Agent-Skills

Review changed code for behavior-preserving simplification by removing dead code, eliminating duplication, extracting shared helpers, improving names, and tightening tests. Use when a user asks for code review, refactor, clean up PR, simplify, tidy up code, review my changes, or maintainability cleanup before merge.

Check if a repository or agent-facing product surface is ready for AI coding agents. Use when you need to audit repo agent compatibility, review AGENTS.md, find missing test/build commands, evaluate docs quality, assess tool/action parity, or produce a file-evidence scorecard with specific fixes.

Use when reviewing, creating, shrinking, or refactoring AGENTS.md and directly linked instruction guidance that need scoped routing, deduplication, contradiction resolution, or progressive disclosure.

Review code architecture, code quality, dependency graphs, coupling, technical debt, modularization, ownership, and test seams. Use when refactors, restructuring, tightly coupled code, or architecture decisions need proof-backed options.

Choose validation proof for tests, CI, coverage, evals, and closeout evidence: map changed files to repo-native commands, classify pass/fail/blocked ownership, preserve trace/regression artifacts, and keep local, CI, Tessl, external-review, tracker, and runtime truth separate. Use when users ask what tests to run, why validation failed, what proof is enough, or whether command evidence supports a claim.

Defensive review of AI-agent skills, plugins, and tools using Liran Tal's security principles. Use when assessing provenance, permissions, data exposure, sandboxing, or approval boundaries before adoption.

Creates release-ready skill packages from proven workflows by drafting SKILL.md frontmatter, trigger-rich instructions, references/contract.yaml, eval scenarios, task profile, and validation commands. Use when the user asks to create a skill, skillify a workflow, package a process, write SKILL.md, make a reusable skill template, or prepare a skill for release checks.

Run structured AI code review as an advisory closeout gate for local diffs, PR branches, or commits when the user asks for autoreview, Codex review, second-model review, or pre-ship validation.

Use when evaluating LLM or RAG outputs: audit eval coverage, analyze failed traces, write binary judge prompts, validate judges against labels, generate targeted synthetic cases, evaluate retrieval quality, or plan review tooling. Do not use for ordinary software test implementation.

Automate until-green PR review, CI, merge, and cleanup follow-through. Use when open project PRs need GitHub, CodeRabbit, CircleCI, Context7, autofix, heartbeat, and branch/worktree pruning.

Builds a project glossary that maps everyday wording to canonical terms, repository actions, and reusable engineering rules. Use when a user asks "what does X mean here?", "define our terms", "standardize this naming", "turn this phrase into a repo action", "grill the domain language", or whether a local correction should apply to similar code.

Assists with questions about Guy Podjarny's talk "Skills are the new Code". Use when the user wants to understand, apply, audit, or explore frameworks from this keynote — including the five engineering disciplines for skills (static analysis, evals, security testing, dependency management, observability), the three challenge buckets, the agentic development stack, or concepts like skill authoring, context engineering, agent harnesses, and skill quality scoring.

Reviews and improves SKILL.md packages by fixing audit findings, triggers, examples, evals, token budget, release proof, safety verdicts, comparator/baseline choices, and bounded code-lens hardening. Use when the user says improve a skill, fix a skill file, review SKILL.md, raise Tessl score, reduce context cost, add skill evals, or prepare a plugin skill for release.

Use when a Codex goal/task is stuck, hanging, not finishing, or needs status. Reads goal.md, state.yaml, receipts.jsonl; syncs reported status with board files; fixes invalid state.yaml; classifies blockers; decides done. Not for ordinary reviews or one-off fixes.

Create, review, and maintain gold-standard Skills SDK eval scenarios before internal evals, dry Tessl staging, or live private Tessl scoring. Use when creating or updating a skill, writing skill tests, adding eval cases, importing KnowledgeOS or Tessl suggestions, checking scenario drift, or hardening evals that are too easy.

Audit, rewrite, and validate README, runbook, code-doc, config-doc, package-evidence, and public trust-surface documentation against live repository evidence. Use when documentation needs proof-backed correction, reader-focused review, generated-document ownership, public-access checks, or legacy docs-expert routing.

Improve existing Harness Engineering skills, references, contracts, and evals from concrete evidence such as failed evals, repeated review findings, usage traces, or documented regressions. Use when a bounded hardening pass is required; do not use for speculative redesign.

