Figma Make: From AI Generator to Trusted AI Product Development OS

I’ve been using Figma Make extensively and thinking about what the next evolution of AI-powered product development could look like.

First, I want to acknowledge that Figma has already made significant progress with Make, AI models, skills, design-system context, editing, code workflows and agent capabilities.

My feedback is therefore not simply “add more AI features.”

I believe the next major challenge is trust, control, verification and orchestration.

Today, the basic experience is often:

Prompt → AI generates → Designer reviews → Prompt again → Iterate

I believe the next evolution could be:

Understand → Plan → Confirm → Design → Build → Verify → Test → UAT → Approve → Release

1. AI Planning & Understanding Gate

Before Make starts building, let AI first analyze the request and create an understandable plan.

For example:

AI Understanding

Product goal:
What I believe you are trying to achieve.

Users:
Who I believe will use it.

Requirements:
What I believe needs to be built.

User journeys:
How I believe users will accomplish their goals.

Business rules:
Rules I identified from the request.

Data:
Data and relationships required.

Edge cases:
Cases that need to be handled.

Assumptions:
Things I had to assume.

Unknowns:
Things I don't know.

Ambiguities:
Things that could have multiple interpretations.

Then show:

AI Understanding Confidence: 86%

And allow:

Confirm & Build
Edit Understanding
Ask AI to Clarify

This could significantly reduce hallucination before any AI credits are consumed.

2. AI PRD Generator + Approval Gate

Allow Make to transform an idea or conversation into a structured PRD containing:

• Problem

• Goals

• Non-goals

• Personas

• User stories

• User journeys

• Functional requirements

• Business rules

• Data requirements

• Permissions

• Edge cases

• Empty/loading/error/success states

• Accessibility

• Responsive requirements

• Analytics

• Acceptance criteria

• Dependencies

• Risks

• Assumptions

Most importantly:

Do not build until the user approves the PRD.

This creates:

Idea → PRD → Human Confirmation → Build

rather than immediately generating UI from an ambiguous prompt.

3. AI Clarification / Product Interview Mode

AI should behave like a Product Manager / Business Analyst when requirements are incomplete.

Instead of expecting users to write perfect prompts, AI should ask:

• Who are the primary users?

• What is the main business goal?

• What are the critical workflows?

• Who can perform each action?

• What permissions are required?

• What happens when there is no data?

• What happens when an action fails?

• What are the business rules?

• What integrations are required?

• What should happen on mobile?

• What are the acceptance criteria?

The goal should be better understanding before better generation.

4. Specialized AI Agent Team

Instead of one generic AI, allow users to select specialized agents:

Product Manager Agent

Goals, scope, priorities and product strategy.

PRD Agent

Requirements, user stories and acceptance criteria.

Business Analyst Agent

Business rules, data, dependencies and edge cases.

UX Researcher Agent

Research, personas, journeys and insights.

UX Designer Agent

Flows, IA, interaction and experience.

UI/Design Agent

Visual design, components and design-system usage.

Developer Agent

Architecture, code, logic and integrations.

QA Agent

Functional, visual, responsive and regression testing.

Accessibility Agent

WCAG and accessibility validation.

Data/Analytics Agent

Metrics, calculations, events and data validation.

UAT Agent

Acceptance testing and stakeholder validation.

Release Manager Agent

Readiness, approvals, release notes and deployment preparation.

5. Agent Orchestration

The agents should work as a coordinated team rather than isolated assistants.

Example:

Product Idea

↓
Product Analyst

↓
PRD Agent

↓
UX Agent

↓
Design Agent

↓
Developer Agent

↓
QA Agent

↓
UAT Agent

↓
Release Agent

All agents should share one Product Source of Truth.

6. Agent Skills + Knowledge + Permissions

Allow users to configure each agent with:

Skills

• Healthcare UX

• Enterprise UX

• Accessibility

• Design Systems

• React

• TypeScript

• Data Visualization

• QA

• UAT

Knowledge

• PRDs

• Design systems

• Brand guidelines

• Research

• Product documentation

• Business rules

• API documentation

• Coding standards

Permissions

Control whether an agent can:

• Read

• Analyze

• Edit

• Build

• Test

• Approve

• Release

This would make agents much more useful for real product teams.

7. AI Output Verification / Trust Layer

I would love to see a dedicated verification layer for everything Make generates.

Validate:

UI

• Layout

• Typography

• Spacing

• Components

• Design tokens

Design System

• Correct components

• Correct variants

• Correct tokens

• No invented components

UX

• User flows

• Navigation

• Missing states

• Dead ends

• Error handling

Logic

• Conditions

• State transitions

• Business rules

• Validation

Data

• Data types

• Relationships

• Totals

• Percentages

• Calculations

• Consistency

Accessibility

• Contrast

• Keyboard navigation

• Focus states

• Labels

• Touch targets

Responsive

• Desktop

• Tablet

• Mobile

Then provide:

AI Trust Score: 92/100

with clear issues and suggested fixes.

