Claude Code skills that augment existing CLI tools with AI intelligence. The tools do the heavy lifting; AI handles context, decision-making, and error recovery.
Use the right tool for the job. Use AI for what the tool cannot do.
Every skill in this repo wraps a real CLI tool. The tool executes the work; the skill adds the layer of intelligence that makes the tool genuinely useful in an agentic context: interpreting output, diagnosing failures, choosing the right flags, and recovering from errors.
This keeps token usage efficient. If acli can create a Jira ticket, it does — the skill is there to translate intent into the right command, not to reinvent the wheel.
| Skill | Wraps | What AI adds |
|---|---|---|
| act-workflows | act |
Parallel job execution, failure diagnosis, and fixes for common act compatibility issues |
| commit | git |
Conventional commit formatting, issue number detection, logical change splitting |
| fix-github-workflow | gh + act |
CI failure diagnosis, fix application, local validation loop, watching the outcome |
| gabo-workflows | gabo + act |
Workflow type selection, generation orchestration, then hands off to act-workflows for testing |
| ralph-loop | claude + Docker sandboxes |
Assesses Ralph conformance, guides setup across all five pillars, enforces the principles |
pnpm dlx skills@latest add RyanGannon/skillsTo update:
pnpm dlx skills@latest updateSkills that encode expert knowledge follow a three-step pipeline:
- Refine the source — Run a raw transcript (video or talk) through a transcript refiner script to clean it up
- Research and synthesise — Pass the refined transcript to Claude with a prompt that expands knowledge via web search, resolves conflicts by favouring the creator's own words, and produces a structured introduction document
- Generate the skill — Pass the document to
/write-a-skill
The ralph-loop skill was built this way, starting from Geoffrey Huntley's talk "The Ralph Wiggum Loop from 1st principles".