previous step.
• Generate the opportunity frame by pasting this prompt into the conversation:

Use only the approved synthesis in this conversation. Act as a product design strategist, but keep conclusions traceable to the feedback IDs.

Create:
1. Five user needs written as "People choosing a dataset need... so that..." Rank them using qualitative impact, frequency in this packet, and evidence confidence. Explain each ranking and cite feedback IDs. Do not turn features into needs.
2. Three "How might we" opportunities. Include one about findability, one about transparent trust, and one about distinguishing similar assets.
3. One problem statement naming the user, decision, obstacle, consequence, and evidence boundary. Make clear that this fictional packet suggests a direction but does not validate prevalence.
4. Three distinct solution concepts:
   - one search-first direction that combines discovery with transparent trust evidence,
   - one metadata-comprehension direction,
   - one comparison-first direction.
5. A concept tradeoff matrix using evidence fit, speed to prototype, support for contextual trust, cognitive load risk, and likely implementation complexity. Use Low, Medium, or High with one-sentence reasoning. Do not select a winner for me.
6. A skeptical challenge to each concept: one reason an experienced designer might reject it.

End with a blank "Human decision" section containing: chosen direction, rejected alternatives, rationale, assumption to test, and non-goal.

How the Prompt Protects Your Synthesis

The evidence constraints keep each need, opportunity, and concept traceable to F1 through F10.

The blank Human decision section reserves concept ownership for you.

• Send the prompt to Claude.
• Wait for Claude to finish the complete response.
• Check that all five needs follow the People choosing a dataset need... so that... structure.
• Check that every need includes a ranking rationale with feedback IDs.
• Confirm that the three opportunities cover findability, transparent trust, and similar assets.
• Confirm that the problem statement includes an evidence boundary.
• Confirm that the tradeoff matrix compares three distinct concepts without selecting a winner.

Good progress. You now have three directions to compare without handing the choice to Claude.

Did Claude Choose a Winner?

• Check that your submitted prompt still includes Do not select a winner for me.
• Ask Claude to regenerate the concept comparison with the human decision section left blank.
• Get help restoring the decision boundary.

Make and record the design decision

A tradeoff matrix makes evidence fit and design risk visible. Your judgment turns that comparison into a bounded product direction.

• Review the evidence-fit reasoning for each concept.
• Check whether the search-first concept addresses both discovery and transparent trust.
• Choose Trust Finder as the search-first direction.
• Keep comparison as a secondary direction.
• Preserve contextual trust as a product boundary.

Before you record the decision, do you expect this wording to keep Claude from treating status as a universal recommendation?

• Add the completed decision to the same Claude conversation by pasting this record:

Human decision:
Chosen direction: Trust Finder, a search-first catalog that places transparent trust evidence in the selection flow.
Rejected alternatives: A metadata-comprehension-only view is too narrow for the findability evidence; a comparison-first experience adds useful depth but is secondary to helping users reach a credible shortlist.
Rationale: This direction addresses the strongest cross-theme decision: finding a plausible asset and understanding whether its evidence fits the intended use.
Assumption to test: Users can interpret trust evidence quickly enough to choose what to inspect next.
Non-goal: The interface will not declare one dataset universally correct or make the decision for the user.

What Does This Decision Record Do?

• The Chosen direction field names the concept you will prototype.
• The Rejected alternatives field makes the tradeoffs visible.
• The Assumption to test field identifies what still needs evaluation.
• The Non-goal field protects the interface from automated dataset selection.

• Send the decision to add it to the project record.
• Confirm that the earlier response contains five ranked user needs.
• Confirm that the earlier response contains three how-might-we opportunities.
• Confirm that the earlier response contains one bounded problem statement.
• Confirm that the earlier response contains three distinct concepts.
• Confirm that the tradeoff matrix leaves the final selection to you.
• Confirm that the completed decision includes rejected alternatives, a rationale, an assumption, and a non-goal.

