| name | bm-skill-builder |
| description | Build a new agent skill (or restructure an existing one) following Builder Methods conventions. Use when the user wants to create a skill, turn a repeatable process into a skill, or restructure a skill that has outgrown its shape — "create a skill", "make this a skill", "turn my proposal process into a skill", "restructure this skill". |
bm-skill-builder
Turns a repeatable process into a well-built agent skill: instructions in plain markdown and folders, portable across any agent harness. Works for brand-new skills and for restructuring existing ones that have outgrown a single SKILL.md.
The process
- Understand the process —
steps/understand.md: ask the user to describe the process the skill is for, then infer what you can — job, inputs, artifact, raw material.
- Interview the user —
steps/interview.md: design the skill with the user — description, name, inputs (if any), its own interview (if any), realistic examples (if applicable), and the step plan. Recommend-then-confirm format throughout.
- Design the shape —
steps/design.md: simple or structured, per the conventions.
- Build it —
steps/build.md: write the files against the conventions checklist.
- Save it —
steps/save.md: ask where the skill should live — this repo's .agents/skills/, global ~/.agents/skills/, or a .claude alternate — and place it there.
- Verify & hand off —
steps/verify.md: run it from its saved location, then teach the user the improvement loop.
Hard rules
- At the start of every invocation of a multi-step skill, present a simple numbered overview of the process with one very short sentence per item, then immediately begin the first step without asking permission or pausing.
- A skill's instructions are plain markdown. SKILL.md and every instruction file (steps, phases, shared, reference docs) are .md — no formats a harness has to interpret. Bundled
scripts/ and reference material can be whatever the job needs (a Python script, a JSON schema, a template file). Nothing harness-specific in the skill's core — it must work anywhere.
- Start as simple as the process allows. Structure (steps/, phases/, shared/) must be earned by real complexity, never added on spec.
- Order lives in the orchestrator, never in the steps. SKILL.md (or a phase's own file) carries the numbering and points at each step file; step files are topic-named, order-agnostic, and say nothing about their position in the sequence.
- The full conventions live in
reference/conventions.md — read them before designing, and check the finished skill against them before handing off.