Recallx Harness Self Improve

RecallX Harness Self-Improve

Use this skill only after a meaningful RecallX task is already complete.

The goal is not to reopen the task. The goal is to decide whether the harness needs one small reusable improvement.

Inputs To Review

Look at only the minimum evidence needed:

1. the user task and what was actually changed
2. failures, retries, or drift that happened during the run
3. validation commands that were run or skipped
4. current harness files:
   • AGENTS.md
   • relevant files under .codex/skills/
   • relevant files under .codex/hooks/

Decision Rule

Ask these in order:

1. Did the task expose a repeated preflight miss?
2. Did drift happen that a better checklist or reminder could have prevented earlier?
3. Did validation guidance prove wrong, too vague, or too broad for this repo?
4. Did I repeat the same explanation or manual decision that belongs in a reusable skill or hook?

If all four are "no", do not change the harness.

If any answer is "yes", make only one improvement that is:

• small
• reusable across future RecallX tasks
• easy to validate

Where To Put The Improvement

• AGENTS.md for repo-wide policy, guardrails, and definitions of done
• .codex/skills/ for reusable agent workflows or task-specific reasoning loops
• .codex/hooks/ for lightweight command-line reminders or validation routing

Choose the narrowest location that solves the problem.

Hard Limits

• Make at most one bounded harness improvement per task.
• Do not rewrite the harness wholesale.
• Do not create a new skill if an existing skill only needs one short rule.
• Do not use this skill to sneak in unrelated cleanup.
• After one self-improvement pass, stop.

Finish

Record one of these outcomes:

• "No reusable harness lesson emerged."
• "Applied one bounded harness improvement: ."

Source: jazpiper/RecallX — distributed by TomeVault.