You should now be able to trace Trust Finder back to the approved synthesis. The record also confirms that status evidence supports judgment without choosing a dataset for the user.

Did the Direction Drift?

• Reject any response that turns certification into a universal recommendation.
• Reassert that the interface explains status signals without calculating universal trust.
• Get help checking the decision against your approved synthesis.

Your direction is documented with evidence, tradeoffs, and clear limits. Next, you will turn Trust Finder into an interactive first prototype.

Build the First Trust Finder Prototype

You have chosen Trust Finder through an evidence-backed human decision. That direction now needs a working interface you can test against the original trust problem.

A Claude Artifact turns the search-first direction into something you can use. This first version gives you concrete behavior to test before you decide whether the interface resolves the trust question.

In this step, get ready to:

• Generate a self-contained Trust Finder prototype with eight fictional datasets.
• Test its search, filters, result count, empty state, and dataset details.
• Evaluate whether the status badge provides enough evidence for a contextual decision.

Generate Trust Finder version 1

The prototype prompt fixes the scope before generation begins. It specifies the data, required interactions, visual direction, and boundaries that protect your human-owned decision.

• Switch back to the dedicated Claude project conversation from earlier.
• Create Trust Finder version 1 by pasting this complete prompt into the message composer:

Build a self-contained interactive Claude Artifact named "Trust Finder" for an enterprise data catalog. This is a functional concept prototype, not production software. Do not ask questions before building.

Design goals:
- Professional enterprise visual design with strong hierarchy, restrained color, clear spacing, and readable status indicators.
- Desktop-first layout that works in the Artifact preview.
- No external images, APIs, packages, storage, or network requests.
- Do not recommend one dataset as universally best.

Required interactions:
1. A working search field that filters by display name, technical name, description, and tags.
2. Working filters for domain and status, plus a clear-all action and visible result count.
3. Clicking a result opens a dataset-details panel with description, intended use, owner, freshness, expected refresh, quality-check pass rate, access note, and tags.
4. Include a useful empty state when no results match.

Important designed limitation for version 1:
Show status badges such as Certified, Verified, Experimental, and Deprecated, but do not yet add a "why this status" explanation or trust-evidence section. This opacity is intentional for the next critique step.

Use exactly this fictional sample data:

1. Customer 360 Canonical | customer_360_prod | Customer | Certified | Owner: Customer Data Products | Updated: 2 hours ago | Expected refresh: Daily at 06:00 UTC | Quality checks: 98% passing | Access: Internal, approval required | Intended use: Customer analytics and segmentation | Description: Governed customer profile with resolved identities and core lifecycle attributes | Tags: customer, profile, segmentation, canonical
2. Customer 360 Legacy | customer360_legacy | Customer | Deprecated | Owner: Data Migration | Updated: 410 days ago | Expected refresh: No longer refreshed | Quality checks: 72% passing | Access: Internal, read only | Intended use: Historical migration validation only | Description: Frozen customer extract retained for reconciliation during migration | Tags: customer, legacy, migration, archive
3. Customer Master Gold | cust_master_gold | Customer | Verified | Owner: MDM Platform | Updated: 6 hours ago | Expected refresh: Daily | Quality checks: 94% passing | Access: Restricted | Intended use: Identity resolution and master-data workflows | Description: Golden records focused on identity matching and survivorship | Tags: customer, master data, identity, gold
4. Customer Profiles Enriched | customer_profiles_enriched | Marketing | Experimental | Owner: Growth Data | Updated: 1 day ago | Expected refresh: Weekly | Quality checks: 86% passing | Access: Restricted marketing use | Intended use: Personalization experiments | Description: Customer profiles augmented with modeled interests and engagement features | Tags: customer, marketing, enrichment, experimental
5. Monthly Revenue Finance | revenue_monthly_finance | Finance | Certified | Owner: FP&A Data | Updated: 3 days ago | Expected refresh: Monthly after close | Quality checks: 99% passing | Access: Finance approved | Intended use: Official management and financial reporting | Description: Closed-period revenue aligned to the finance chart of accounts | Tags: revenue, finance, monthly, reporting
6. Monthly Revenue Sales | revenue_monthly_sales | Sales | Verified | Owner: Sales Ops Analytics | Updated: 1 day ago | Expected refresh: Daily | Quality checks: 93% passing | Access: Internal | Intended use: Pipeline and sales-performance analysis | Description: Operational revenue view aligned to sales territories and opportunities | Tags: revenue, sales, pipeline, daily
7. Orders Core | orders_core_prod | Commerce | Certified | Owner: Commerce Data | Updated: 1 hour ago | Expected refresh: Hourly | Quality checks: 97% passing | Access: Internal | Intended use: Order operations and product analytics | Description: Current order events and normalized line-item facts | Tags: orders, commerce, operations, core
8. Orders Month-End Snapshot | orders_snapshot_month_end | Finance | Verified | Owner: Financial Data Engineering | Updated: 8 days ago | Expected refresh: Monthly at close | Quality checks: 96% passing | Access: Finance approved | Intended use: Reconciliation and period-close analysis | Description: Immutable month-end snapshot of orders for financial reconciliation | Tags: orders, finance, snapshot, close

