Skills & agents
38 skills and four subagents. Skills stay out of context until a task calls for them, so the always-on footprint stays small.
Updated
The flow
grill → plan → execute → verify → review → ship. Bugs start at debug. In Claude Code, call a skill with /honey:<skill>; in Codex, use $<skill>.
Chosen for you
The agent picks these itself when the situation matches.
| Skill | Use |
|---|---|
grill | Pressure-test an idea when a product decision is missing. |
plan | Scope multi-file or multi-session work into a plan, spec or tickets. |
execute | Work through a plan, ticket set or TODO list continuously. |
debug | Find the root cause of a failure before touching the fix. |
tdd | Build or fix test-first with a red-green loop. |
verify | Run the checks that prove a change before calling it done. |
review | Review diffs, branches and PRs; handle review feedback. |
ship | Commit, push, open PRs, resolve conflicts, finish branches. |
deploy | Cloud and PaaS deploys, containers, IaC and CI/CD. |
test-design | Design the test matrix and hard edge cases before writing tests. |
tooling | Monorepos, lint, format, typecheck, hooks and CI wiring. |
database | Schemas, online migrations, indexes, N+1 and concurrency. |
perf | Profile the real bottleneck and prove the speedup with numbers. |
observability | Logs, metrics, tracing, SLOs and alerting, OpenTelemetry-first. |
api-design | REST, GraphQL, gRPC contracts, versioning, errors, pagination. |
upgrade | Dependency bumps, CVE patches, breaking changes and EOL runtimes. |
incident | Mitigate first, investigate safely, write the postmortem. |
frontend | UI, layout, components, responsive and React or Next.js work. |
stageshow | Mocked product illustrations and animated cursor demos. |
qa | Check a change in a real browser with visual proof. |
security | Auth, secrets, payments, user data, webhooks and audits. |
docs | READMEs, changelogs, ADRs and PR text that match the change. |
Yours to call
These run only when you invoke them, so their descriptions never expand the always-loaded workflow.
| Skill | Use |
|---|---|
honey | Pick the right skill or flow, or switch response modes. |
rsi | Review what it learned, coach, save or forget learnings. |
unslop | Cut AI tells and filler from any text. |
reverse-engineer | Study software you are authorized to analyse, with evidence. |
handoff | Write or restore a handoff so a fresh session can continue. |
prototype | Throwaway prototypes that settle a design question. |
architecture | Deepening opportunities and behaviour-preserving refactors. |
domain | Sharpen domain terms; record decisions in CONTEXT.md and ADRs. |
teach | Explain a concept, or how and why some code works. |
write-skill | Write or patch a skill to the quality bar, with evals. |
principles | Engineering principles index, read before citing. |
opensrc | Library context from package.json and official docs. |
project-verify | A project-local verification harness and feature map. |
triage | Qualify issues, PRs and suggestions into actionable briefs. |
research | Answer questions from primary sources with citations. |
statechart | Model a stateful flow and check its invariants. |
Subagents
Four focused agents handle delegated work with their own tools and reports.
| Agent | Role |
|---|---|
planner | Turns a request into a scoped plan in honey plan format. Read-only. |
implementer | Implements one scoped task with focused tests and a DONE / BLOCKED report. |
reviewer | Checks spec compliance, then code quality, before commit or PR. Read-only. |
verifier | Chooses and runs the checks that prove a change, and reports risk. |
Workflow highlights
project-verify creates or maintains a project-local verification skill — launch, readiness checks, feature journeys, observable side effects and safe cleanup. A generated harness counts as ready only after its recipe runs successfully. verify and qa reuse it.
triage reads complete reports, checks duplicates and existing fixes, and produces evidence-backed states and agent briefs. Its reproduce-and-fix branch requires before/after proof and delivers through ship.
research answers technical and product questions from primary sources, with constraints, source dates, citations and unknowns. statechart checks queue, retry and workflow invariants with a model checker and explains counterexamples — a checked model is not proof of the production code.
qa uses the tester-army/e2e CLI and project MCP server by default, and discovers existing WebMCP tools for setup and contract checks. Real interactions, visual checks and backend assertions still prove their own requirements. The fast QA path schedules independent flows across bounded workers while keeping failure screenshots.