Create evidence-backed HE reframe migration programs. Use when structural drift, routing ambiguity, or source-prompt gaps need phased rollback-safe execution.

Compress HE cognition artifacts into evidence-backed strategy. Use when intent, review, triage, ADR, core, or source-prompt comparison evidence needs durable direction.

Create, validate, install, fold, or troubleshoot Codex subagent role TOML, agents-table config, discoverability wiring, and duplicate-role merges. Use when a user asks for a Codex agent role, reviewer agent, role config, TOML role file, subagent setup, or overlapping agents to merge.

Binary calibration pack for recurring workflow guardrail recommendations.

Analyzes Codex skill-management requests, selects the workflow lane, and returns selected_lane, mode, next_step, and blockers. Use when the user says create a skill, add/update/fix/review a skill, install/sync/list skills, choose a workflow, or merge/retire a skill.

Plan AI-native hackathon scope, demo pitch, and judge Q&A. Use when the user needs 1-hour pitch prep, 24-hour build scoping, AI Native DevCon-style track choice, idea pressure tests, or a spec-led hack plan. Do not use for code generation, implementation support, or non-hackathon DevRel content.

Analyzes bounded skill evidence, classifies root causes, and recommends a lifecycle lane such as keep, observe, improve through Skill Factory hardening, merge with approval, or retire with approval. Use when a skill is not working, a skill is not triggering correctly, evals or Tessl disagree, repeated failures need debugging, or skill performance issues need evidence-backed repair handoff items.

Use when creating or updating a Codex skill package with focused triggers, progressive disclosure, evals, evidence, or tool integration.

Use when listing or installing Codex skills from curated sources, GitHub repo paths, private repos, or local package locations.

Fixture skill with promotable eval scenario metadata.

Turn Slack #code-fixes, CodeRabbit, Codex Review, CI, and check-status noise into a repo-and-PR action queue. Use when Jamie asks for a daily code-fixes digest, recent review-noise triage, or what needs fixing across active engineering repos.

Analyze recent Codex session evidence for repeated manual workflows and route them to skills, subagents, validators, or no artifact when Jamie asks what he keeps doing manually.

Create mission-grounded study plans, lessons, quizzes, references, resources, and learning records. Use when the user asks to learn, study, be taught, continue a course, build a curriculum, or maintain a multi-session teaching workspace.

Minimal valid Codex skill source fixture for JSC-391 scaffold tests.

Ship skill changes to PRs when Codex skills need source edits, projection sync, strict audit, reviewer evidence, commit, push, and PR status.

Use when hardening, converting, auditing, or pre-release checking a Codex plugin package by verifying manifest paths, bundled skills, hooks, MCP/app config, validation gates, and release blockers.

Apply approved fixes for unresolved CodeRabbit review comments, Codex P1-P3 findings, PR feedback, and code review issues with validation evidence. Use when asked to address review comments, fix review findings, clear unresolved comments, or autofix PR feedback.

Run bounded automated experiment iterations by recording baselines, applying hypothesis patches, comparing metrics, protecting regression guards, and deciding keep, discard, rollback, or block. Use when automated research is requested or a repo/skill needs evidence-backed research, metric tracking, or safe optimisation loops.

Create, diagnose, and validate a local dev bootstrap. Use when the user asks to clone a repo, install toolchains, install dependencies, and prove the project runs.

Use when designing, reviewing, or updating Codex app automations, cron jobs, scheduled tasks, recurring runs, or heartbeat follow-ups.

Scaffold hook packs, validate hooks.json schema, verify hook script permissions, migrate hook configuration, and troubleshoot Codex hook execution errors. Use when creating, auditing, upgrading, or validating Codex hook packs, hooks.json files, hook scripts, SubagentStart/SubagentStop lifecycle hooks, PreToolUse/PostToolUse/PreCompact hooks, Stop claim checks, or repo-local/user-level .codex hook installs.

Review local dirty changes, committed branches, and PR diffs with Codex CLI; report findings, validation, blockers, and merge-readiness evidence. Use when the user asks for Codex review, autoreview, independent model review, or pre-ship validation.

Use when users need to install, bootstrap, upgrade, audit, diagnose, or explain @brainwav/coding-harness in a repository, including harness init/upgrade, CI migration, governance gates, command discovery, and Codex environment action sync.