Mirror

Armor for coding agents: a playbook of Claude Code skills that make an agent prove its work, not describe it.

Install

npx -y skills add yiyaw-lab/agent-armor --skill mirror

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Sync a skill from ~/.claude/commands/<x>.md to the public skills repo (yiyaw-lab/agent-armor; local clone ~/code/claude-code-skills) as <x>/SKILL.md — disclosure+secrets gate first, README bookkeeping (badge count, playbook diagram, catalog row) handled, then a PR with CI green; merge only on explicit approval. Curated: only named skills, never a bulk export. Modes: <skill-name> | sweep | audit.

SKILL.md

4.9 KB, as published. Nobody here has run it

The law this skill runs on: a public mirror rots silently. The suite's skills are sharpened in place (~/.claude/commands — /hone edits them, lessons fold in), but the public copies in the skills repo update only when someone remembers — so the repo's contract ("this armor updates when my own workflow does") decays into stale copies and a wrong skill count. mirror makes that update a checked procedure: one command takes a named skill from the private commands dir to a reviewed PR on the public repo, with the disclosure gate and the README bookkeeping nobody remembers by hand.

Three hard invariants:

Curated, never bulk. Mirror ONLY the skill(s) the user names, or already-public skills that drifted. Presence in ~/.claude/commands is NOT admission — the repo's bar is "survived contact with real work," and the user decides what's public (some skills stay private on purpose: personal pipelines, unproven drafts).
Nothing leaves the machine unscanned. Before writing into the clone: a secrets scan (keys/tokens/bearer strings) and a disclosure scan (private paths, client/project names not already public, identifiers beyond the repo's own voice). A hit = stop and surface; never silently redact.
PR-gated; merge is a per-run human call. Push a branch, open the PR, wait for the repo's validate-skills CI. Merge only when the user says so in THIS run (a standing "publish it" in the invocation counts). Never force-push, never commit to main directly.

/mirror <skill-name>

Sync the clone first: git -C ~/code/claude-code-skills pull --ff-only — a mirror from a stale clone re-introduces what was already fixed upstream.
Read the source ~/.claude/commands/<x>.md and run both scans (invariant 2).
Transform = copy, honestly. <x>/SKILL.md carries the same frontmatter (description, argument-hint) and body verbatim. Private-convention references (memory files, private/ taxonomy, TASTE.md) STAY — the README already tells readers to swap in their own conventions; flag anything that would genuinely confuse a public reader rather than silently rewriting it. If the skill ships a script, copy it under <x>/scripts/. When the public copy MUST intentionally diverge (genericized private module names, references to unpublished skills, personal-detail sanitization), make the divergence machine-visible: put  as the first body line of the public SKILL.md, so a future sweep never "fixes" it back to the private text.
README bookkeeping (the part nobody remembers): a NEW skill bumps the skills-N badge AND the prose count, joins its playbook-diagram line, and gets a catalog row in the section that owns it, written in that table's register. An UPDATE changes no counts. Badge, prose, diagram, and table must agree — CI validates skills, not the README; mirror owns that agreement.
Branch, commit, PR: branch skill/<x>, explicit git add of only the touched files (never -A), commit, push, gh pr create. Wait for the validate check; report its verdict + the PR URL.
Merge on approval only (invariant 3). After merging, confirm the default branch actually contains the change and report.

/mirror sweep

For every repo dir with a same-named ~/.claude/commands/<x>.md, diff <x>/SKILL.md against the command body. A public copy containing a mirror:transform marker is bucketed TRANSFORMED, not drifted: report its diff for awareness but never propose a verbatim sync over it (a naive diff cannot tell stale from deliberately-divergent — the marker can). Report five buckets: IN-SYNC · TRANSFORMED · DRIFTED (public copy stale — summarize what changed privately) · PRIVATE-ONLY (in commands, not mirrored — candidates; the user picks) · REPO-ONLY (mirrored but no local command — flag it, NEVER delete; it may be another machine's skill). Propose which drifted ones to sync; execute per approval, batched or per-skill as the user prefers.

/mirror audit

Read-only: the same four-bucket report; propose nothing, write nothing.

Rules

Lane discipline: /publish owns yiya.dev content; /harmonize converges internal parts on a named standard; a skill-body EDIT is /hone's job. mirror owns exactly the cross-repo publication mechanics for the skills repo: scan → copy → README bookkeeping → PR.
Never delete on sweep; REPO-ONLY is a flag, not a cleanup target.
Honest empty result: nothing named and nothing drifted = say so and write nothing.