review-bench/scotthamilton77_agents-config/commands

Optimize Skill

Audit and improve existing SKILL.md files via the optimize-my-skill skill.

Tip: For creating new skills from scratch, prefer the writing-skills skill. This command is for auditing and improving existing skills.

Arguments

$ARGUMENTS accepts one non-flag target and optional flags:

/optimize-my-skill [<target>] [--deep] [--max-iterations N] [--model <name>]

Parse by splitting $ARGUMENTS on whitespace and walking tokens left-to-right:

• --deep is a boolean flag (consumes no value).
• --max-iterations and --model are value-taking flags: the next token

after the flag is its value and MUST NOT be considered as <target>. Treat --max-iterations=N / --model=<name> (equals-form) as a single token.

• The first unclaimed (non-flag, non-value) token is <target>. Any further

unclaimed tokens are an error — reject with a usage message.

• Unknown --flags are an error — reject with a usage message.

Collect flags separately from <target> and pass both to the skill. Empty <target> (no unclaimed token after flag/value consumption) triggers the probe in the next section.

Empty-argument resolution

When $ARGUMENTS is empty, probe in this order and pick the first that satisfies the predicate:

Predicate: directory exists AND contains at least one immediate (depth-1) subdirectory that contains a SKILL.md file (case-sensitive; regular file or symlink to regular file). An empty-but-existing SKILL.md satisfies the predicate (file presence, not content validity). A non-existent directory fails the predicate.

1. ~/.claude/skills/
2. .claude/skills/

If both probes fail, emit exactly:

No skills found. Pass a path: /optimize-my-skill <path>

and stop.

Invoke the skill

With the resolved target and any flags, invoke the optimize-my-skill skill and pass them in its invocation context. The skill owns the full methodology (discover, assess, propose, optionally empirical-optimize + iterate, then confirm/apply).

Quick mode (no --deep) is unchanged from prior behavior. Deep mode prompts the user for a cost confirmation before invoking model-heavy phases.

Report

After the skill completes, surface its summary table and the list of files modified (with diffs already shown by the skill).