Skill Builder

Build lean SKILL.md-based skills without building a skill factory.

skill-builder helps agent builders decide when a skill is actually justified, draft a minimal skill skeleton, and audit existing skills for bloat and drift.

It is built for people who keep hitting the same questions:

• Should this be a skill, a tool, a subagent, or just a plain edit?
• What belongs in SKILL.md versus references/ versus scripts/?
• How do I build a useful skill without spawning README/changelog/template junk?
• How do I tighten an existing skill without rewriting everything from scratch?

Why this exists

Most skill-authoring advice falls into one of two bad buckets:

1. Too vague — “just write a good SKILL.md”
2. Too ceremonial — giant templates, governance layers, dashboards, policy engines, and file sprawl

skill-builder takes the middle path:

• builder-first enough to produce a real starting skeleton
• reviewer-guarded enough to stop you from overbuilding

It is explicitly designed to help you ship a small, usable skill fast.

What it does

skill-builder operates in three modes:

1. Decide

Choose between:

• plain edit
• existing tool or existing skill
• subagent
• lean new skill

2. Build

When a skill is justified, it helps build the smallest useful package:

• minimal folder tree
• draft frontmatter
• SKILL.md outline or draft
• 1–3 justified references/ files only if needed
• optional scripts/ stub only if deterministic execution is clearly warranted

3. Audit

Tighten an existing skill by checking for:

• bloated SKILL.md
• weak trigger descriptions
• duplicated guidance
• stale instructions
• unnecessary files
• scope drift

What makes it different

This skill is intentionally not:

• a meta-agent
• a governance engine
• an autonomous skill factory
• a repo/dashboard/lifecycle system
• a boilerplate generator that creates ten files before proving one is needed

It is optimized for one thing:

helping you build smaller, sharper skills that are actually worth keeping

Typical use cases

Use skill-builder when you want to:

• create a new OpenClaw / AgentSkills-compatible skill
• decide whether a repeated workflow should become a skill at all
• restructure a skill that has grown into a wall of text
• split a large skill into SKILL.md plus focused references
• tighten a published skill before pushing to GitHub or ClawHub

Example prompts

• “Should this become a skill or just stay a plain edit?”
• “Build me a lean skill skeleton for OpenClaw config troubleshooting.”
• “Audit this skill for bloat and spec drift.”
• “Tighten this SKILL.md and move detail into references.”
• “Decide what belongs in SKILL.md versus references/ versus scripts/.”

Example build output

A typical minimal result looks like this:

Recommendation: Use a skill

my-skill/
├── SKILL.md
└── references/
    └── checklist.md

name: my-skill
description: Helps with X. Use when the user asks for Y or needs Z.

Then SKILL.md only needs:

• scope / non-scope
• default workflow
• pointer to the optional reference file

That’s the point: usable first, elaborate later only if earned.

Repository contents

• skill-builder/SKILL.md — main skill entrypoint
• skill-builder/references/decision-tree.md — choose skill vs tool vs subagent vs plain edit
• skill-builder/references/frontmatter-patterns.md — naming and description patterns
• skill-builder/references/supporting-files-guide.md — what belongs in each folder
• skill-builder/references/audit-checklist.md — fast bloat/spec-drift review

Who this is for

This is for people building skills in:

• OpenClaw
• AgentSkills-compatible environments
• Claude Code / Codex-style SKILL.md workflows
• local agent workspaces that need discipline, not more abstraction theater

If you care about:

• progressive disclosure
• smaller context footprint
• reusable workflows
• anti-bloat design
• practical skill authoring

this skill is for you.

Installation

ClawHub

clawhub install lean-skill-builder

Manual

Copy the skill-builder/ folder into your skills directory.

Positioning in one sentence

skill-builder helps you decide when to build a skill, build the smallest viable one, and keep it from turning into prompt sludge.

License

MIT