A collection of agent skills for Claude Code, Codex, Copilot, and other AI coding assistants.
Note
These skills are designed to be useful across agents. Unknown frontmatter fields (like disable-model-invocation) are ignored per the agentskills.io standard, so Claude-Code-specific fields are safe to include in any compliant agent.
| Skill | Description |
|---|---|
| pressure-test | Interviews you relentlessly about a plan, branch by branch, and ends with a consolidated decision artifact you can optionally save to a .decisions.md file for to-plan. |
| to-plan | Synthesizes a decisions record and a repo read into a pure implementation plan on disk — file paths, modules, phase-gated slices, and anchored verification — shaped to be hardened by plan-clash. |
| plan-clash | Hardens an existing implementation plan by looping a Codex (GPT) critic and a Claude (opus) adjudicator over it for N rounds — cross-model, fresh-eyes review that improves the plan against the real code. Runs on CLI subscriptions; improves a plan, never certifies one. |
| commit | Project-agnostic git commit workflow — triages secrets, runs the project's test/lint gate, groups changes, and writes conventional commits after you approve. Tailorable per project. |
| next-slice | Closes out a just-finished implementation slice and hands the next off to a fresh session — gates that the slice landed (clean tree, green build, committed), mines what the build revealed that reshapes the next slice, re-anchors the plan's stale references against current code, and emits a self-contained handoff prompt. |
Important
A skill is instructions your agent runs on your behalf — anyone can publish one, and GitHub doesn't verify them. Inspect before you install: gh skill preview bitcoin21ideas/skills <name>, or read the skill's SKILL.md on GitHub. The skills here are MIT-licensed and meant to be auditable; apply the same caution to skills from anywhere.
Three ways, easiest first: the GitHub CLI (one command, cross-agent), ask your agent, or by hand.
gh skill (GitHub CLI ≥ 2.90.0) installs a skill into whichever agent you use — Claude Code, Copilot, Cursor, Codex, or Gemini CLI — resolving the right location for you and recording where it came from:
gh skill preview bitcoin21ideas/skills pressure-test # inspect first
gh skill install bitcoin21ideas/skills pressure-test # then installSwap pressure-test for any skill in the table above.
No CLI? Hand your agent the skill's folder URL and ask it to install:
I like this skill — install it in my project: https://github.com/bitcoin21ideas/skills/tree/main/skills/pressure-test
This needs an agent that can fetch a URL — Claude Code and Gemini CLI do by default; Cursor and Copilot may need web access turned on; Codex CLI needs a fetch tool (MCP) wired up. Tell it two things: a skill is a multi-file folder (not one file), and read the raw files (raw.githubusercontent.com), not the rendered blob page. On a model with no skills runtime (a bare Kimi / OpenAI-compatible endpoint), just paste SKILL.md into your system prompt.
git clone https://github.com/bitcoin21ideas/skills.git
cd skillsDrop a skill folder where your agent looks. .agents/skills/ is the vendor-neutral location honored by Codex, Cursor, Gemini, and Copilot:
ln -s "$PWD/skills/pressure-test" .agents/skills/pressure-test # project-level
ln -s "$PWD/skills/pressure-test" ~/.agents/skills/pressure-test # user-levelPrefer an agent-specific home?
| Agent | Project path | User path | Invoke |
|---|---|---|---|
| Claude Code | .claude/skills/ |
~/.claude/skills/ |
/pressure-test |
| Codex CLI | .agents/skills/ |
~/.agents/skills/ |
$pressure-test |
| Cursor | .cursor/skills/ |
~/.cursor/skills/ |
/ menu |
| Copilot (VS Code) | .github/skills/ |
~/.copilot/skills/ |
/pressure-test |
| Gemini CLI | .gemini/skills/ |
~/.gemini/skills/ |
/skills |
Most agents also auto-load a skill when its description matches, so explicit invocation is optional.
As of June 2026 the Agent Skills standard is natively supported across the major coding agents (Claude Code, Codex, Cursor, Copilot, Gemini CLI, and dozens more), so these skills are portable without per-agent rewrites.
The disable-model-invocation: true frontmatter field prevents Claude Code and Copilot from auto-invoking a skill based on its description — the user must call it explicitly. Agents that don't recognize the field ignore it safely, falling back to description-based triggering.
Invocation mode is chosen per skill:
pressure-testis explicit-only (disable-model-invocation: true) — you rarely want to be auto-grilled.plan-clashis explicit-only (disable-model-invocation: true) — each run spawns long, subscription-billed CLI subprocesses, so it should fire only when you ask for it.commitis both user- and model-invocable (no such field) — natural-language requests like "commit this" can trigger it, and it always proposes a plan and waits for approval before committing, so the model can never make a surprise commit.
Either way, skill descriptions are phrased with explicit user-action triggers ("when the user says...") so even agents that ignore the field don't fire them spuriously.
This repo is agent-first. The easiest way to add or fix a skill is to ask your coding agent: point it at AGENTS.md, describe the change, and have it open a PR. See AGENTS.md for the layout and conventions. Issues and PRs welcome — from an agent or by hand.
MIT — see LICENSE.
Commits in this repo are made with its own
commitskill.