Build Trustworthy AI Feedback Triage

Introduction

30 Second Summary

A polished summary of product feedback can still hide the comments behind its conclusions. Without visible evidence or a reviewer’s judgment, the result is hard to trust.

In this project, you will use Figma Make to build Signal Review, a functional desktop prototype for reviewing themes from fictional product feedback. You will transform an unverifiable first draft into a review flow where every conclusion can be checked, corrected, approved, or escalated.

What You'll Build

You will demo Signal Review from a shareable Figma URL, where a reviewer can trace each generated theme back to fictional comments before deciding its outcome.

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

• A functional feedback triage flow where eight fictional comments become four simulated themes after you click Analyze 8 comments.
• A traceable evidence view that reveals the exact comments behind each theme. Relative confidence cues make uncertainty visible without claiming false precision.
• A set of human review controls that lets a reviewer edit generated conclusions, approve useful themes, or route weak output to a review queue.
• Secret Mission: Add an ambiguous fictional comment that the system refuses to group until a reviewer assigns it to a theme or leaves it ungrouped.

Are there any prerequisites?

You need occasional prompting experience plus a paid Figma Full seat. The project uses included AI credits and requires no additional purchase.

Before We Start

Before any hands-on work begins, this is your moment to commit to the product judgment you want to practise. You are building a feedback triage prototype to constrain, critique, and supervise AI-generated output instead of accepting it at face value.

Create a Signal Review Starter

A full feedback flow would spend AI credits before you know whether the tool is ready. A tiny working screen gives you a fast readiness check.

In this step, you will create a new Figma Make file. You will generate a minimal Signal Review starter with one responsive button.

In this step, get ready to:

• Create a new Figma Make file in Drafts.
• Generate the Signal Review Lab starter from a focused prompt.
• Test the Start prototype button in the interactive preview.

Create the Figma Make file

A Figma Make file keeps your prompt conversation beside the interactive preview it generates. This gives the prototype a dedicated place to grow across the project.

• Switch back to Figma through the browser or desktop access you used earlier.
• Complete Figma's sign-in flow with the account that has your paid Figma Full seat.
• Select Drafts in the Figma file browser.
• Click Make in the upper-right corner.

You'll see a new Figma Make file with an AI chat area and an interactive preview. The file is ready for its first prompt.

Can't See Make?

Confirm that you signed in with the account connected to your paid Figma Full seat. Figma Make is available to Full seats on paid plans.

If the option is still missing, help me check why Figma Make is unavailable for my account. You can also share what you see with the NextWork community.

Generate the starter screen

A structured prompt tells Figma Make what to create. Its constraints also prevent unrelated features from expanding the scope.

• Paste the following setup prompt into the AI chat:

Create a minimal desktop interface titled "Signal Review Lab". Show the subtitle "Trustworthy AI feedback triage prototype" and one button labeled "Start prototype". Make the button respond visually when clicked. Do not add a backend, external API, login, image, or extra page.

What Does This Prompt Define?

• The title gives the starter screen a clear identity.
• The subtitle states the trust-focused purpose of the prototype.
• The button provides a small interaction that proves the preview works.
• The final constraints keep the result limited to one local prototype screen.

• Click Send in the AI chat.
• Wait for Figma Make to finish generating the interactive preview.

You'll see a generation summary once the prompt completes. The interactive preview will display the new starter screen.

Preview Missing or Off-Scope?

Check that the complete prompt reached the AI chat before you clicked Send. A missing sentence can remove the interaction or its constraints.

If the preview still does not match, help me compare my Figma Make preview with the Signal Review starter prompt. You can also show the mismatch to the NextWork community.

✔️ Awesome, I've got everything!

Your Figma Make file now contains the minimal Signal Review starter. Keep the file open for the interaction check.

ⓧ I'd like to double check the full code

Compare the prompt in your AI chat with this complete version.

Create a minimal desktop interface titled "Signal Review Lab". Show the subtitle "Trustworthy AI feedback triage prototype" and one button labeled "Start prototype". Make the button respond visually when clicked. Do not add a backend, external API, login, image, or extra page.

What Should Match?

Every sentence should appear in the same order. The final sentence is essential because it keeps the starter free from backends, external APIs, login flows, images, and extra pages.

Test the starter interaction

A visible button response proves that Figma Make produced an interactive prototype. The check also confirms that the starter stayed on one focused screen.

