micro-skill-pipeline
Stop writing 200-line skills that Claude skims. Split them into 5 small steps with hard gates.
The Problem
Large SKILL.md files (100-300+ lines) get partially read, partially followed, and produce inconsistent output. Claude skips sections, forgets constraints, and rationalizes "good enough" as done. There's no checkpoint, no verification, no learning from past mistakes.
The Solution
This plugin converts any monolithic skill into a 5-stage gated pipeline:
| Stage | What happens | Gate |
|---|
| 1. Scope | Extract constraints from the request | Can I state the deliverable in one sentence? |
| 2. Plan | Design the approach, check failure log | Does every constraint map to a component? |
| 3. Build | Execute with micro-checks per component | Did every component pass its micro-check? |
| 4. Check | Hard gate -- binary YES/NO questions | ALL questions YES. Any NO = fix first. |
| 5. Deliver | Clean up and present output | Output in final location? Summary concise? |
Each stage is a separate file under 40 lines. Claude reads one at a time, completes it, passes its gate, then reads the next. No skipping ahead.
The Check step is the hard gate. It contains specific, binary questions -- not "ensure quality" but "does every endpoint have error handling? YES or NO." Any NO means fix before proceeding.
A failure log compounds learnings over time. Every time a Check gate catches something, the pattern gets recorded. Next run, Claude reads those patterns before starting.
Installation
From local directory
claude install-plugin /path/to/micro-skill-pipeline
From GitHub marketplace
# Add the marketplace (one-time)
claude plugin marketplace add github:stevesolun/micro-skill-pipeline
# Install the plugin
claude plugin install micro-skill-pipeline
Quick Start
1. Convert an existing skill
/skill-convert api-design
Just pass the skill name -- the plugin searches your project skills, user skills, agent skills, and installed plugins to find it automatically. You can also pass a full path if you prefer:
/skill-convert ~/.claude/skills/api-design/SKILL.md
This reads the monolith, splits it into ~10 files (each under 40 lines), extracts YES/NO gate questions, and seeds a failure log.
2. Use the converted pipeline
/skill-use api-design
Claude now runs through Scope -> Plan -> Build -> Check -> Deliver with gates between each step. If the original skill has been updated since conversion, you'll be asked whether to reconvert.
3. Or just use any skill normally
When you invoke a skill that has a converted pipeline, a hook notifies you:
> Micro-skill pipeline: A converted pipeline exists for 'api-design'.
> Use /skill-use api-design to run the gated pipeline, or proceed with the original.
You choose -- the original is never modified or overridden.
Commands
| Command | What it does |
|---|
/skill-convert <name-or-path> | Convert a monolithic skill into a gated pipeline |
/skill-use <name> | Activate and run a converted pipeline |
/skill-new <name> | Scaffold a blank pipeline from scratch |
/skill-check [name] | Run the Check gate against current work |
/skill-list | List all converted skills with staleness status |
/skill-fail <desc> | Log a failure pattern to both project and skill logs |
/skill-bypass <name> | Use the original skill, skipping the pipeline |
/skill-convert <name-or-path>
Convert a monolithic skill into a micro-skill pipeline. Pass a skill name or full path -- the plugin auto-discovers the skill location.
/skill-convert api-design # by name (auto-finds it)
/skill-convert ~/.claude/skills/my-big-skill/SKILL.md # by path
Skill search order: project skills -> user skills -> agent skills -> installed plugins. If multiple matches are found, you pick which one.
What it does:
- Reads the source skill, computes SHA256 hash
- Classifies sections: instructions -> Build, checks/avoid -> Gate questions, reference data -> separate files
- Creates the pipeline in persistent storage (
${CLAUDE_PLUGIN_DATA}/converted/<name>/)
- Extracts YES/NO gate questions into
check-gates.md
- Seeds
failure-log.md with "avoid" items from the original
- Registers in
registry.json for staleness tracking
/skill-use <name>
Activate a converted pipeline. This is how you run a converted skill instead of the original.
/skill-use api-design
What happens:
- Staleness check -- compares the original skill's current hash to the stored hash
- If the original has been updated since conversion, asks: "Reconvert to pick up changes?" (your failure-log is preserved)
- If you agree, reconverts automatically; if you decline, uses the existing pipeline
- Loads the pipeline -- runs through Scope -> Plan -> Build -> Check -> Deliver with gates
