Skip to content

program.md

Latest commit

 

History

History
63 lines (41 loc) · 2.05 KB

program.md

File metadata and controls

63 lines (41 loc) · 2.05 KB

Running the Evolution Loop with Coding Agents

This guide explains how to manually drive the prompt evolution loop using an AI coding assistant like OpenCode or Cursor.

How It Works

You take the role of the human evaluator in the modify → evaluate → keep/revert loop:

  1. Modify — Improve the prompt in prompt.txt
  2. Evaluate — Run python eval.py to see the score
  3. Keep or Revert — Commit improvements, revert regressions
  4. Repeat — Iterate until the prompt is excellent

Instructions

Step 1: Open the Project

Open this repo in OpenCode or Cursor. The file you'll edit is prompt.txt at the repo root.

Step 2: Improve the Prompt

Edit prompt.txt to make it better at generating AI agent projects. Some ideas:

  • Add more specific instructions about project structure
  • Request specific libraries or patterns (LangGraph, Pydantic, etc.)
  • Clarify output format requirements (markdown code blocks)
  • Add constraints about code quality, testing, or documentation

Step 3: Evaluate

After each change, run:

python eval.py

The evaluator scores your prompt from 0–100 based on quality signals.

Step 4: Keep or Revert

Score Action
Higher than before Commit: git add prompt.txt && git commit -m "Better prompt - score X"
Same or lower Revert: git checkout prompt.txt and try a different approach

Step 5: Repeat

Keep going for many iterations. Each generation brings you closer to the optimal prompt.

Automated Evolution

For a fully automated version (no human needed), use run_evolution.sh. It uses a genetic algorithm with mutation, crossover, and selection — no manual editing required.

chmod +x run_evolution.sh
./run_evolution.sh

Tips

  • Make small, focused changes rather than rewriting the whole prompt
  • Check results.log to track your score history
  • The best prompts are specific, structured, and mention concrete libraries and patterns
  • Focus on instructions that lead to production-ready, local-first agent projects