Before you click the button, what kind of visual response do you expect the preview to show?

• Click Start prototype in the interactive preview.

You'll see the button respond visually. You'll remain on the Signal Review Lab screen.

• Confirm that the screen title reads Signal Review Lab.
• Confirm that the subtitle reads Trustworthy AI feedback triage prototype.
• Confirm that the screen contains one button labeled Start prototype.
• Confirm that clicking the button does not open an extra page.

That first interaction is working. Your Figma Make file is ready to become the Signal Review feedback triage experience.

Button Not Responding?

Confirm that you are clicking the button inside the interactive preview. If the button remains static, send the original prompt again so Figma Make can restore the requested visual response.

For more help, help me fix a Start prototype button that does not respond in Figma Make. You can also share the preview behavior with the NextWork community.

Your starter screen now proves that the prototype environment can generate a focused interaction. Next, you will turn it into the first Signal Review experience.

Generate the Untrustworthy Draft

Your starter proved that Figma Make can produce a working interaction. The same file now needs enough detail to resemble a product manager's feedback workflow.

A polished summary can feel credible before it has earned trust. This step creates that first impression so you can test what a reviewer can actually verify.

In this step, get ready to:

• Replace the starter with a structured product brief.
• Generate the simulated grouping flow from eight fictional comments.
• Test whether the Performance result can be verified.

Replace the starter with the product brief

Structured prompting gives Figma Make a clear job before it generates the interface. This brief also fixes the fictional data so every learner evaluates the same scenario.

Sending this prompt uses credits included with your existing paid seat. You do not need to buy more credits for this step.

• Switch back to the AI chat in the same Figma Make file.
• Paste this replacement prompt into the AI chat:

Replace the starter with a desktop functional prototype called "Signal Review" for a product manager reviewing fictional product feedback.

Always show these labels:
- "Fictional test data"
- "Simulated AI output"

Show these eight feedback comments in an input list:
FB-101: The dashboard is slow every Monday morning.
FB-102: Reports take too long to load when the date range is more than a month.
FB-103: Can I export reports as CSV?
FB-104: I need a PDF export for client meetings.
FB-105: The filter labels are hard to understand.
FB-106: I can never find my saved filters.
FB-107: Dark mode would help when I work late.
FB-108: Please add more chart colors.

Add a primary button labeled "Analyze 8 comments". When clicked, show a short loading state and then display these four simulated groupings:
- Performance: FB-101 and FB-102
- Export options: FB-103 and FB-104
- Filters and findability: FB-105 and FB-106
- Visual preferences: FB-107 and FB-108

For this first version, each result card must show only the theme title, a one-sentence summary, and the number of comments. Do not show source comments, source IDs, confidence, uncertainty, explanations, edit controls, approval controls, or review controls.

Use local hardcoded prototype data only. Do not use a backend, external API, real customer data, generated images, or additional packages. Keep the app desktop and full-screen friendly. Use labeled controls, readable text, and visible keyboard focus states.

How is this prompt structured?

• The opening defines the product manager's task.
• The comment list fixes the fictional input data.
• The grouping list fixes the simulated output.
• The local-only constraints prevent integrations.
• The card rule limits each result to three visible fields.

Figma Make needs a moment to rebuild the interactive preview. The preview updates when the prompt finishes.

• Click Send in the AI chat.
• Wait for the generation to finish.

You should see a desktop interface titled Signal Review. The input area should contain all eight comments from FB-101 through FB-108.

The labels Fictional test data plus Simulated AI output should remain visible. The primary action should read Analyze 8 comments.

Preview not updating?

• Confirm that the full replacement prompt appears in the AI chat.
• Check that Send completed before trying again.
• Keep the Figma Make file open until the generation finishes.

Figma Make did not update my preview after I sent the replacement prompt. Help me troubleshoot without changing the project requirements.

Run the limited analysis

The generated screen is now a testable draft. Its analysis button converts the eight hardcoded comments into a fixed set of simulated groupings.

Before you click it, ask yourself this: will four polished cards be enough to verify the Performance conclusion? Keep your prediction in mind.

• Click Analyze 8 comments in the interactive preview.

You should briefly see a loading state. The results should settle into four cards.

The card titles should be Performance, Export options, Filters and findability, plus Visual preferences. Each card should show a one-sentence summary plus a count of 2 comments.

No card should show source comments, source IDs, confidence, uncertainty, or explanations. You should also see no edit, approval, or review controls.

