Bmad Builder Skill for Claude

BMAD Builder

Function: Scaffold and validate custom planning/orchestration skills for the BMAD Planning & Orchestrator plugin. Produces compliant SKILL.md files, shell scripts, and templates — pre-wired to this plugin's path conventions — and runs scope-violation checks to keep new skills inside the PLAN/ORCHESTRATE boundary.

Scope (PLAN, never build)

This skill produces planning artifacts and skill skeleton files. It does NOT write application code, run tests, lint, check coverage, or execute builds. If a skill being designed is tempted to "implement", "fix the code", "run the suite", or "review the diff" — that is a scope violation. Plan and hand off instead.

Three intents

Always clarify which intent applies before starting.

Create — new planning skill from scratch

• Gather requirements (use TodoWrite to track):
• Skill name (lowercase-hyphen, prefixed bmad-; e.g. bmad-example)
• What planning/orchestration problem it solves
• Trigger phrases users will say
• Allowed tools (subset of: Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch, TodoWrite)
• Which BMAD track(s) it applies to (Quick Flow / BMad Method / Enterprise)
• Upstream BMAD counterpart (e.g. bmad-risk-assessment)

• Run the scaffold script to create the directory:

   bash ${CLAUDE_PLUGIN_ROOT}/skills/bmad-builder/scripts/scaffold-skill.sh <skill-name>

This creates skills/<skill-name>/ with scripts/ and templates/ subdirectories and a starter SKILL.md.

• Fill the SKILL.md using the skill template:
• Reference: ${CLAUDE_PLUGIN_ROOT}/skills/bmad-builder/templates/skill.template.md
• Keep body under 5K tokens; push long reference detail to a sibling REFERENCE.md in the new skill's own folder
• Use ${CLAUDE_PLUGIN_ROOT} for all paths to bundled scripts/templates
• Artifacts go to bmad-output/ (honor user-configured folder)
• End with the mandatory attribution footer (see template)

• Validate the new skill:

   bash ${CLAUDE_PLUGIN_ROOT}/skills/bmad-builder/scripts/validate-skill.sh \
     ${CLAUDE_PLUGIN_ROOT}/skills/<skill-name>/SKILL.md

Fix all errors; review all warnings before declaring done.

Validate — check an existing SKILL.md

Run the validator directly against any SKILL.md path:

bash ${CLAUDE_PLUGIN_ROOT}/skills/bmad-builder/scripts/validate-skill.sh <path-to-SKILL.md>

The validator checks:

• name field present and lowercase-hyphen
• description field present with trigger phrases
• allowed-tools field present (warns if missing)
• No dev/lint/build/coverage scope violations in allowed-tools or body
• File size within the ~5K token target (~20KB)
• Attribution footer present

Scaffold — directory structure only

Use when you want an empty directory skeleton without generating content:

bash ${CLAUDE_PLUGIN_ROOT}/skills/bmad-builder/scripts/scaffold-skill.sh <skill-name>

Then fill SKILL.md manually using the template as a guide.

Scope law for new skills

Every skill created by this builder MUST comply with the Scope Law:

• The plugin PLANS and ORCHESTRATES. It NEVER writes application code, runs tests, lints, checks coverage, builds, or reviews implemented code.
• The last artifact any skill may produce is a ready-for-dev story file or a handoff manifest. Implementation is handed to EXTERNAL dev tools/plugins.
• Acceptance Criteria, Testing STRATEGY, and Dev Notes are planning — allowed. Executing tests or writing implementation code is out of scope.

The validate-skill.sh script enforces this automatically by flagging forbidden terms.

BMAD fidelity reminders

When designing skills, track the BMAD Method conventions:

| Convention | Rule |

|------------|------|

| Scale Tracks | Quick Flow / BMad Method / Enterprise — user confirms interactively |

| Story size | ~2-8h, one dev-day max; split if larger |

| Delivery tracking | Stories remaining / completion rate — NO Fibonacci points, velocity, or burndown |

| Story name | {epic}.{story}.{slug}.story.md |

| Story status | backlog → ready-for-dev → in-progress → review → done |

| Decision log | Decisions thread through decision-log.md |

| Project constitution | project-context.md loaded across skills |

| Three intents | Create / Update / Validate where applicable |

Available scripts

| Script | Purpose |

|--------|---------|

| scripts/scaffold-skill.sh <name> | Creates skills/<name>/ with subdirs and starter SKILL.md |

| scripts/validate-skill.sh <SKILL.md> | Validates frontmatter, scope, size, and attribution footer |

Invoke with the ${CLAUDE_PLUGIN_ROOT} prefix shown above. The orchestrator marks scripts executable; you may also run them via bash.

Templates

| Template | Use for |

|----------|---------|

| templates/skill.template.md | Skeleton for a new planning/orchestration SKILL.md |

| templates/document.template.md | Generic planning document (PRD section, brief, spec section, etc.) |

Subagent strategy

For creating a full skill package in parallel:

| Agent | Task | Output |

|-------|------|--------|

| Agent 1 | Draft SKILL.md body from gathered requirements | skills/<name>/SKILL.md |

| Agent 2 | Write domain scripts (validators, checklists) | skills/<name>/scripts/ |

| Agent 3 | Write domain templates | skills/<name>/templates/ |

| Agent 4 | Write REFERENCE.md if body exceeds 5K tokens | skills/<name>/REFERENCE.md |

Coordination: gather requirements first (sequential), write spec to bmad-output/skill-spec.md, then fan out. Main context validates all outputs with validate-skill.sh and assembles the final package.

Notes for LLMs

• Use TodoWrite to track which components have been created.
• Never create a skill that contains test runners, linters, coverage tools, build commands, or diff/code-review steps — validate-skill.sh will flag these.
• Always end SKILL.md with the attribution footer block (the template includes it verbatim).
• Keep SKILL.md under 5K tokens; use REFERENCE.md for lengthy patterns.
• Use ${CLAUDE_PLUGIN_ROOT} for all internal paths — never hardcode ~/.claude or absolute machine paths.
• Output artifacts (non-skill files) go under bmad-output/ by default.

> ---

> Part of the BMAD Planning & Orchestrator plugin — a Claude Code harness for the BMAD Method by the BMAD Code Organization (https://github.com/bmad-code-org/BMAD-METHOD). Implements the spirit of bmad-bmb-builder. All methodology credit belongs to the BMAD Code Organization.