Battle-tested skill prompts that make AI coding agents follow through — each skill encodes one hard-won workflow behind Iron Laws and HARD-GATEs so "the agent should do X" becomes "the agent will not proceed without X."
- A copy-paste prompt pack or chat snippet library
- A replacement for your IDE, model, or git workflow
- An engineering archive of plans, audits, or design notes
- A claim that every host adapter is day-to-day tested (only Claude Code, Cursor, Codex, and Grok Build are)
npx @henryavila/atomic-skills installDocs: https://atomic-skills.henryavila.com/ [PRODUCT_END]: #
Build the static product site from a checkout (no network after install):
npm run generate-siteThen open site/dist/index.html in a browser. Details: site/DEPLOY.md.
Tested (real sessions): Claude Code, Cursor, Codex, Grok Build.
Theoretical (install + adapters ship; not day-to-day QA): Gemini CLI, OpenCode, GitHub Copilot, and other profiles.
Support column in the table:
- Tested — install, discovery, invoke, and skill workflows exercised in real agent sessions.
- Theoretical — installer + adapter present; not verified in day-to-day use.
[IDES_TABLE_START]: #
| IDE | Profile | Directory | Format | Support |
|---|---|---|---|---|
| Claude Code | claude-code |
.claude/commands/atomic-skills/ |
Command (slash) | Tested |
| Cursor | cursor |
.cursor/skills/atomic-skills/ |
Markdown | Tested |
| Gemini CLI (Skills) | gemini |
.gemini/skills/atomic-skills-<skill>/ |
Markdown | Theoretical |
| Gemini CLI (Commands) | gemini-commands |
.gemini/commands/ |
TOML (Slash commands) | Theoretical |
| Codex | codex |
.agents/skills/atomic-skills/ |
Markdown | Tested |
| OpenCode | opencode |
.opencode/skills/atomic-skills/ |
Markdown | Theoretical |
| GitHub Copilot | github-copilot |
.github/skills/atomic-skills/ |
Markdown | Theoretical |
| Grok Build | grok |
.grok/plugins/atomic-skills/skills/ |
Markdown | Tested |
| [IDES_TABLE_END]: # |
Compact index (name · one-liner · Iron Law).
Dual docs view (intentional): the product docs site (online or offline site/dist/ via npm run generate-site) is the canonical human product surface (modules, long skill write-ups, project guide). Generated Markdown under docs/skills/ is a secondary offline / agent / GitHub-browsing reference, kept drift-checked by npm run generate-skill-docs (still part of npm run check-docs). Engineering archive paths (docs/kb/, docs/design/, plans, audits) are not product docs and are not published on the site.
[SKILLS_TABLE_START]: #
| Skill | One-liner | Iron Law | |
|---|---|---|---|
| 🔧 | fix |
Diagnose root cause → write test → fix → verify | NO FIX WITHOUT ROOT CAUSE. |
| 💾 | save-and-push |
Scan for secrets, group commits, save learnings, push safely | NO PUSH WITHOUT FRESH VERIFICATION. |
| 🔍 | review-plan |
Adversarial plan review with local/codex/both mode picker | NO APPROVAL WITHOUT EVIDENCE. |
| 🔬 | review-code |
Adversarial code review with local/codex/both mode picker | NO APPROVAL WITHOUT EVIDENCE. |
| 📊 | project |
Plan / Initiative / Task state your agent reloads every session | NO IMPLEMENTATION WITHOUT ANCHORED INITIATIVE. |
| 📝 | prompt |
Generate a self-contained prompt with exact paths and guardrails | NO PROMPT WITHOUT CODEBASE ANALYSIS. |
| 🎯 | hunt |
Adversarial tests from the spec, not the code — depth over breadth | NO HUNT WITHOUT BOUNDED SCOPE. |
| 🚀 | parallel-dispatch |
Dispatch a task list to N parallel sessions with verified isolation | NO LAUNCH WITHOUT MECHANICAL SCOPE ISOLATION. |
| 👁️ | parallel-dispatch-audit |
Verify each batch deliverable on disk; fix or escalate with evidence | NO CONCLUSION WITHOUT EVIDENCE FROM DISK. |
| 💡 | brainstorm |
Diverge, decide, then write a critic-gated design.md before any plan | NO PLAN WITHOUT AN APPROVED DESIGN. |
| 🎨 | design-brief |
Generate DS + screens prompts for a design agent, contamination-free | NEVER SILENCE BEHAVIOUR OR PHILOSOPHY — SILENCE IS FOR VISUAL FORM ONLY. |
| 🎭 | debate |
Roundtable of independent subagent personas for divergent thinking | NO SYNTHESIS WITHOUT INDEPENDENT VOICES. |
| ⚙️ | implement |
Drive plan tasks to done; optional --mode=automate pure maestro | CODING STAYS SINGLE-THREADED (ONE WRITER PER WORKTREE). |
| ✅ | verify-claim |
No success claim without fresh verification — run it, cite it | NO SUCCESS CLAIM WITHOUT FRESH VERIFICATION. |
| 🧠 | init-memory |
Consolidate scattered memory into .ai/memory/ and wire it to the IDE | NO DELETION WITHOUT CONFIRMED BACKUP. |
| [SKILLS_TABLE_END]: # |