Agent
Read-only auditor for a repo's agent-instruction system. Used by /instructions-audit (or directly) to sweep CLAUDE.md, .claude/, and .agent/ for broken pointers, dead/duplicate rules, contradictions with reality, cap breaches, and enforcement drift. Returns a severity-ranked findings list with file:line evidence…
Agent
Prior-art lookup — searches the project's .agent/lessons/ (plus .docs/) for lessons matching an error, symptom, or component, and returns the matches with relevance. Use PROACTIVELY before debugging a non-trivial error, so you reuse what past sessions already learned instead of re-deriving it. Never proposes fixes …
Command
De-rot the instruction system — pointers, liveness, truth-vs-reality, duplicates, caps, enforcement; T2 applied, T3 proposed (runs the instructions-audit skill).
Command
Load the instruction-system constitution before changing CLAUDE.md / .claude / .agent (runs the instructions-maintenance skill).
Command
Capture or curate a durable lesson in .agent/lessons/ — one file, tiered + indexed (runs the lessons skill).
Command
Wrap the session — harvest learnings, route to the right stores, apply-safe/propose-risky, commit (runs the retro skill).
Command
The sanctioned way to change a rule/cap/enforcement — explicit approval, coherence check, re-test, changelog (runs the rules-change skill).
Skill Claude CodeCodex
De-rot the agent instruction system — after many sessions of accretion, keep it SMALL, TRUE, and NON-DUPLICATED. Sweeps for broken pointers, dead/duplicate rules, truth-vs-reality drift, cap breaches, and enforcement drift; applies safe fixes (T2) directly and batches risky ones (T3) into one proposal. Use when a size…
Skill Claude CodeCodex
The constitution of the agent instruction system — the rules for how CLAUDE.md, .claude/ (skills, rules, hooks, settings, agents), and .agent/ evolve. Load this BEFORE creating or modifying any instruction file, when adding a skill/rule/hook/agent, changing a convention, or capturing a lesson. Owns the design…
Skill Claude CodeCodex
Capture and curate durable lessons in .agent/lessons/ — one file per lesson, indexed in its README, surfaced by tier (enforced hook / routed doc / indexed). Use whenever the user corrects a repeatable mistake ('don't do that anymore', 'I already told you', 'remember this'), asks to save a lesson, right after a hard…
Skill Claude CodeCodex
Harvest durable learnings from this session into the instruction system — with provenance, dedup, and pruning. Use at the end of a significant session, after a user correction worth keeping, when a convention drifted or a footgun was hit, or when the user says "retro", "wrap up", "save learnings", "what did we learn"…
Skill Claude CodeCodex
The sanctioned way to make a T3 change to the instruction system — add/remove a rule, change a cap, or change any hook/skill/agent/settings behavior. Use ONLY on an explicit user request or approval (including approving a retro/audit proposal, or granting a one-time exception). Never invoke as a side effect of another…
Agent
Verify that recent code changes comply with the project's pattern registry at .agent/patterns/. Read-only audit. Different from pattern-verifier (which checks the project's coding-style/convention rules) — this one audits against the formal per-topic pattern bodies in .agent/patterns/{topic}.md, checking shape…
Agent
Verify that recent code changes comply with the project's own coding-style / convention rules. Read-only review pass — outputs violations grouped by rule. Use as a fast code-shape check (e.g. the third pass in a fix → verify → fix-pass-2 chain) — lighter than the deeper pattern-compliance registry audit. Reads the…
Command
Author or curate the .agent/patterns/ registry — walk a topic through the 5-phase flow, flip status, split/merge (runs the manage-patterns skill).
Skill Claude CodeCodex
Author and curate a per-topic pattern registry under .agent/patterns/. Use when: adding a new pattern topic, modifying one (status flip TODO → decided → blessed, body rewrite, anti-pattern addition, cross-reference update), splitting/merging topics, or when the user says "bless a pattern", "decide a pattern", "walk a…
Command
Rename a file with the full reference cascade — git mv + update every pointer (runs the rename skill).
Skill Claude CodeCodex
Rename a file with full reference cascade — git mv the file, then update every non-archive reference across the project. Use when the user says: "rename X to Y", "move X to Y", "X should be called Y", "the file should be Z instead", or signals a path/name change. Catches the dozen-references-cascade problem (renames…
Agent
Read-only drift audit of a repo's binding docs against code reality. Use after large refactors, before a release, or when docs feel stale — verifies CLAUDE.md, rules, agent memory, docs/, and any deploy contract still tell the truth. Returns a drift table with file:line evidence; never edits. Broader than…
Agent
Adversarial multi-tenant isolation review of a diff or module. Use proactively after changes to API routes, auth, tenant lifecycle (join/leave/delete), sharing, tokens, or offline/cache sync — and before any release of a multi-tenant app. Reports findings with file:line evidence and severity; read-only.
Agent
Run a semantic-bug review against changed files using a 7-item checklist (resource lifecycle, error handling, unbounded growth, idempotent cleanup, empty input, shutdown, unvalidated casts). Catches bugs that grep-based pattern checks miss. Use after architectural refactors or anytime resource/lifecycle code changed.…
Agent
Given a feature or topic name, verify the implementation matches what's documented in the project's specs (.docs/). Read-only diff between docs and code. Use when you suspect a feature has drifted from spec, before opening a PR, or as a verification step after a refactor. More targeted than a full drift audit — one…
Agent
Given a wireframe/mockup file (HTML) and the implementation file paths, identify functional gaps between the wireframe and the code (missing buttons, unhandled states, dropped error feedback, unimplemented modes). Use when a UI screen has been implemented and you want a third-party check that the wireframe's…
Command
Run the full done-gate — typecheck, unit, e2e, visual — before claiming done (runs the qa-gate skill).