8. AI Assumptions & Evidence

For important decisions, AI should tell us where the decision came from:

Source

• User prompt

• PRD

• Design system

• Attached file

• Connected source

• Existing code

• AI assumption

Example:

“I used the Primary Button component because it exists in your connected design system.”

Or:

“I assumed USD because currency was not specified.”

This would make AI behavior much more transparent and reduce hallucination.

9. Requirement → Design → Code → QA → UAT Traceability

This could be extremely valuable for professional product teams.

Example:

Requirement R-024

→ User Story

→ User Flow

→ UI

→ Component

→ Code

→ QA Test

→ UAT Test

→ Release

Then show:

Implemented ✓
Validated ✓
UAT Approved ✓

This would give teams confidence that AI didn't simply create something visually convincing but incomplete.

10. AI-Generated QA & UAT

Once the PRD is approved, AI should automatically create test cases.

For each requirement:

• Happy path

• Negative path

• Edge cases

• Permission cases

• Error states

• Responsive behavior

• Accessibility

• Data validation

Then provide a dedicated:

UAT Workspace

Where stakeholders can:

• Run test cases

• Pass/fail

• Add comments

• Attach screenshots

• Create defects

• Assign issues

• Retest

Ideally, UAT should happen directly against the generated prototype/application.

11. Release Management

After development and UAT:

Release Readiness

Requirements: 100%
Design: 100%
Development: 96%
QA: 94%
UAT: 92%
Accessibility: 100%
Critical Issues: 0
Blockers: 0

AI Recommendation

READY FOR RELEASE

But require configurable human approval before release.

12. AI Credit Intelligence

I know Figma is already improving AI credit visibility, so I don't want to simply ask for “more credits.”

I would like to see task-level credit intelligence:

Task: Create workforce dashboard
Estimated: 250–350 credits
Actual: 312 credits

Then:

Version 4

• Initial generation: 210

• Layout refinement: 42

• Logic: 35

• Validation: 25

Also provide:

• Project budget

• Version budget

• Task budget

• Credit estimate before execution

• Actual usage

• Usage history

• Cost by agent/model/task

This would make AI usage much more predictable.

13. Full Manual Editing as a No-Credit Safety Net

AI should never make the product feel unusable when credits are exhausted.

Manual editing should progressively become closer to the flexibility of Figma Design:

• Layout

• Components

• Variants

• Variables

• Typography

• Responsive behavior

• Interactions

• Structure

• Data

• Logic

Manual editing should not consume AI credits.

This gives users:

AI → Manual → AI

rather than:

AI → Credits exhausted → Workflow blocked

14. Versioning, Change Impact & Rollback

Every significant AI operation should capture:

• Prompt

• Agent

• Skill

• Model

• Credits

• Context

• Requirements affected

• Components affected

• Changes

• Validation

• Tests

• Approval

Before a major change:

“This change will affect 12 screens, 34 components and 7 UAT cases.”

Then:

Review → Apply

And provide:

Compare / Branch / Rollback

15. Production Mode vs Exploration Mode

I would love two modes:

Exploration Mode

AI can experiment freely.

Production Mode

AI must:

• Follow the approved PRD

• Follow the design system

• Respect locked components

• Preserve existing functionality

• Validate changes

• Run tests

• Maintain traceability

• Require approval for high-risk changes

This would make Make much safer for real products.

The Bigger Opportunity

I don't think the next evolution of Figma Make should only be:

“Generate better UI.”

I think it can become:

AI Product Development OS

IDEA

↓

AI ANALYST

↓

PRD

↓

AI UNDERSTANDING

↓

HUMAN CONFIRMATION

↓

AGENT TEAM

↓

DESIGN

↓

BUILD

↓

VERIFY

↓

QA

↓

UAT

↓

HUMAN APPROVAL

↓

RELEASE

With one shared source of truth containing:

Requirements + Design System + Skills + Agents + Data + Business Rules + Versions + Tests + UAT + Approvals + Releases

My core message to the Figma team

Figma Make has already made AI generation increasingly powerful. I believe the next major opportunity is not just making AI build faster — it is making AI trustworthy enough for professional product development.

The question I would love Figma to solve is:

“How can AI prove that it understood what I meant before it builds, prove that what it built is correct after it builds, and help my team safely move that work through QA, UAT and release?”

If Figma can solve that, Make could move from an impressive AI prototyping tool to a true AI-native product development platform.

I'd love to hear how the Figma team is thinking about this direction and where the community can help shape it.