AI behavior packs that coding agents can load for specialized roles — spawning a dialectic council to debate an implementation, or a Playwright tester that exercises every feature of an app without writing a single assertion. Skills are closer to workflows than prompt templates: they define phases, roles, and completion signals, not just what to say.
First, install the skill CLI to your system:
./install.shThis copies skill to a directory on your PATH so you can use it from anywhere.
skill install-all # install every skill in this repo
skill install <path|url> # install one skill from a local dir or git URL
skill remove <name> # remove one
skill list # list installed skills
skill update <name> # pull latest for git-cloned skillsSee INSTALL.md for Python and system requirements.
Every skill declares when it activates and when it stays silent:
description: Does X. TRIGGER on: "audit my design", "fix the UI".
SKIP on: "performance issues", "add a backend endpoint".When your message matches a TRIGGER phrase, the agent loads that skill. When it matches a SKIP phrase, the agent ignores it. That's the idea.
You: "audit my app's design"
Agent: [loads frontend-ux-designer skill]
Agent: "Evaluating layout consistency, typography, color, spacing..."
Orchestration — skills that spawn and coordinate multiple sub-agents:
| Skill | What it does |
|---|---|
| implementer | Implement → blind Skeptic pass + scored Review → Fix loop until 10/10 quality |
| one-shot-project | Full project builder: Architect, PM, parallel implementers, tester |
| code-smellz | Parallel bug-hunter, simplifier, optimizer, and security auditor |
| brainstormers | Dialectical thinker sub-agents debate ideas until convergence |
| dialectic-council | Multi-model debate: Claude + DeepSeek, OpenAI, Gemini, MiniMax |
Quality — skills that demand excellence through adversarial refinement:
| Skill | What it does |
|---|---|
| student-counsel | Student works, philosophers review in dialectic rounds until consensus |
| look-for-flaws | Submit work to a dialectic of sub-agents who find flaws in the codebase |
Frontend & Testing — browser-based QA and design auditing:
| Skill | What it does |
|---|---|
| frontend-twerkin | Starts the app, tests every feature with Playwright, auto-fixes failures |
| frontend-ux-designer | Design audit: layout, typography, color, spacing; produces critique + fixes |
Knowledge & Reference — skills that bring deep domain expertise:
| Skill | What it does |
|---|---|
| teach-me | Learn any codebase with Mermaid diagrams |
| rust-evm | EVM internals: revm, Foundry, bytecode, Yul, gas optimization |
| prediction_markets | Polymarket and Kalshi: APIs, CLOB trading, auth, order types |
| web-scraping | Playwright, httpx, anti-bot evasion, pagination, structured extraction |
Meta — skills that create, curate, and constrain other skills:
| Skill | What it does |
|---|---|
| skill-creator | Create, test, and publish new skills |
| stakeholder-of-last-resort | Speaks for absent parties facing foreseeable harm in decisions |
A few things I've found useful when building skills:
Composable. Skills can call skills. Every skill emits a
DONE:completion signal so orchestrators know when a sub-agent has finished and what it produced. One skill can spawn another —code-smellzdispatches four parallel auditors, each of which is itself a skill-like role.
Dialectic. Structured disagreement often surfaces what agreement misses. Several skills use adversarial review — sub-agents that genuinely argue with each other — to surface flaws a single pass would miss.
student-counselwon't accept work until a council of philosophers reaches consensus that it is beautiful.implementerincludes a blind Skeptic pass that reviews the output without seeing the original question — flaws that survive context collapse under it.
Self-critical. Several skills —
skill-creator,code-smellz,student-counsel— include mechanisms for revising not just the output but the process itself. A skill that can't improve itself misses its own potential.
Skills are SKILL.md files with YAML frontmatter that declare their name,
description, and activation contract:
---
name: my-skill
description: What it does. TRIGGER on: "these phrases". SKIP on: "those phrases".
---
## Phase 0: Understand the problem
## Phase 1: Do the work
## Phase 2: Verify
DONE: <output> — <summary>To create one properly, use the skill that knows how:
Use the skill-creator skill
Or read skills/skill-creator/SKILL.md directly.
MIT License. See LICENSE.