This repository is my personal collection of agent skills — self-contained SKILL.md files that teach AI agents a specific workflow, convention, or behavior. They are:

• Bounded. One skill = one job. No kitchen-sink prompts.
• Composable. Skills stack. A feature-manager roadmap can be reviewed by spec-review; self-healing runs underneath every task.
• Harness-agnostic. Any agent that loads SKILL.md can use them.
• Installable everywhere. A single shell script publishes them globally and to every git repo on your machine.

Think of this repo as a toolbox. Pull only the SKILL.md you need; ignore the rest.

• trump makes sure the agent never loses sight of what matters.
• self-healing makes sure the agent gets better every time it runs.
• spec-review makes sure the agent does what you actually meant.
• feature-manager makes sure big ideas become small, shippable tasks.

Every skill in this repo follows the same minimal shape:

skills/<name>/
├── SKILL.md            # the only required file
├── evals/              # (optional) trigger & behavior evals
└── references/         # (optional) deeper theory, tables, history

SKILL.md opens with a YAML frontmatter block that drives discovery:

---
name: skill-name
description: >
  Plain-language trigger conditions. When should this skill fire?
  Which phrases, file types, and user intents should activate it?
---

A good skill is specific about when to fire and rigorous about how to behave. The four in this repo model that.

A single script installs every skill globally and seeds it into every git repo under ~/src/.

git clone https://github.com/blkst8/skills.git ~/src/github/blkst8/skills
cd ~/src/github/blkst8/skills
./install-default-agents.sh

What it does:

1. Copies every skills/*/ directory into ~/.agents/skills/.
2. Walks ~/src/ and copies ~/.agents/skills/ into each repo's .agents/.
3. Installs a post-checkout git hook so freshly-cloned repos auto-populate .agents/.

Tip: Override TARGET_DIR to seed a different root, or just symlink ~/.agents/skills/ to skills/ for live updates.

Want only one skill? Copy a single directory:

cp -r skills/feature-manager ~/.agents/skills/

Or symlink it from a checkout:

ln -s "$(pwd)/skills/trump" ~/.agents/skills/trump

These are skills I use daily that live elsewhere. See other-skills.md for install snippets.

• ast-grep — structural code search for refactors and codemods.
• caveman — token-saving compressed communication mode.
• goalbuddy — structured goal intake for long-running Claude Code work.

Long-form thinking that backs the skills lives in specs/:

The fastest path:

1. Copy any skills/<name>/ directory.
2. Rewrite the frontmatter description to be specific about trigger conditions.
3. Write the body in three parts: When to use / When NOT to use / Algorithm.
4. Add at least one eval under evals/ to lock the trigger phrase.
5. Open a PR.

The skills in this repo are the same shape you'd ship.

Issues and PRs welcome — but read spec-review first. It will save you a round trip.

MIT — do what you want, just keep the attribution.