Harness

Judge the harness by the outcomes it gives the agent and the user, not by the files it has today. Nothing earns protection by already existing: add, rewrite, merge, move or delete whatever makes the harness better, and preserve the useful behavior, not the component that carried it. write-skill owns skill-level consolidation and architecture owns structural alternatives when either helps.

Pick the posture the request needs:

• Assess — measure the current harness and report drift, conflicts, overlap and waste. Stays local: no upstream fetch or checkpoint change.
• Reassess or redesign — after a model, host or recurring-failure change, or when asked to reorganize or simplify.
• Absorb upstream — on an explicit request to review sources.

Both deeper postures load Review, which owns coverage, native-host checks, the single approved plan and ledger checkpoints, and uses Behavior probes when comparison evidence is needed.

Inspect every layer

Each layer earns its keep differently. Judge it on its own test, and against the host's current native features: a component built around a host gap becomes waste when a release closes the gap.

• Skill — gives a useful stance on its intended turns. Tell a weak trigger from missing discovery, selection or observation before retiring it.
• Agent preset — holds a role the fleet cannot already cover, and returns evidence the parent could not gather inline.
• Extension, hook or tool — owns state, commands, UI, gates or lifecycle that a skill cannot carry, as one shared piece rather than one per caller.
• Setup and package surface — keeps source, installed copies and runtime resources separate, and preserves unrelated user files.
• Doc — states truth the environment cannot be asked for. Anything a script, config or --help answers is a cache that will go stale.
• Test — catches real regressions in executable contracts, rather than pinning prose or certifying behavior it never exercised.

Also look across layers for instruction drift between AGENTS.md/CLAUDE.md copies, permission and preset conflicts, stale memory, weak navigation pointers, and two owners of one fact.

For a repeated mechanical mistake, inspect existing tests, lint, scripts and CI: a missing, unwired or broken check may be the defect. Prefer a focused deterministic check over another prose rule. Keep judgment and non-obvious authority boundaries in guidance; do not install a hook merely because one is absent.

Measure the skill layer

Resolve SKILL_DIR to the directory containing this loaded SKILL.md, not a checkout-relative path. For an isolated source inventory, run:

bun "$SKILL_DIR/scripts/skill-cleaner.ts" --root PATH --root-only --no-logs

Use --help for options. Log scans read private history and need that scope; --log-root PATH replaces the default Pi/Claude roots. Do not widen roots merely to fill gaps. Default inventory omits project .claude/skills and package resources; add them only as explicit roots. Prefer an actual host-exported inventory: filesystem presence and frontmatter eligibility are not effective selection.

Read coverage and skipped/error/byte counts before interpreting results. Missing signals, partial logs and inherited context are unknown, not proof of disuse. Token figures are character-ratio estimates of listings and bodies, not native prompts, retention or savings; reference and inherited-context costs stay unmeasured.

Ground rules

Current config and command output outrank memory. Fix the owned source, not an unexplained deployed copy. Similarity is a lead, not identity or permission to overwrite. Preserve user files; deletion of user-owned material needs authority. A clean assessment writes nothing.

Deliver: posture, layer coverage and evidence limits, ranked findings with owning paths, the proposed changes (including removals and moves) with preserved or retired behavior, and the validation still unrun.