mh

WCAG 2.2 AA accessibility, ARIA patterns, React a11y fixes for forms/focus/keyboard nav. Use when building web UI. Don't use for React architecture (mh:frontend-patterns).

Triage + respond to open PR review comments (fetch, classify, fix via mattpocock-skills:diagnosing-bugs, reply). Use when replying to reviewer feedback; say 'address review/แก้ตามรีวิว'. Don't use for review (mattpocock-skills:code-review) or merge (mh:ship-merge).

Scan 12-layer agent stacks, regression, memory pollution, tool discipline, repair loops. Use when debugging a misbehaving harness (stuck loops, rot). Don't use for code review.

Backend architecture, API design, and DB optimization for Node.js/Next.js. Use when building a Node/TS backend. Don't use for Python/Go/Rust backends, or the client half (mh:frontend-patterns).

Catalog of 7 highest-yield blind-spot shapes (cross-file, framework-behavior, data-flow-asymmetry, identity, scope-mismatch, emitted-string, vacuous-test). Auto-loads when blind-spot-hunter runs. Don't use for escalation/output-format or standalone hunting.

Sweep: N parallel agents (default 5) each hunt one small bug, report-only. Use when hunting for small bugs across a codebase. Don't use for PR review (mattpocock-skills:code-review) or session audit (mh:deep-audit).

Detect the project build system and incrementally fix build/type errors with minimal safe changes. Use when a build or type-check is failing. Delegates to the build-error-resolver agent. Don't use for logic bugs (mattpocock-skills:diagnosing-bugs).

Compact a bloated markdown doc for tokens; verify-before-overwrite, grammar stays full. Use when over harness-audit's 20K threshold. Don't use for content grading or suggest-only scans.

Scan context-window consumption across agents/skills/MCP/rules; flag bloat + top savings. Use when context feels full or costs climb. Don't use for one-off response trimming.

Compact LLM-pipeline cost: model routing, prompt caching, retry. Use when designing a multi-model pipeline where cost/latency matter. Don't use for single calls or prompting tips.

Generate a local Claude Code cost report from the cost-tracker metrics log. Use when checking session spend. Don't use for scheduling or budget alerts (none exist).

Post-implementation adversarial audit: reconstruct session state, verify every claim against evidence, score before/after on a defined rubric, implement only evidence-backed fixes, re-score. Use after a significant implementation pass to check it actually improved something. Don't use for a first-pass code review — see mattpocock-skills:code-review.

Scan a design system for token/visual consistency and AI-slop detection, or generate one. Use when starting a project. Don't use for scraping other sites.

Drizzle ORM patterns: schema, type inference, migrations, query builder, relations, transactions. Use when building Drizzle apps on PostgreSQL/SQLite. Don't use for Prisma or TypeORM.

Eval-driven development (EDD) framework for Claude Code. Use when setting up EDD, building graders, or measuring AI-assisted workflow quality. Don't use for end-user feature work.

Load a working-frame: dev/review/research (posture-setter, not a workflow or voice change). Use when switching task posture; say 'dev mode/โหมด dev/ตั้งโหมด'. Don't use for skills or /config's output-style picker.

Frontend-design-direction: typography, layout, tone, and motion. Use when building or restyling UI. Don't use for React architecture (mh:frontend-patterns) or HTML artifacts (plannotator-effective-html).

Frontend architecture, component design, and rendering optimization for React/TS. Use when building a React/TS frontend. Don't use for Vue/Svelte/Angular, backend (mh:backend-patterns), or WCAG/a11y audits (mh:accessibility).

Compact a /goal completion condition: done-when check, one-way-door screen, turn bound. Use when drafting a /goal condition. Don't use for single-turn tasks (do it directly).

gRPC patterns for Node/Bun: proto, @grpc/grpc-js client/server, TypeScript codegen, streaming, deadlines/metadata. Use when building gRPC services in Node/Bun. Don't use for REST/HTTP or non-Node gRPC.

Harness-state surface, two modes: fleet/schema audit, --health for session token cost. Use for harness audits or cost checks. Don't use for repo lint/security (mh:security-auditor).

Parallel divergent ideation (5 isolated agents, rotating frames, novelty/viability/fit scoring). Use when the question is open-ended. Say 'brainstorm/ระดมความคิด/คิดไอเดีย'. Don't use for syntax, lookups, or closed-phrasing asks.

Search past mh:ideate runs via the local qmd collection. Use when recalling a prior ideation run. Say 'ideate search/ค้นหาไอเดีย/หาไอเดีย'. Don't use for a new session (mh:ideate) or code/web research (research).

Incident: run a production incident incl. hotfix. Use when alerts fire or user asks for hotfix. Thai: 'เหตุฉุกเฉิน'. Don't use for non-prod bugs or post-mortem.

Diagnosis + design for latency-sensitive systems, realtime dashboards, market data, streaming, queues, caches, HFT-like infra. Use when designing/reviewing/debugging them. Don't use for batch or offline.