• Press Tab to move keyboard focus through the controls.

You should see a visible focus indicator move between the labeled controls.

✔️ I see four limited cards

You have the intended first draft: the loading state resolves into four simulated themes. Every card stays limited to a title, summary, and comment count.

ⓧ I see extra trust controls

Figma Make can add features that sound useful even when the prompt excludes them. This correction restores the intended evaluation draft.

• Remove the extra controls by pasting this correction into the AI chat:

For this evaluation draft, remove every source, confidence, uncertainty, edit, approval, and review control. Keep only title, summary, and comment count on each theme card.

What does this correction change?

The correction narrows each theme card to the three required fields. It preserves the existing data plus the four groupings.

• Click Send in the AI chat.
• Wait for the preview to update.
• Click Analyze 8 comments again.

You should now see four cards containing only a title, one-sentence summary, and count of 2 comments.

Still seeing extra controls?

• Confirm that the correction appears as a separate message in the AI chat.
• Check the updated preview after the generation finishes.

Help me remove extra controls from my Figma Make theme cards while preserving the four required groupings.

Test the trust gap

The Performance card presents a concise conclusion about two comments. Test whether the card gives you enough evidence to validate that conclusion.

• Locate the Performance card in the results.
• Read its one-sentence summary.
• Try to inspect which two comments support the summary using only the result card.

The card does not identify FB-101 or FB-102 as its sources. The interface also provides no uncertainty cue or review path.

That shortfall is intentional. You have found the designed trust gap in the first draft.

Why is this a trust gap?

Source traceability links a generated conclusion to the exact inputs behind it. The input list contains comments, but the result card never states which ones support its conclusion.

This prototype uses hardcoded simulated output. The exercise evaluates interface behavior. It does not establish model accuracy.

You found the exact weakness this draft was designed to reveal: its polished Performance summary cannot be traced to evidence. Next, you will make each grouping inspectable and communicate uncertainty without false precision.

Expose Evidence and Uncertainty

Your first Signal Review draft now turns eight fictional comments into four polished themes. Because their evidence and uncertainty are hidden, unsupported conclusions look final.

You'll use Figma Make annotations plus a structured follow-up prompt to add source traceability and uncertainty cues. This gives reviewers evidence before they use a simulated theme.

In this step, get ready to:

• Annotate the trust gap on an existing theme card.
• Add exact source comments plus relative confidence cues.
• Verify that all eight fictional comments remain traceable and unchanged.

Annotate the trust gap

An annotation connects your critique to the exact part of the interface that needs attention. This edit uses included Figma AI credits, so the focused message below helps avoid unnecessary retries.

• Click Annotate for agent above the interactive preview.
• Place an annotation on any existing theme card.
• Enter the annotation by pasting this message:

This polished summary cannot be verified. Add access to exact source comments and a relative confidence cue without changing the four groupings.

Why start with an annotation?

The annotation ties the trust critique to a specific theme card. It identifies the polished summary as the part that needs supporting evidence.

The message also preserves the four existing groupings. Your next prompt can improve transparency without changing the simulated analysis.

• Click Send to agent.
• Wait for the annotation to appear in the AI chat.

You should see the annotation attached to the selected card in the agent context.

Annotation missing from the chat?

• Select the same theme card in the interactive preview.
• Place the annotation again.
• Click Send to agent again.

Help me check why the annotation was not sent.

Add evidence and uncertainty

Relative confidence labels communicate which simulated groupings appear stronger within this fixed prototype. A plain-language limitation keeps those labels from implying measured model accuracy.

• Add the implementation details to the AI chat by pasting this prompt:

Keep the existing layout, data, and groupings. Add a visible banner above the results that says: "AI-generated grouping. Verify source comments before using this output."

On every theme card:
- Label the summary "AI-generated draft".
- Add a relative confidence label. Use High for Performance and Export options. Use Medium for Filters and findability and Visual preferences.
- Beside confidence, show the helper text "Illustrative, not model-calibrated."
- Add a control labeled "View 2 source comments" that expands to show the exact source IDs and verbatim feedback already stored for that theme.

Never invent, rewrite, or paraphrase source comments. Do not show numeric confidence percentages.

What does this prompt change?

• The warning banner asks reviewers to verify evidence before using a grouping.
• The AI-generated draft label identifies every summary as generated output awaiting review.
• The High and Medium labels communicate relative uncertainty without numeric precision.
• The source controls preserve the exact fictional comments behind every theme.

