skills-lab

A library of reusable Agent Skills for Claude Code (and compatible agent harnesses) — plus the authoring principles behind them.

Each skill is a self-contained, model-invoked capability: a SKILL.md with YAML frontmatter that tells the agent when to reach for it and how to use it, sometimes paired with a small zero-dependency script. Every skill here is generic and portable — no proprietary tooling, no company-internal systems, nothing you can't run on a laptop.

This repo is as much about how to write good skills as it is about the skills themselves. See docs/authoring-skills.md.

Skills

Research & information

Skill	What it does
deep-research	Decompose a question, fan out searches across web/code/docs, read the best sources, refine on gaps, and synthesize a cited answer.
researching-ai-startups	Research an AI startup category or one company: explain the product, identify founders, collect interviews and transcripts, separate claims from evidence, and organize reusable company notes.
fact-check	Verify a claim against multiple independent sources and return a structured verdict with confidence and citations.
search-strategy	Search any corpus effectively — pick the right source, craft precise queries, iterate, filter, and switch keyword↔semantic.
web-search	DuckDuckGo search from the terminal — zero-dependency Node script, text or JSON output.
web-fetch	Fetch a page and extract readable text (or raw HTML, or all links) — zero-dependency Node script.

Code, review & delivery

Skill	What it does
git-workflow	Feature-branch git flow, conventional commits, rebasing, conflict resolution, and PRs via the gh CLI.
code-review	Understand intent → review against an 8-lens rubric → post actionable, severity-labeled inline comments.
design-review	Review a technical design (or defend your own) against a rubric: problem, alternatives, failure modes, operability, cost.
call-chain-tracing	Trace a request/value hop-by-hop across services and code to find where it's transformed, filtered, or dropped.
test-coverage-analysis	Interpret coverage metrics and CI test-run failures; prioritize the gaps that actually matter (risk-based, not 100%-chasing).
ci-cd-triage	Diagnose red pipelines — classify the failing stage, read logs first-error-first, triage build/test/deploy/infra failures.

Operations & infrastructure

Skill	What it does
tmux-runner	Run builds, servers, watchers, and parallel agent tasks in background tmux sessions without blocking the main agent.
aws-cli-safety	Safe, effective AWS CLI use — read-before-write, least privilege, prod caution, --dry-run, --query/jq, pagination.
change-management	Author and run production change records: what/why/blast-radius/rollback/verification, risk tiers, approval flow.
operational-readiness-review	Run or prepare for a launch-readiness review: monitoring, runbooks, on-call, capacity, rollback, durability, load testing.
security-review-gate	Pause on security-sensitive requests (secrets, PII, auth, IAM), consult the standard before acting, never emit insecure code.

Oncall & incident response — a coordinated set; oncall-agent is the orchestrator that composes the rest into one on-call assistant (see Composing skills into an agent).

Skill	What it does
oncall-agent	Orchestrator — routes each situation (shift start, triage, alarm, live incident, handoff) to the right specialist skill below, coordinating them through shared context files.
oncall-context-system	A file-based shift-memory system so oncall context never evaporates — ticket ledger, deadline calendar, working-on tracker, gaps log, daily logs, and handoff lifecycle.
ticket-triage	Categorize, prioritize, and destale a ticket queue; investigate 3+ tickets in parallel with read-only sub-agents, then reconcile in one pass.
alarm-correlation	Cluster related alarms, assess systemic-vs-transient severity, and trace cross-service dependency chains to a root cause.
incident-response	Handle a live sev1/sev2 — impact quantification, primary/secondary linking, the resolution-summary gate, and outage comms.
runbook-authoring	Write agent-consumable service runbooks: alarm-definition blocks, ranked-likelihood investigation tables, quality tiers, and agent-routing metadata.
postmortem	Write and facilitate a blameless postmortem — timeline, impact, root cause, and concrete preventive action items.
evidence-discipline	Ground every operational claim in a traceable source, label its evidence tier, and never present correlation as causation.

