pi-gepa

Genetic Evolution for Prompts and Agents. Evolve Pi skills and prompts through mutation, crossover, and selection.

Install

pi install npm:@artale/pi-gepa

What it does

GEPA applies evolutionary algorithms to prompt/skill optimization:

1. Mutate — create variants of a skill with targeted changes
2. Crossover — combine strengths of two skills
3. Select — evaluate variants against a fitness function
4. Evolve — repeat until convergence

Results

Tested on real Pi skills:

• canvas-design: −69% token reduction while maintaining quality
• pptx: −60% token reduction
• xlsx: −63% token reduction

Commands

/gepa evolve <skill>         — start evolution of a skill
/gepa status                 — show evolution state
/gepa history                — past evolution results

How it works

Uses the LLM itself as the mutation operator. Each generation:

1. Reads the current skill SKILL.md
2. Generates N mutated variants (shorter, restructured, different examples)
3. Evaluates each variant on a benchmark task
4. Selects the best performer
5. Repeats for M generations

The fitness function measures: output quality + token efficiency + instruction clarity.

License

MIT