• Click Apply in the AI chat.
• Wait for the interactive preview to update.
• Run the revised prototype by clicking Analyze 8 comments.

What should I see?

You should see AI-generated grouping. Verify source comments before using this output. above the results.

Every theme card should label its summary AI-generated draft. Each card should also show a relative confidence cue.

Every card should include Illustrative, not model-calibrated. beside confidence. You should also see a View 2 source comments control.

🙋‍♀️ Missing a trust cue?

• Check every card for the exact summary label, confidence helper text, and source control.
• Return to the AI chat if a required element is missing.
• Apply the same implementation prompt again without changing its constraints.

Help me correct the missing evidence or uncertainty cue.

Verify every source trail

Traceability works when a reviewer can move from each summary to the exact comments that support it. This check confirms that the new interface preserves every source without rewriting it.

Before you rerun the analysis, do you expect every theme to reveal its original comments? Hold that prediction for the check.

• Rerun the simulated analysis by clicking Analyze 8 comments.

After the loading state, you should see exactly four existing grouping cards. The warning banner should sit above them.

• Expand Performance by clicking View 2 source comments on its card.

You should see FB-101: The dashboard is slow every Monday morning. It should also show FB-102: Reports take too long to load when the date range is more than a month.

The Performance card should show High beside Illustrative, not model-calibrated.

• Expand the remaining three cards one at a time with their View 2 source comments controls.

What should the other cards show?

Export options should show FB-103: Can I export reports as CSV? It should also show FB-104: I need a PDF export for client meetings.

Filters and findability should show FB-105: The filter labels are hard to understand. It should also show FB-106: I can never find my saved filters.

Visual preferences should show FB-107: Dark mode would help when I work late. It should also show FB-108: Please add more chart colors.

• Inspect the relative confidence label on every card.

Performance should show High. Export options should show High.

Filters and findability should show Medium. Visual preferences should show Medium.

• Scan the expanded cards for numeric confidence percentages or altered source comments.

You should find no numeric confidence percentages. Every expanded comment should match the fictional input list exactly.

That closes the trust gap: every simulated theme now exposes its evidence plus the limits of its confidence cue. Next, you'll give the reviewer direct control over generated conclusions.

Add Human Review Controls

Your Figma Make prototype already exposes the evidence behind each simulated theme. Every generated draft is now open to inspection.

Evidence alone does not create accountability. Human review gives someone authority over each generated conclusion.

The reviewer needs a correction path. Separate controls must cover approval plus escalation.

In this step, get ready to:

• Add a reviewer workflow to every theme card.
• Exercise approval, editing, and escalation across separate cards.
• Reset the workflow before recreating the target demo state.

Add review controls to every theme

A default status creates a shared baseline for review. Local state keeps the interaction repeatable during the demonstration.

• Paste the following prompt into the AI chat from earlier:

Add human review controls to every theme card without changing the source data or groupings.

Each card starts with status "Not reviewed" and has three controls:
- "Approve"
- "Edit theme"
- "Send to review"

Behavior:
- Approve changes the status to "Approved by reviewer".
- Edit theme opens editable title and summary fields with Save and Cancel. Saving updates only that card.
- Send to review changes the status to "Needs review" and increases a visible review queue count in the page header.
- Add a secondary control labeled "Reset review state" that restores the original titles, summaries, statuses, and queue count.

Keep all actions local to the current prototype session. Do not add login, persistence, a backend, or an external API.

What does this prompt add?

• The Not reviewed status gives every card a shared starting point.
• The Approve control records a reviewer decision.
• The Edit theme control limits corrections to one card.
• The Send to review control creates a visible escalation path.
• The Reset review state control restores a clean testing baseline.
• Local session state keeps the demo repeatable without a backend.

• Click Send in the AI chat.

A short pause while the preview rebuilds is expected.

• Wait for the interactive preview to update.

The updated preview is ready for a fresh analysis.

• Click Analyze 8 comments in the preview.

You'll see every theme card start with Not reviewed. Each card also shows the three reviewer controls.

Good progress. The prototype now makes review decisions visible at the card level.

Missing a review control?

• Check every result card against the required controls in the prompt.
• Describe the missing control in the same AI chat if the preview omits one.
• Wait for the preview to rebuild before running the analysis again.