Writing & communication

Skill	What it does
technical-writing	Write clear long-form docs (design docs, proposals, narratives) — structure, lead with the point, data over adjectives, self-review pass.
document-review	Review someone else's prose: does it answer its question, is the structure sound, what's missing — with severity-labeled comments.
writing-style-transfer	Extract and reproduce a person's or brand's writing voice from samples, via a reusable style card.
diagrams-as-code	Generate clean architecture diagrams as code (Mermaid / Graphviz / PlantUML / SVG) — with a tool decision table.
html-notes	Produce polished, self-contained HTML reading layers for specs, reports, research notes, diagrams, and long-form wiki notes.

Thinking, judgment & meta

Skill	What it does
socratic-questioning	Ask focused questions to reach shared understanding before acting — with a grill-vs-act decision model.
engineering-judgment	Operate at senior/staff/principal level: scope, ambiguity, influence without authority, and when "good enough" wins.
behavioral-interview	Run or prepare for competency/behavioral interviews — STAR, competency areas, probing for real signal, level calibration.
technical-interview	Run or prepare for coding and system-design interviews — round types, calibration rubric, interviewer prep doc, debrief writing, candidate coaching.
terminal-recording	Record polished terminal demos via asciinema + tmux, with idle-time compression and post-hoc trimming.
agent-retrospective	An agent analyzing its own session transcripts to find friction and missed corrections, then proposing durable improvements.

Composing skills into an agent

Skills aren't only standalone — several here are designed to compose into a single agent. The oncall set is the worked example: seven specialist skills (oncall-context-system, ticket-triage, alarm-correlation, incident-response, runbook-authoring, postmortem, evidence-discipline) that share one set of context files, coordinated by an orchestrator (oncall-agent).

The pattern generalizes to any multi-skill agent:

1. Specialists do one thing well. Each skill is independently useful and independently testable.
2. An orchestrator skill routes situations to specialists (a situation → skill table) and documents the phase order for an end-to-end flow.
3. Shared state is a contract. The orchestrator defines who owns the shared files — e.g. one writer (only the main agent writes the ledger; parallel sub-agents are read-only), journal ≠ source of truth, and cross-cutting rules like evidence discipline that every specialist obeys.
4. The orchestrator's description names the whole domain, so the agent loads it as the entry point and pulls in specialists on demand.

Install the whole set together and load the orchestrator first — see oncall-agent for the routing table, the shift lifecycle, and how to point the specialists at your ticketing / monitoring / CI-CD / chat stack.

Installing

Skills live in a skills/ directory that your agent reads. For Claude Code, symlink (or copy) any skill into your skills directory:

git clone https://github.com/jingchaodev/skills-lab.git
cd skills-lab

# Install a single skill for your user
ln -s "$PWD/skills/tmux-runner" ~/.claude/skills/tmux-runner

# ...or install them all
for d in skills/*/; do
  ln -s "$PWD/$d" ~/.claude/skills/"$(basename "$d")"
done

Then the agent picks them up automatically — it reads each SKILL.md frontmatter and invokes the skill when the description matches the task.

Requirements

Most skills are pure Markdown and shell. The ones with extra needs:

• web-search / web-fetch — Node.js 18+ (built-in global fetch; no npm install).
• terminal-recording — asciinema, tmux, python3, optionally agg for GIF export.
• tmux-runner — tmux.
• git-workflow / code-review — git and the gh CLI.
• aws-cli-safety — the AWS CLI.
• diagrams-as-code — any of Mermaid, Graphviz, or PlantUML (your choice).

Anatomy of a skill

skills/web-search/
├── SKILL.md        # frontmatter (name, description) + usage docs the agent reads
└── search.js       # optional: a self-contained script the skill drives

The description field is the most important line in the file — it's the only thing the agent sees when deciding whether a skill is relevant, so it names the concrete trigger conditions. See docs/authoring-skills.md for the full rationale.

License

MIT. Use them, fork them, adapt them.