Google Stitch 2.0 can generate production-quality UI in seconds.
But the moment you try to refine a single component, the system starts to break.
This post is a hands-on evaluation of Stitch while building a real product flow, focusing on where it excels and where it fails under practical usage.
TL;DR
- UI generation quality is exceptionally high (near senior-level output)
- Iteration speed with Gemini 3 Flash is extremely fast
- DESIGN.md improves consistency across outputs
- Component-level editing is unreliable
- Agent can become unresponsive with no feedback
Conclusion: Strong for generation, weak for controlled refinement
Demo Video
Context
This evaluation was done while building a real product flow for an AI-powered agriculture application.
The workflow included:
- Prompt → UI generation
- Multi-screen navigation
- Component-level edits using "Edit with AI"
- Export pipeline testing
What Works Well
High-quality UI generation
Using Gemini 3.1 Pro ("Thinking" mode), Stitch produces layouts with:
- Strong visual hierarchy
- Clean spacing and alignment
- Coherent component structuring
The output often matches what a mid-to-senior designer would produce as a first iteration.
Fast iteration loop
With Gemini 3 Flash:
- Multiple UI variations are generated within seconds
- Exploration cost is significantly reduced
This compresses early-stage design cycles into a single interaction loop.
Accurate instruction-to-design translation
Compared to earlier versions:
- Prompts are interpreted more consistently
- Layout intent is preserved
- Component grouping is largely correct
DESIGN.md improves consistency
Defining design rules such as:
- Colors
- Typography
- Component constraints
results in more consistent outputs across iterations.
Export layer is well structured
Available outputs include:
- AI Studio
- Figma
- Code export
This creates a clear bridge between design and implementation.
Where It Breaks
Case Study: Scan Button Component
While refining a floating "Scan Crop" button, the following behavior was observed.
Intended behavior
- Default state: circular icon button
- First interaction: expands into pill with "Scan" label
- Second interaction: triggers scan action
Step 1: Edit with AI
Result:
- Button collapsed into circular form (expected)
- Scan icon was removed (unexpected)
This resulted in a visually ambiguous control with no clear affordance.
Step 2: Targeted correction
Instruction:
"Add scan icon while keeping all other properties unchanged"
Result:
- No visible update
- No feedback
- No error signal
Step 3: Repeated attempts
- Same instruction issued multiple times
- No change observed
Outcome
- Edit pipeline became unresponsive
- Component could not be refined further
- Required abandoning the edit flow
Analysis
This behavior indicates a system-level limitation rather than a simple generation error.
Weak component-level editing
The system struggles with scoped modifications such as:
- Modify one attribute while preserving structure
Lack of deterministic control
Edits are:
- Non-reliable
- Not precisely applied
Agent state inconsistency
The system appears unable to:
- Maintain prior edit context
- Apply incremental changes
Silent failure
There is no:
- Error feedback
- Retry mechanism
- Recovery guidance
The system simply stops responding.
Why This Matters
This directly impacts practical usability:
- Reduces trust in iterative refinement
- Forces regeneration instead of controlled editing
- Slows down later-stage prototyping
Stitch is strong in generation, but unreliable in refinement.
Missing Capabilities
For production-level workflows, the following are required:
Deterministic edit mode
Apply scoped, predictable changes.
Component locking
Preserve structure while editing specific attributes.
Feedback system
Provide clear signals when edits fail.
Visual diff
Show before and after changes.
Retry mechanism
Automatically handle failed edits.
Strategic Observation
Stitch 2.0 has largely solved the generation problem.
The remaining challenge is control.
Final Assessment
Strengths:
- High-quality UI generation
- Fast iteration
- Strong direction toward AI-native design workflows
Limitations:
- Fragile editing
- Lack of deterministic control
- Agent reliability issues
Conclusion
Stitch is already a highly capable prototyping system.
However, until component-level editing becomes reliable, it cannot replace traditional design workflows for production use.
For now:
- Use it for rapid generation
- Avoid relying on it for precise refinement