Real AI Coding Workflows: Beyond Figma Dumps
Real AI Coding Workflows: Beyond Figma Dumps
Stop feeding entire Figma files to your AI coding agents. You're burning tokens and ruining your codebase.
Like many, I initially treated AI coding tools as magic wands. I'd connect the Figma MCP, point the agent to my data sources, feed it a full design file, and tell it to build.
The result? A massive 1-2 million token context burn per run. The AI constantly messed up my project-specific API structures, failed to reuse existing components, and generated code that stood out like a sore thumb.
I briefly reverted to writing 80% of the code myself, creating hard-coded skeletons, and only using AI for data integration. But that was painfully slow.
I realized the problem wasn't the AI—it was my workflow.
The 400-Line Master Guideline
Instead of letting the agent blindly explore the whole project to understand the context, I used a heavy model (Sonnet/Opus) to analyze the project once.
I had it generate a strict guideline detailing:
- How components should be structured
- Where imports live
- How our APIs are handled
Now, the agent only reads this doc. Massive token savings, and the AI stays consistent with our patterns.
Micro-Component Prompting
I stopped feeding the agent entire screens. I put my architect hat on and broke down the UI logically into distinct files:
- Define the exact props and placement for each component
- Let AI build isolated components one at a time
- Stitch them together manually
This gives you:
- Clean, reusable components
- Consistent design language
- AI that doesn't hallucinate UI patterns
The Automated "Add Icon" Skill
AI is notoriously bad at handling icons—hallucinating them or duplicating SVGs. Doing it manually took 1-2 minutes per icon.
So I built a custom automated skill:
Script 1: Figma MCP Integration
Uses Figma MCP to extract icon metadata and SVG paths from existing components.
Script 2: SVG Optimization
Cleans up the SVG, removes Figma-specific attributes, and standardizes the format.
Script 3: Project Integration
Adds the icon to the correct location in the codebase with proper naming conventions.
Total time per icon: 3 seconds.
The Core Principle
Large context = Low signal
Small, targeted context = High signal
Instead of explaining your entire codebase to AI, explain exactly what each piece needs to do.
What Doesn't Work
❌ Feeding entire Figma files ❌ Pointing AI at a repo and saying "build this" ❌ Letting the model explore freely ❌ No boundaries or constraints
What Works
✅ Micro-component prompting ✅ Strict coding guidelines ✅ Automated skills for repetitive tasks ✅ Clear boundaries and expectations ✅ Token-efficient context
The Meta-Lesson
The bottleneck in AI-assisted development isn't the AI. It's the workflow around it.
Build systems that help AI help you:
- Structured guidelines
- Clean component boundaries
- Automated skills for low-level tasks
- Proper context management
AI coding tools are only as good as the systems you build around them.