Help me fix a missing or unresponsive review control.

Exercise all three review paths

Each control captures a different reviewer decision. Testing them on separate cards proves that one action does not overwrite another card.

• Click Approve on the Performance card.

You'll see the Performance status change to Approved by reviewer.

• Click Edit theme on the Export options card.

You'll see editable title plus summary fields. You'll also see Save plus Cancel controls.

• Replace the title with Reporting exports.
• Click Save.

You'll see Reporting exports saved on that card. The other card titles remain unchanged.

• Click Send to review on the Visual preferences card.

You'll see the card status change to Needs review. The header review queue count increases to 1.

The reset control provides a clean baseline for repeatable task testing.

• Click Reset review state.

You'll see the original Export options title return. Every card returns to Not reviewed.

The review queue count returns to 0. Your testing baseline is clean again.

Recreate the target review state

A deliberate final state makes the reviewer workflow easy to demonstrate. Each changed card should communicate one distinct human decision.

• Click Edit theme on the Export options card.
• Replace the title with Reporting exports.
• Click Save.
• Click Approve on the Performance card.
• Click Send to review on the Visual preferences card.

Before you inspect the final state, consider this question: what statuses plus count should your actions produce.

• Inspect the three changed cards in the interactive preview.
• Inspect the review queue count in the page header.

You'll see Performance marked Approved by reviewer. You'll see Reporting exports saved as the renamed card.

You'll see Visual preferences marked Needs review. You'll see the review queue count at 1.

• Click View 2 source comments on the Reporting exports card.

You'll still see FB-103 plus FB-104 with their original wording. Renaming the theme has not changed its evidence.

Your prototype now supports correction, approval, escalation, and reset without hiding the source evidence. Next, you'll test the complete reviewer task before sharing the experience.

Test and Publish the Prototype

Your prototype now exposes the evidence behind each generated theme. It also gives a reviewer clear control over each conclusion.

A generated interface remains a draft until its core task works from start to finish. Another person must also be able to open it.

In this step, you will use task-based evaluation to test the full Figma Make flow. You will then publish the prototype through a shareable URL.

In this step, get ready to:

• Complete the core review task in the fullscreen preview.
• Publish the prototype or share a permission-controlled preview.
• Open the saved URL to retest the complete reviewer flow.

Test the complete review flow

A task script gives every interaction a concrete success condition. This one checks the reviewer journey from analysis through escalation.

• Switch back to the Figma Make file from earlier.
• Open the fullscreen prototype by clicking Open preview in a new tab in the upper-right corner.
• If the new tab retains the previous review state, restore the original cards by clicking Reset review state.
• Run the simulated analysis by clicking Analyze 8 comments.

You will see a brief loading state followed by four theme cards. You will also see the warning AI-generated grouping. Verify source comments before using this output.

• Expand the Performance evidence by clicking View 2 source comments on that card.

Under Performance, you will see FB-101: The dashboard is slow every Monday morning..

You will also see FB-102: Reports take too long to load when the date range is more than a month..

• Open the Export options editing fields by clicking Edit theme on that card.
• Replace the title with Reporting exports.
• Save the edited card by clicking Save.

The card title will now show Reporting exports. The other three theme titles will remain unchanged.

• Approve the Performance card by clicking Approve.
• Route the Visual preferences card to the queue by clicking Send to review.

Performance will show Approved by reviewer. Visual preferences will show Needs review.

The review queue count will show 1. You have now tested all three reviewer decisions in one flow.

Is a review control missing?

• Return to the Figma Make file from earlier.
• Wait for the interactive preview to finish loading.
• Reopen the fullscreen preview through Open preview in a new tab.

If the same control still does not respond, help me troubleshoot the broken review interaction.

Publish or share the prototype

Publishing makes your prototype available on the public web when your workspace permits it. Your planned version contains only fictional test data.

Public sharing deserves one careful check before you continue. The prototype must stay free of customer feedback or personal data.

• Return to the Figma Make file from earlier.
• Confirm the prototype still displays Fictional test data.
• Confirm the prototype still displays Simulated AI output.
• Confirm the file contains no real customer feedback.
• Confirm the file contains no personal data.
• Confirm the file contains no API keys.
• Confirm the file contains no content that you lack permission to publish.
• Begin the publishing workflow by clicking Publish.

Which publishing result do you see?