After building, give me a three-item manual test checklist outside the Artifact. Do not critique or fix the designed limitation yet.

How does this prompt control the prototype?

• The eight fixed records keep the generated experience grounded in the same fictional evidence.
• The required interactions translate the chosen search-first direction into a testable flow.
• The boundaries prevent external dependencies, new data, automated selection, and universal recommendations.
• The manual checklist gives you an initial testing aid outside the Artifact.

• Submit the prompt to Claude.
• Wait for Claude to finish creating the Artifact.
• Read the three-item manual test checklist beneath the finished Artifact.

That is your first working Trust Finder draft. Eight fictional datasets now sit inside an interactive data-catalog Artifact beside the conversation.

Artifact not rendering?

• Select Try fixing with Claude if the Artifact displays an error.
• Return to Settings from earlier. Confirm that Cloud code execution and file creation remains enabled under Capabilities.
• Ask Claude to preserve the exact eight records if an automatic repair changes the sample data.
• Help me troubleshoot my Trust Finder Artifact.

Check the filters and supporting states

Manual testing confirms that the interface changes state when you use its controls. Start with the supporting interactions so the final trust test can focus on the product decision.

• Check the initial result count in the Trust Finder Artifact.
• Open the domain filter.
• Select Finance.

You should see two results: Monthly Revenue Finance and Orders Month-End Snapshot.

• Use the clear-all action.

The visible result count should return to eight. This confirms that the clear-all action restores the full catalog.

• Select the search field.
• Enter supplier.

You should see the empty state because none of the eight records match that query.

• Use the clear-all action again.

You should see all eight datasets again. Search, domain filtering, empty-state handling, and clearing now have visible evidence behind them.

Filters producing unexpected results?

• Check that the Artifact contains exactly the eight datasets from the prompt.
• Ask Claude to repair the existing Artifact if clearing leaves a search term or filter active.
• Help me diagnose the Trust Finder filters.

Test the core trust decision

The final flow tests the interaction at the center of your chosen direction. Before you begin, predict whether the details panel can justify the status badge and its fit for a marketing use case.

• Select the search field.
• Enter customer.

You should see four matching customer datasets in the results. The visible count confirms that the search recognizes the term across the records.

• Open the status filter.
• Select Certified.
• Select Customer 360 Canonical from the filtered results.

You should see one filtered result with its details open. The panel shows its description, intended use, owner, last updated value, expected refresh, quality-check pass rate, access note, and tags.

• Clear the status filter.
• Compare the visible names of Customer 360 Canonical, Customer 360 Legacy, and Customer Master Gold.
• Open Customer 360 Canonical again.

