A portable AI skill that turns source material (code repositories, notes, specs, or business briefs) into decision-ready documents (PRD, BRD, FRD, SOW, Technical Design, Audit/Roadmap/Gap Analysis) through a multi-role council debate.
This package is framework-free. It contains only Markdown files — no code, no dependencies, no runtime. Any AI IDE or agent shell that can read Markdown can consume it.
skill/
├── SKILL.md ← root entrypoint (read this first)
├── README.md ← this file
├── COMPATIBILITY.md ← platform compatibility matrix
├── roles/ ← 10 council role prompts (one per role)
├── workflows/ ← 10 stage definitions (input/output/rules)
├── templates/ ← 6 canonical document structures
├── providers/ ← 23-provider registry + routing rules
├── validators/ ← 3 quality-gate rule sets
├── examples/ ← sample input + sample PRD output
└── adapters/ ← 5 IDE-specific setup guides
Given a repository URL, pasted notes, or a product brief, the skill:
- Understands the source — extracts context map, tech stack, architecture clues, explicit and implicit requirements, likely missing information.
- Researches (optional) — competitors, feature recommendations, cost estimates, sprint plan, architecture simulation.
- Debates — eight expert roles (Product, Business, Architecture, Engineering, QA, Security, UX, Delivery) each propose independently, then respond to each other, then a moderator synthesizes consensus.
- Discovers requirements with an inference engine (e.g. "auth → password policy, forgot/reset, MFA, session timeout, audit logging").
- Surfaces gaps and risks across 10 gap categories and 6 risk categories.
- Picks the right artifact (PRD vs BRD vs FRD vs SOW vs Tech Design vs Audit) with explicit justification.
- Generates a complete Markdown document with tables, Mermaid diagrams, and Confirmed/Inferred/Assumption labels.
- Validates against 10 quality gates; revises once if any fail.
- Emits metadata — traceability matrix, assumptions, confidence ratings, open questions, next actions.
Single-prompt generation produces generic output. The council forces adversarial reasoning: the security reviewer challenges the UX reviewer, the engineering lead challenges the product manager, the QA lead challenges everyone. The moderator resolves or honestly reports unresolved disagreements. The result is a document that has already survived the cross-functional review it would normally get in week three.
- Read
SKILL.md— it is the contract. - Pick your IDE's adapter from
adapters/. - Run the sample input in
examples/sample-input.mdand diff againstexamples/sample-output-prd.md.
| Type | Best for |
|---|---|
| PRD | Product behavior, user needs, features, measurable outcomes |
| BRD | Business need, stakeholder goals, organizational impact |
| FRD | Functional behavior, validations, business rules |
| SOW | Scope, deliverables, timeline, commercial execution |
| Tech Design | Architecture, components, APIs, deployment |
| Audit | Quality / readiness review of an existing system |
| Roadmap | Sequencing, milestones, phased delivery |
| Gap Analysis | What's missing in an existing product or repo |
| AUTO | Let the strategy stage decide |
🎯 Product Manager · 💼 Business Analyst · 🏗️ Solutions Architect · ⚙️ Engineering Lead · 🧪 QA Lead · 🔐 Security Reviewer · 🎨 UX Reviewer · 📦 Delivery Manager · ⚖️ Moderator · 📜 Compliance Reviewer
See roles/ for full prompts.
The skill is model-agnostic. The host IDE supplies the LLM. Twenty-three
providers are pre-profiled in providers/registry.md
covering free tiers, trial credits, and OpenAI-compatible endpoints. Routing
rules in providers/routing-rules.md match
each pipeline stage to the right provider class (low-latency / large-context /
quality-critical).
Validation runs ten gates: completeness, consistency, clarity, testability,
feasibility, ownership, risk coverage, scope boundaries, traceability, and
document correctness. Rules in validators/. A failing gate
triggers at most one revision pass — the pipeline never loops forever.
See COMPATIBILITY.md for the full matrix across
Claude Code, Cursor, Codex, Kiro, GitHub Copilot, Gemini CLI, OpenCode,
local agent shells, and generic project-instruction systems.
- Skill version: 1.0.0
- Source of truth:
SKILL.md - All other files elaborate on the contract defined there.
Inherits the license of the host project. No proprietary dependencies are introduced by this skill package — it is pure Markdown.