Monday, September 7

Building Your Own Tool with Claude Starts with Decision Rules | XPLORED

Monday, September 7

Building Your Own Tool with Claude Starts with Decision Rules

Building Your Own Tool with Claude Starts with Decision Rules — Image source: Medium · Sophia Hee

1. Explain decision rules, not just the desired appearance

Results improved when “make it smoother” became specific rules for classifying shapes.

Classify outlines as straight lines, arcs, or Bézier curves
- Rather than tracing blurry pixel boundaries directly, ask the AI to identify the underlying shape first.
Use a similar typeface as the starting point for logo lettering
- Use the typeface as a framework, then adjust it to match the original’s position and weight.

2. Mark what is wrong and preserve the reasoning behind the fix

Visual feedback becomes clearer when it communicates both the problem’s location and the intended shape.

Circle errors directly on the output
- Point out a stair-stepped outline and explain that it is actually one diagonal line.
Incorporate the corrected decision rule into the code
- Improve the tool’s behavior so the same explanation is not needed for the next logo.

3. Create a known answer to validate the results

An SVG you create yourself provides a reference for identifying where and how much the result deviates.

Convert an SVG to PNG, then trace it back
- Give Claude only the PNG and compare its output with the original SVG.
Create the conditions you want to test
- Introduce blur, small lettering, or stair-stepped edges to test specific problems.

4. Use the plugin for routine work and Claude for exceptions

Package validated rules into a Figma plugin that fits your everyday workflow.

Figma plugin: handle everyday tracing
- Reduce the need to move images between apps or start a new conversation each time.
Claude: handle exceptions in complex logos
- Refine the incorrect parts together, then incorporate the resulting rules back into the plugin.

XPLORED View

This case shows how explaining and testing a designer’s visual judgment can lead to a tool that works repeatedly.

When to read

When you keep giving AI similar revision requests but get inconsistent results
When you want to build your own tool for a frequent design task

Takeaways

Turn the feeling that something looks off into observable criteria such as spacing, alignment, weight, and corners
Go beyond one successful result by keeping tests with known answers alongside reusable rules