The similar names make the selection problem visible. The surrounding metadata helps you inspect the record after opening it.

• Inspect the panel for evidence explaining why the asset is certified.
• Check whether the displayed information establishes that certification fits a marketing use case.

The designed gap is visible

The badge states that the asset is Certified. The panel does not expose the evidence behind that status.

The interface also cannot establish whether certification makes the asset appropriate for marketing. This is the intended shortfall in version 1.

You now have a functional Trust Finder prototype with a specific UX weakness you observed firsthand. Next, you will critique that weakness and apply only the changes that survive human review.

Refine Trust Evidence and Handoff

Your first Trust Finder Claude Artifact proves that the search-first direction works as an interactive experience. Its opaque status badges also expose the unresolved question of why an asset earned its status.

An AI critique can widen the review by spotting gaps across several tasks. Your judgment must decide which recommendations fit the fictional evidence before Claude changes the prototype.

In this step, get ready to:

• Critique the current Artifact without allowing automatic edits.
• Triage the findings through explicit human decisions.
• Refine Trust Finder and prepare a learner-edited handoff.

Request and triage a skeptical critique

The current trust gap gives the critique a precise target. Claude should evaluate the experience against the approved synthesis without rewriting your strategy or changing the Artifact.

• Request five evidence-linked findings by pasting this prompt into the message field in your project conversation:

Act as a skeptical enterprise UX evaluator. Review the current Trust Finder Artifact and the approved synthesis in this conversation. Do not modify the Artifact yet.

Evaluate these tasks:
- Find a trustworthy customer dataset for segmentation.
- Distinguish three similarly named customer assets.
- Understand whether Certified or Verified is relevant to the intended use.
- Identify ownership, freshness, expected refresh, and access constraints before requesting access.

Return exactly five findings in a table with: task, observed issue, evidence from the Artifact or feedback IDs, severity from 1 to 3, proposed change, and risk of applying that change blindly.

Then add:
- One thing the prototype gets right.
- One recommendation you are deliberately not making because the evidence is insufficient.
- One question that requires actual usability testing rather than another AI opinion.

Do not rewrite the problem statement, invent user behavior, or modify the prototype.

What This Critique Protects

• The task list keeps the review focused on decisions that Trust Finder should support.
• The evidence column forces every concern to point to the Artifact or the approved feedback.
• The risk column exposes how a plausible recommendation could distort the product direction.
• The final constraints preserve your problem statement and the tested Artifact.

• Submit the critique prompt.

You should receive exactly five findings in a table. The response should also identify one strength, one unsupported recommendation, and one question for real usability testing.

Did Claude Change the Artifact?

• Tell Claude to restore the tested first version before you continue if it changed the Artifact.
• Keep the critique as analysis until your triage record authorizes specific changes.
• Ask for help with an unexpected critique response.

A critique becomes a design decision only after you evaluate its evidence and consequences. Your triage record makes that accountability visible in the conversation.

• Review the evidence cited for each finding.
• Mark transparent status evidence as Accept.
• Mark stronger differentiation among similarly named assets as Accept.
• Mark any automated universal dataset choice as Reject or Constrain.
• Choose the least certain remaining finding for Defer.
• Record the research needed to resolve the deferred finding.
• Submit the completed triage record in the same conversation.

That is the control point established. Your conversation now shows which recommendations the evidence supports and which decisions remain yours.

Apply only the accepted changes

The accepted findings address the visible trust gap without expanding the product into automated decision-making. The next prompt keeps every tested interaction while adding evidence and neutral differentiation.

