Agent Skills

A personal collection of agent workflows I am building and maintaining for the way I plan, build, review, and ship software.

These are opinionated on purpose. They favor evidence from the repository, durable decisions over conversation memory, explicit human gates, and output another agent can pick up cold. I keep changing them as I learn what actually works.

Feel free to use them, adapt them, fork them, or pull out the pieces that fit your own workflow.

The collection

Skill	What it does
⚔️ slice	Shapes a rough idea into a fully grilled, acceptance-driven vertical slice and a cold-start implementation handoff.
grill	Stress-tests a plan against the project's domain language and durable architecture decisions.
uat	Writes a concise, executable manual acceptance plan around observable user behavior and demo readiness.
prd	Builds a product requirements document through interview, codebase exploration, and module design.
review	Reviews diffs for correctness, architecture, security, performance, maintainability, and testing risks.
threat-model	First-principles threat model — source→sink paths, ranked risks, and controls — strongest on hard-to-reverse surfaces: money, secrets, access, agent actions, deletes, publishing.
pr	Derives PR copy from the final diff and triages review feedback into address, defer, or dismiss.
typescript-cli	Builds, migrates, and reviews reliable, distributable Node.js CLIs in TypeScript.
product-website	Builds and audits clear, credible, conversion-focused product marketing sites.

Each directory has its own README for humans and a SKILL.md containing the actual agent workflow.

Using a skill

Copy or link the skill directory into whichever skills location your agent uses, then invoke it by name—for example:

Use $review to check this branch before I merge it.

The descriptions in SKILL.md are written to support automatic triggering too, when the agent environment provides it.

Anatomy

skill-name/
├── README.md            # Human-facing overview
├── SKILL.md             # Trigger metadata and executable workflow
├── agents/openai.yaml   # Optional UI metadata
├── references/          # Detailed guidance loaded only when needed
├── scripts/             # Repeatable deterministic helpers
└── assets/              # Templates and output resources

Not every skill needs every directory. The core workflow stays in SKILL.md; supporting material exists only when it earns its keep.

Building these

• Keep the workflow portable and discover project-specific rules from the repository instead of baking in one application's architecture.
• Prefer code, diffs, tests, and durable artifacts over conversation memory.
• Make write boundaries and human approval gates explicit.
• Use progressive disclosure so specialized guidance does not crowd every task.
• Exercise skills on real work and revise them when reality finds the weak spot.