✔️ Publishing is available

• Set the prototype title to Signal Review: Trustworthy AI Feedback Triage in the Publish modal.
• Publish the prototype by clicking Publish again.

Figma will generate a public URL for the published prototype.

• Record the generated URL here: your shareable Figma URL.

ⓧ Workspace policy blocks publishing

Some workspaces restrict public publishing. You can still share the fullscreen preview with permission controls.

• Return to the Figma Make file from earlier.
• Open the fullscreen prototype by clicking Open preview in a new tab.
• Open the share modal by clicking Share.
• Set appropriate permissions for the person reviewing your prototype.
• Copy the preview link from the share modal.
• Record the copied URL here: your shareable Figma URL.

Retest the shared prototype

A saved link proves that another person can reach the prototype. Repeating the core task proves that the shared version preserves its interactions.

Before you open the saved link, do you expect the same evidence and reviewer controls to work outside the Make file?

• Open a new browser tab.
• Paste your shareable Figma URL into the browser address bar.
• Load the shared prototype by pressing Enter.

You will see Fictional test data near the comments. You will also see Simulated AI output on the shared prototype.

• If the shared tab retains a previous review state, restore the original cards by clicking Reset review state.
• Run the shared analysis by clicking Analyze 8 comments.
• Expand the Performance evidence by clicking View 2 source comments.

You will see the unchanged FB-101 source comment. You will also see the unchanged FB-102 source comment.

• Open the Export options editing fields by clicking Edit theme.
• Replace the title with Reporting exports.
• Commit the title change by clicking Save.
• Approve the Performance card by clicking Approve.
• Route Visual preferences to the queue by clicking Send to review.

You will see Reporting exports remain saved. Performance will show Approved by reviewer.

Visual preferences will show Needs review. The header review queue count will show 1.

That is the handoff complete. Your shared Signal Review prototype now supports the complete reviewer flow for another person.

Secret mission

Route Ambiguity to Review

Add one ambiguous fictional comment that could support conflicting interpretations. The prototype will pause the simulated analysis so a reviewer can decide where the comment belongs.

Clean Up Your Resources

Clean Up Your Resources

This project added no separately billed resource beyond your existing subscription. Choose the cleanup option that matches what you want next.

Cost warning

Every prompt you send in Figma Make consumes included AI credits. Usage varies by model, task complexity, and context volume.

Cleanup does not require buying more credits. If your included allowance is exhausted, avoid sending more prompts.

Resources you used:

• The persistent Figma Make file titled Signal Review: Trustworthy AI Feedback Triage.
• The published Figma URL or permission-controlled preview link for the prototype.

Keep everything running

No action needed. Choose this if you want the prototype to remain available for future edits or demonstrations.

• Leave the published or permission-controlled URL available.
• Keep the Figma Make file in Drafts for future edits.
• Continue using only fictional data in the prototype.

Pause - I'll come back to this later

This option pauses public access while preserving the Make file for future edits.

• Switch back to the Figma Make file if you published a public URL.
• Click Make settings in the upper-right corner for the public version.
• Click Unpublish for the public version.
• Keep the existing restrictions in place if you use a permission-controlled preview link.

Your Make file remains available for future edits after the public version is unpublished.

Delete - I don't want to use this again

This option ends access to the shared prototype. It also moves the Figma Make file to trash.

• Switch back to the Figma Make file if you published a public URL.
• Click Make settings in the upper-right corner for the public version.
• Click Unpublish for the public version.

The permission-controlled preview path has no published version in this project. File removal is the remaining cleanup for that path.

• Return to the Figma file browser.
• Open Drafts.
• Right-click the file titled Signal Review: Trustworthy AI Feedback Triage.
• Select Move to trash.
• Select Move file to trash to confirm.

You should no longer see the file in Drafts.

Nice Work!

Nice Work!

You did it. Your published Figma Make prototype now turns fictional feedback into traceable drafts under reviewer control.

Here's what you learned:

• Built Signal Review as a functional prototype. Used structured prompting to constrain it to fixed fictional data.
• Exposed a deliberate trust gap in polished AI output. Added source traceability to every theme. Added uncertainty cues with an illustrative limitation.
• Added human review controls to each generated theme. Used task-based evaluation to verify the complete workflow from a shareable Figma URL.
• Completed the optional Secret Mission by designing an abstain-and-review path for FB-109. Kept the ambiguous source comment verbatim after each reviewer decision.

Ready to quiz yourself?