• Request the approved refinement by pasting this prompt into the same conversation:

Update the existing Trust Finder Artifact. Preserve its sample data, visual language, working search, filters, result count, empty state, and dataset-details interaction. Apply only these accepted changes:

1. Add an expandable "Why this status?" section inside dataset details. Explain that status is one signal, not a universal recommendation. Show fictional evidence appropriate to each status, using the existing owner, expected refresh, quality-check rate, intended use, and access fields. For Certified assets, include a fictional governance-review note. For Deprecated assets, clearly state that refresh has stopped and use is limited.
2. Add a "Similar assets" area in customer dataset details. Show the closest two customer assets and summarize the most decision-relevant differences: intended use, status, refresh pattern, owner, and access. Do not rank them or label one as best.
3. Strengthen scanability for owner, last updated, expected refresh, and access. Keep last updated distinct from expected refresh so users cannot confuse recency with cadence.
4. Add a small note near status filters: "Status supports judgment; it does not replace fit-for-use review."

Do not add dashboards, recommendations, scores, AI chat, external dependencies, or new data. After updating, summarize what changed and what remains intentionally unresolved.

What This Refinement Changes

• The expandable explanation connects each status to evidence already present in its fictional record.
• The Similar assets area reduces the effort required to distinguish customer datasets with overlapping names.
• The revised metadata hierarchy separates recent activity from the expected refresh cadence.
• The guardrails preserve contextual judgment by excluding scores and automated recommendations.

• Submit the refinement prompt.

Claude should update the existing Artifact beside the conversation. Its summary should separate completed changes from questions that still need research.

Before you retest, form a prediction about whether the refined panel can explain status without choosing a dataset for you.

• Enter customer in the Artifact search field.
• Select Certified in the status filter.
• Open Customer 360 Canonical from the filtered results.
• Expand Why this status? in the details panel.

What Should You See?

The expanded section should explain the status through owner, refresh cadence, quality checks, intended use, access, and a fictional governance review. It should also state that status remains one input to a fit-for-use decision.

The details panel should still show the existing description and metadata. Search plus the Certified filter should continue to narrow the results.

• Inspect the Similar assets area for the closest two customer records.
• Compare the intended-use differences shown for those records.
• Check that last updated remains distinct from expected refresh.
• Scan the details panel for any score or universal recommendation.
• Use the clear-all action.

You should see neutral differences across intended use, status, refresh pattern, owner, and access. You should not see a winner selected for you.

After clearing the filters, the result count should return to all eight fictional datasets. The original search, filters, empty state, and details interaction should remain available.

Did a Core Interaction Break?

• Tell Claude which tested interaction changed after the refinement.
• Ask Claude to restore that original behavior while preserving only the accepted trust-evidence changes.
• Get help diagnosing a broken Trust Finder interaction.

Create and rehearse the handoff

A polished prototype does not show who remained accountable for the design choices. The handoff should separate AI speed from human judgment while naming every claim that still needs evidence.

• Generate the handoff note and timed demo outline by pasting this prompt into the same conversation:

Create a concise handoff note for experienced UX and product designers based only on this conversation.

Output:
1. "AI accelerated": five bullets covering pattern grouping, structured framing, concept breadth, prototype production, and critique breadth.
2. "Human judgment remained essential": five bullets covering evidence validity, theme correction, priority, concept selection, and critique triage.
3. "Still unvalidated": three bullets naming claims that require research or usability testing.
4. A 3-minute demo outline with timestamps. It must show one search, one filter, one dataset-details interaction, the before-and-after trust explanation, and one explicit example of rejecting or constraining AI output.

Do not claim that fictional feedback validates real user prevalence or that this prototype is production ready.

What This Handoff Captures

• The acceleration section records where AI reduced production time.
• The accountability section records decisions that required human ownership.
• The unvalidated section prevents fictional feedback from becoming a claim about real users.
• The timed outline turns the process into a concise design narrative.

• Submit the handoff prompt.
• Review the five bullets under the AI acceleration section.
• Review the five bullets under the human judgment section.
• Check that the unvalidated section contains three claims.
• Check that the demo outline covers the full three-minute sequence.

The generated structure gives you a presentation baseline. Your edit turns it into a handoff you can defend in your own voice.

