coding-mastermind-help

Overview

A task-oriented answer to "how do I use this kit," derived from the kit's real
structure (list the installed coding-mastermind-* skills, read the MANIFEST) rather
than a hand-maintained copy that drifts. Keep the answer to under a screen.

When to Use

• Someone asks how to use the kit, what is automatic, or how to harden a repo.
• You need a quick orientation before reaching for a specific kit command.

When NOT to use: to actually run a command (invoke that command's skill directly).

Process

1. List what is installed (so the answer is current, not from memory):

   ls ~/Documents/agent-skills/skills | grep '^coding-mastermind'
   sed -n '1,20p' ~/Documents/agent-skills/coding-mastermind/MANIFEST.md

2. Answer these, briefly:

   What is AUTOMATIC + global (the advisors). The synced rules and skills every
   agent reads, distributed by the dotfiles rail to Claude Code, Codex, and Gemini. No
   prompt needed. These NAME invariants; they do not enforce them.

   What is PER-REPO (the enforcers). The mechanical checks: a repo's INVARIANTS.md
   registry, its scripts/... checks + CI gates, and (Claude Code only) a Stop hook
   that re-runs them in-loop. The invariant LIVES in CI (vendor-neutral); each agent
   gets the best in-loop enforcement its own harness supports.

   How to harden a new repo. Run coding-mastermind-init (scaffolds a short layered
   CLAUDE.md + an INVARIANTS.md stub + a DoD + a PR template), then author that repo's
   first real invariant and its enforcer. The init skill walks you through it.

   The hands-off loop. The locator reports which INVARIANTS.md rows the touched
   files are subject to; compliance-guard proposes drafting the enforcer in-PR when an
   invariant recurs or a new surface appears (the promotion trigger is the SECOND
   recurrence). coding-mastermind-invariant mints a NEW row plus its enforcer when a
   recurring problem has no row yet (or scaffolds one you name); coding-mastermind-audit
   is the low-frequency backstop that catches what the in-loop check missed.

   The cross-vendor cross-check (coding-mastermind-cross-check). For RARE
   high-stakes or large-diff decisions only: it asks Codex + Gemini to REFUTE a claim
   (read-only/sandboxed) and synthesizes a proposal that preserves the disagreement. Not
   a majority vote. Skip it on Q&A and small edits.

   Maintenance. coding-mastermind-update upgrades the agent CLIs (each via its real
   install channel) and then re-stamps the MANIFEST baseline + re-verifies the capability
   facts; the Wiki's own refresh skill keeps the knowledge wiki current. The MANIFEST
   and wiki edits propose, never auto-apply (the CLI upgrade is the one explicit action).

   Canonical homes. Concepts/why = the system wiki. Cross-repo rules =
   global-rules/. Skills + MANIFEST = agent-skills/. Per-repo invariants = that
   repo's INVARIANTS.md. Decisions = that repo's docs/decisions/.

Red Flags

• Reciting a stale, hand-maintained command list. List the installed skills live.
• Telling someone to run the cross-check for routine work (it is rare-high-stakes only).

Verification

• The answer fit in about a screen, named the actually-installed commands, and pointed
  at the canonical homes.

Related

• Every coding-mastermind-* skill (init, invariant, cross-check, audit, update).
• The kit's SPEC.md, PLAN.md, MANIFEST.md.