Scan a session transcript for cross-turn patterns ambient auto-memory misses. Use when wrapping up a session; batch-gate via AskUserQuestion. Don't use for single known memories.

Pre-flight gate + review checklist against loop failure modes — spinning, verifier-gaming, wrong-answer completion. Use when designing/reviewing an agent loop. Don't use for one-off tasks.

Catalog of UI-polish details — spacing, borders, shadows, motion, hit areas, text wrapping. Use when a UI feels flat. Don't use for direction choices (mh:frontend-design-direction).

Scan memory store for dangling [[links]], orphans, index drift; --trim archives bloat. Use when MEMORY.md over cap. Don't use for semantic review or harness health.

MySQL/MariaDB schema, query, indexing, transaction, replication, and pool patterns. Use when designing or troubleshooting MySQL/MariaDB. Don't use for non-MySQL databases.

Triage competing tasks and route each to inline/parallel/sequential/drop. Use when the user lists tasks or says 'จัดสรรงาน'. Don't use for single-issue triage or PR review.

Catalog of performance-optimizer's 14-row algorithmic-complexity pattern table. Auto-loads when performance-optimizer runs. Don't use for other agents.

Draft a post-mortem for a resolved bug (trigger/mechanism/patch/validation known). Use after mattpocock-skills:diagnosing-bugs; say 'เขียน post-mortem/บันทึกบั๊ก/incident report'. Don't use for in-progress or non-technical incidents.

PR the branch on GitHub, templated body previewed before submit. Trigger on 'open a PR/เปิด PR'. Don't use for merging (`mh:ship-merge`) or review replies (`mh:address-review`).

Scan production readiness pre-launch. Use when asked whether an app is ready to ship. Don't use for in-flight feature work (use /mattpocock-skills:implement).

Cage: human-gated, anti-unattended harness loop. Use when the user asks to improve or audit the harness. Don't use for bug fixes or new surfaces.

Safely identify and remove dead code (JS/TS, Python, Go, Rust) with test verification after each change. Use when cleaning up unused code. Delegates to the refactor-cleaner agent. Don't use for logic refactors beyond dead-code removal.

Catalog of requirement-analyst's self-consistency checklist, Output Format template, and Anti-Patterns list. Auto-loads when requirement-analyst runs. Don't use for other reviewer agents or standalone requirement analysis.

Dispatch 2 independent staff-eng agents to adversarially review skill-creator-style fixture outputs (with_skill vs baseline) for a skill, agent, or command before deciding a fix. Use when fixtures already exist mid an improve+optimize loop. Don't use for PR review (mattpocock-skills:code-review) or skill-creator's own quantitative grading/benchmark step.

Next.js App Router file-convention (error.tsx/loading.tsx/route.ts/parallel routes) and Middleware checklist. Auto-loads when nextjs-reviewer runs. Don't use for caching/Server Actions or standalone review.

Classify a PR's risk as LOW/MEDIUM/HIGH from diff size and sensitive-path signals. Use when scoping review effort. Advisory only — never gates a merge. Don't use for the merge decision (mh:ship-merge).

Score pending decisions on weighted criteria: numeric verdict, pass/fail, confidence, trace. Use when a decision needs a verdict. Don't use for trivial or already-decided choices.

Scan security vulnerabilities, threat-model + remediation (auth, secrets, injection, XSS, traversal). Use when PRs touch auth/APIs/payments/deps. Don't use for quick branch checks or code review.

Catalog of security-reviewer's BAD/GOOD examples (SQLi, IDOR, JWT, mass assignment, SSRF, ReDoS). Auto-loads when security-reviewer runs. Don't use for the deep-audit workflow (security-auditor).

Run AgentShield against agent, hook, MCP, permission, and secret surfaces. Use when auditing harness/agent surfaces for security issues. Don't use for code-vulnerability review — use mh:security-auditor.

Merge a PR safely: validate, server-side merge, cleanup, monitor CI. Use when a reviewed PR is ready to land. Say 'merge PR/รวมโค้ด'. Don't use for failing CI or hotfixes (mh:incident).

Cut a release end-to-end: bump, changelog, review gate, tag, merge, monitor. Use when cutting a release; say 'ship release/ปล่อยเวอร์ชัน'. Don't use for PR merges (mh:ship-merge) or hotfixes (mh:incident).

Catalog of spec-miner's 10 Anti-Patterns FAIL list. Auto-loads when spec-miner runs. Don't use for other agents or standalone spec authoring.

Compress a document, transcript, or pasted text into a BLUF-structured summary. Use when condensing long content for a reader. Delegates to the summarizer agent. Don't use for structured data extraction or code review.

Catalog of summarizer's Output Format templates, word-level compression BAD/GOOD table, and Anti-Patterns list. Auto-loads when summarizer runs. Don't use for other agents or standalone summarization.