Write Document (BMAD)

Create high-signal technical documentation that serves as an expert collaborator for both humans and AI. This skill enforces the BMAD principles to prevent context rot and ensure architectural durability.

Distinct from edit-document: Use this skill to create a document that does not yet exist. Use edit-document when a document already exists and needs restructuring, clarity, or prose improvements.

HARD GATE — Every document must have a clear "Reason for Existence." If a document doesn't provide actionable leverage for a caller or test, do not create it.

The BMAD Principles

Principle	Execution
Bold	Make strong assertions. Define clear boundaries and "Never" rules. No "it might" or "usually."
Minimal	High-density, low-filler. Circuit Breaker: If the file exceeds 300 lines or the session exceeds 20 turns, you MUST run terse-mode and compact state before saving.
Actionable	Link every doc to a verifiable outcome. Architectural Docs: Verify via Gherkin features (specs/verifications/features/) or grep-based structure checks (grep -c "pattern" file) that prove the design's constraints are present.
Durable	Design for the long-term. Scalability: Use "Nested Indexing"—root files link to module-level GEMINI.md indexes; do not list individual sub-files in the root.

Process

1. Identify the Artifact Type & Scope

Choose the correct BMAD-BigPowers artifact:

• Decision Record (ADR): For "Why" decisions (saved to specs/adr/).
• Context Map: For system-wide architectural mapping (specs/tech-architecture/tech-stack.md).
• Technical Guide: For "How-to" with verification (saved to <module>/REFERENCE.md).
• Behavioral Feature: Gherkin-style compliance specs (saved to specs/verifications/features/).
• Project README: Project-facing documentation (saved to README.md at project root).

Cross-Cutting Concerns: If a doc affects multiple modules, place the authoritative source in the lowest common ancestor directory and use "Delegates" (one-line pointers) in sub-directories to maintain the Single Source of Truth without violating the Stepdown Rule.

2. Draft with Semantic Velocity

STREAM CONTINUITY — When writing file content, output in continuous chunks of ~200 lines. Do not pause. Continue immediately until complete. If you need time, emit a placeholder comment rather than going silent.

Write the document focusing on "Expert Collaboration":

• Instructions over Descriptions: Tell the reader (human or AI) exactly how to interact with the system.
• Provenance Links: Link to ADRs, Issues, or Commits to preserve intent.
• The Stepdown Rule: Information should descend exactly one level of abstraction. If a root doc needs to explain a leaf-level detail, it must point to a sub-index first.

Quick README (Project READMEs only)

1. Ask: "Project name? One-sentence description?"
2. Generate README.md at project root using the template in REFERENCE.md — no TOC, no second interview round.
3. Fill gaps from CLAUDE.md / AGENTS.md commands if available (prefer AGENTS.md); use TODO markers otherwise.
4. Output and suggest edit-document for polish.

→ verify: test -f README.md && [ "$(grep -c '^## ' README.md)" -ge 7 ]

3. Apply the 94% Quality Gate

Before finalizing, audit the document against these red flags:

• Filler Language: Are there pleasantries or "I hope this helps"? (Delete them).
• Ambiguity: Are there "usually," "often," or "it depends" without specific conditions?
• Dead Ends: Does the document end without a "Next Step" or "Verification" command?
• Shallow Content: Does it restate the code without explaining the intent or contracts?

4. Sync and Organize

• Big Powers Hierarchy: Place the document in the correct tier (Global -> Project -> Sub-directory). Project READMEs are an exception — they go to project root (README.md), not specs/.
• Nested Indexing: If adding a module-level doc, ensure the module's GEMINI.md is updated. If the module's index is new, add it to the root GEMINI.md.
• Sync: Run scripts/sync-skills.sh if the document is a SKILL.md or affects generated artifacts.

Rules

• Minimalism is a requirement: If a document can be a 5-line table, do not make it a 5-line essay.
• Verifiable outcomes: Every technical document must include at least one verify: command. For architecture, this can be a grep or run_shell_command that validates the existence of required files or patterns.
• No speculative docs: Do not write documentation for features that do not exist yet unless explicitly doing elaborate-spec.

Suggest next skill: audit-code or sync-skills.sh.