Optimized prompts you install once and invoke in any AI IDE. Each skill is an atom — small enough to stay focused, specific enough to leave no ambiguity, capable enough to make the agent actually follow through.

Stop rewriting prompts.

npx @henryavila/atomic-skills install

Why Atomic?

AI agents skip steps, rationalize shortcuts, and ignore vague instructions. Atomic Skills solve this with battle-tested techniques baked into every prompt:

• Small — one skill, one job. No bloat, no dependencies between skills
• Specific — every step names the tool, demands evidence, defines what "done" looks like
• Capable — Iron Laws, HARD-GATEs, Red Flags, Rationalization tables. Techniques that turn "the agent should do X" into "the agent will do X"

Multi-Agent Optimization

Atomic Skills uses a Polyglot Rendering Engine that detects your agent and optimizes tool naming and instructions automatically.

• Claude Code: Native support for Bash, Read tool, Edit tool, and Agent.
• Gemini CLI: Native support for run_shell_command, read_file, replace, and codebase_investigator.
• Generic/Others: Standardized naming for maximum compatibility.

Supported IDEs

| IDE | Profile | Directory | Format | |-----|---------|-----------|--------| | Claude Code | claude-code | .claude/skills/ | Markdown | | Gemini CLI | gemini | .gemini/skills/ | Markdown (Recommended) | | Gemini CLI | gemini-commands| .gemini/commands/ | TOML (Slash commands) | | Cursor | cursor | .cursor/skills/ | Markdown | | Codex | codex | .agents/skills/ | Markdown | | OpenCode | opencode | .opencode/skills/ | Markdown | | GitHub Copilot | github-copilot| .github/skills/ | Markdown |

Skills

Fix & Hunt — the bug-killing duo

| Skill | What it does | |-------|-------------| | as-fix | Diagnoses root cause with evidence, enumerates the test surface (equivalence partitions, boundaries, error inputs), writes a test cluster with TDD, fixes with minimal change. Integrates mental mutation spot-checks to verify no case was missed | | as-hunt | Writes adversarial tests for existing code that lacks coverage. Accepts a file, function, or directory. Single-file: 6-phase deep hunt. Directory: triages by risk, spawns isolated subagents per file (prevents cross-file tautology). Bugs found are handed off to as-fix with the reproducing test already written |

Plan & Review

| Skill | What it does | |-------|-------------| | as-prompt | Generates an optimized, self-contained prompt from a task description — explores codebase, resolves file paths, applies Iron Law, Red Flags, and task-specific Rationalization table | | as-review-plan-internal | Adversarial review of a plan — finds contradictions, broken dependencies, ambiguity. Verifies file/command existence with Glob, not trust | | as-review-plan-vs-artifacts | Cross-references plan against PRD, specs, and artifacts. Requires line numbers from BOTH documents as proof | | as-status | Tracks the current workstream, completed work, remaining work, and stage-level reviews/verifications |

Session Management

| Skill | What it does | |-------|-------------| | as-resume | Investigates project context (git, memory, instruction files), generates a handoff prompt for a clean session | | as-save-and-push | Reviews conversation, saves learnings to memory, formats code, groups commits logically, pushes. HARD-GATE on main/master |

Techniques

Each skill uses a combination of these techniques to prevent agent shortcuts:

| Technique | What it does | Example | |-----------|-------------|---------| | Iron Law | One non-negotiable rule at the top | NO FIX WITHOUT ROOT CAUSE | | HARD-GATE | Mandatory stop before a dangerous action | "If modifying code without a test: STOP" | | Red Flags | Thoughts that mean you're skipping steps | "I already know what the bug is" | | Rationalization Table | Maps tempting shortcuts to why they fail | "The fix is obvious" → "Obvious to whom? Prove it" | | Evidence Requirement | Every claim must cite line numbers or tool output | "Cite file:line, not 'I checked'" | | Escalation Limit | Max attempts before asking the human | "5 hypotheses failed → escalate" | | Test List | Enumerate test surface before writing any test | Regression + partitions + boundaries + errors | | Mental Mutation | For each condition in the fix: "would a test catch the inverse?" | "If I changed >= to >, would a test catch it?" | | Autonomous Mode | Rules for subagents that can't interact with user | "Auto-split >300 lines, always continue on bugs" |

Development & Quality Assurance

To ensure cross-agent compatibility, Atomic Skills includes a specialized test suite that acts as a linter for prompt templates.

npm test

The test suite verifies:

1. Tool Name Abstraction: Ensures no hardcoded tool names (like Bash or Read tool) exist in the source .md files.
2. Conditional Rendering: Validates that agent-specific instructions are correctly included/excluded for Claude and Gemini.
3. Multi-Format Export: Verifies Markdown and TOML generation for all supported profiles.

When creating new skills, always use the variables defined in AGENTS.md.

Modules

Memory

Persistent context across sessions. The agent saves learnings, decisions, and feedback that survive between conversations.

• Configurable path (default: .ai/memory/)
• Adds the as-init-memory skill
• Supports Claude Code's autoMemoryDirectory for direct integration (no redirect needed)
• Available in both project and user scope installations

Install, Update, Uninstall

npx @henryavila/atomic-skills install       # First install or update
npx @henryavila/atomic-skills uninstall     # Remove everything

Languages

• Português (BR)
• English

License

MIT