Optimize Skill

optimize-skill

Make a skill measurably better without uncontrolled drift. This is an agent-in-the-loop loop, not a hands-off run: the CLI owns the deterministic machinery (scoring, edit budget, rejected-edit buffer, the gate, atomic promote/revert); you own the reasoning (running the skill, judging prose outputs, proposing edits). Expect several turns per optimization.

Preconditions

• The skill has an eval manifest (eval.yaml) with real tasks. If it has none, stop and offer to create one (create-skill) — optimization cannot run without an eval set.
• Optimization edits a managed copy, never the installed symlink target.

The loop (repeat until the gate stops improving)

1. Rollout. For each eval task input, run the current skill yourself and collect its output. Write rollouts.json: [{ "taskId": "...", "output": "..." }].

1. Score.

``bash scripts/optimize.sh score --eval eval.yaml --rollouts rollouts.json --skill --json ``

Deterministic tasks are scored for you. agent-judge tasks come back as pending — score those yourself against each task's rubric and fold them into the aggregate. Record the current score.

1. Reflect. Read the failing trajectories. Diagnose why they failed (this is your job — the CLI never judges why). Propose a small set of structured edits to SKILL.md. Write edits.json: an array of { op: append|insert_after|replace|delete, target?, content?, sourceType: "failure"|"success", supportCount? }. Prefer failure-driven edits.

1. Apply (bounded).

``bash scripts/optimize.sh apply --skill --skill-dir --edits edits.json --step --total ``

The CLI caps edits at the budget (annealed over steps), skips edits to the protected SLOW_UPDATE region, and writes a candidate copy. Note the candidate dir it prints.

1. Validate. Re-run rollout + score against the candidate (steps 1–2 pointing at the candidate dir), including the held-out tasks. Then gate:

``bash scripts/optimize.sh gate --current --candidate --best ``

A non-zero exit means reject — add those edits to your rejected set so you don't re-propose them, and try a different reflection. Also reject if any held-out task regressed, even if the aggregate improved.

1. Promote (human gate). Only on a strict improvement with no held-out regression, present the candidate and its score delta to the user. On their approval:

``bash scripts/optimize.sh promote --skill --live --candidate --score ``

The prior version is retained and the change is reversible.

Revert

scripts/optimize.sh revert --skill  --version  --live 

Honesty

• "Optimize automatically" means the loop, budget, buffer, and gates are automated — the intelligence (rollout, reflection, edits, agent-judge) is yours. A weak reasoning pass simply makes less progress; the gate guarantees no regression is ever promoted.
• Never promote without explicit user approval, even when the gate passes.

Source & license

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

• Author: Bennyoooo
• Source: Bennyoooo/skillmaxxing
• License: MIT

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