• Choose one demo line that does not sound like you.
• Rewrite that line in your own voice.
• Send the revised line as a follow-up in the conversation.
• Ask Claude to incorporate your revised line into the handoff note.

Is the Handoff Overclaiming?

• Remove any statement that treats fictional feedback as proof of real user prevalence.
• Remove any statement that presents the Artifact as production ready.
• Ask for help tightening an unsupported handoff claim.

Before your rehearsal, predict whether the story makes your design accountability visible within three minutes.

• Start a three-minute timer on your phone.
• Rehearse the edited demo outline against its timestamps.
• Show one customer search during the interaction segment.
• Apply the Certified filter during the interaction segment.
• Open Customer 360 Canonical during the details segment.
• Expand the trust explanation during the before-and-after segment.
• State one critique recommendation that you rejected or constrained.

You should reach the closing lesson close to the three-minute mark. Your demo should show that AI accelerated the workflow while evidence decisions and product boundaries remained human responsibilities.

Your Final Handoff

• The conversation contains exactly five evidence-linked critique findings.
• The human triage record contains accepted, rejected or constrained, and deferred recommendations.
• The refined Artifact explains status evidence while preserving all eight datasets and tested interactions.
• The learner-edited handoff contains three unvalidated claims plus a timed demo outline.

Strong finish. Trust Finder now exposes the evidence behind its status signals while leaving the dataset decision with the user.

Secret mission

Add Evidence-Based Dataset Comparison

Extend Trust Finder with a neutral comparison view for two similar datasets. Users will see aligned evidence and meaningful differences without receiving an automatic winner.

Clean Up Your Resources

Clean Up Your Resources

Claude Free keeps this project at $0, subject to your account's current usage limits. Your work lives in one source conversation, so you can keep it, close Safari for now, or remove it.

Resources you used:

• The dedicated Claude project conversation containing your research trail, human decision records, critique findings, handoff note, demo outline, comparison evidence, Artifact Readiness Check, and refined Trust Finder Artifact.

Keep everything running

No action is needed. Choose this option if you want to keep refining Trust Finder or revisit your design decisions.

• Leave the dedicated project conversation in your Claude account.
• Return to the same source conversation when you want to rerun the comparison tests.
• Continue refining the Trust Finder Artifact from its preserved context.

Pause - I'll come back to this later

Pausing preserves every prompt, decision record, and Artifact in the source conversation. There are no running cloud resources to stop.

• Close the Claude project tab in Safari.
• Close Safari if you have finished browsing.
• Return later through the same Claude account.
• Select the dedicated source conversation to resume with its full context.

Delete - I don't want to use this again

Deleting the project conversation is a bigger step than simply closing Safari. Claude shows an in-product confirmation, so pause there to check what will be removed.

• Return to Claude's conversation list.
• Locate the dedicated conversation containing Trust Finder.
• Open the conversation's current controls.
• Choose the control that removes the source conversation.
• Read the in-product confirmation to understand what Claude will remove.
• Confirm the removal.
• Return to the conversation list after Claude completes the removal.

The dedicated Trust Finder conversation should no longer appear in your conversation list. That confirms the project work has been removed.

Nice Work!

Nice Work!

You made it! You transformed fictional feedback into Trust Finder, a working Claude Artifact shaped by human judgment.

You've learned how to:

• Create an evidence-backed synthesis from 10 fictional feedback fragments. Preserve source traceability and uncertainty during human review.
• Turn approved themes into prioritized needs, a bounded problem statement, and three distinct concepts. Select Trust Finder through an explicit tradeoff decision with a testable assumption and non-goal.
• Build and test an interactive Trust Finder prototype with search, filters, dataset details, transparent trust evidence, and similar-asset cues. Use a human-owned critique log plus an edited demo outline to separate AI acceleration from design accountability.
• Secret Mission: Add a neutral two-dataset comparison with aligned evidence and metadata. Keep the final choice with the user by avoiding scores, rankings, recommendations, and automatic selection.

Ready to quiz yourself?