Evolve

Evolution pass (prompt-space gradient descent)

You are updating this system's weights: the personalized rules injected into every prompt. Be conservative — a wrong learned rule costs the user on every message.

Procedure

1. Read ~/.claude/command-autopilot/events.jsonl. Also read learned.json if present (schema below).
2. Cluster events by task-type × command/skill. Look for consistent patterns, for example:

• a command suggested repeatedly and consistently dismissed → candidate negative rule
• a skill invoked on the same kind of task again and again → candidate positive rule ("invoke X early for Y-type tasks")
• a habit self-used regularly → mastered, teaching for it should stop

1. Apply the discipline:

• Promote only patterns with ≥3 consistent observations and no contradicting evidence → status: "in_force".
• Demote existing in_force rules contradicted by new evidence (decrement evidence; at 0, set status: "candidate").
• Decay: rules not reconfirmed for ~60 days → delete.
• Max 5 in_force rules; each text ≤ 25 tokens, English, imperative, generic phrasing ("research-type tasks: offer /fork early").
• NEVER write rules that override the safety net (/rewind) or the one-suggestion contract.

1. Write learned.json:

{
  "updated": "",
  "rules": [
    {"text": "...", "evidence": 4, "first": "", "last": "", "status": "in_force"}
  ]
}

1. Archive processed events — rotate FIRST to avoid racing concurrent appends: rename events.jsonl to events-archive-.jsonl, and only then read the renamed file for the distillation. New events land in a fresh events.jsonl untouched. (If you already read before rotating, rotate anyway and accept the tiny overlap.)
2. Report to the user in their language: what was learned (each rule + its evidence), what was demoted or deleted, and one line on what will change. If nothing met the bar, say so plainly — no fabricated learnings.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: WinterDDo
• Source: WinterDDo/claude-code-skill-autopilot
• License: MIT

Install and usage instructions live in the source repository linked above.