PRD Generator

Plugin for Claude Code and Codex. Turns a freeform product idea into a validated, cross-doc-consistent PRD package, ready to hand to a coding agent.

In 30 seconds

> /prd-generator
[skill] Language? > English
[skill] Brainstormed already? > Yes
[skill] Briefly describe what you want to build:
> Multi-tenant blog platform for portfolio companies. Each tenant
> gets isolated content, AI writers, publishing webhooks. Stack:
> Next.js 16 + Supabase. MVP needs auth, brief approval, content
> generation, scheduling.

[skill asks 6 high-leverage questions in 2 batches]
[skill confirms summary, you approve]

[skill] Generated 11 files in projects/blog-saas/.
        16 quality gates passed.
        8 items safe to implement.
        2 need decision before coding.
        3 deferred to v1.1.

Now any coding agent can read your PRD without making things up.

What it generates

                    ┌─────────────────┐
                    │   Main PRD      │
                    │ {slug}-prd.md   │
                    └────────┬────────┘
                             │
        ┌────────────────────┼────────────────────┐
        ↓                    ↓                    ↓
  ┌──────────┐         ┌──────────┐        ┌──────────┐
  │AGENTS.md │         │CLAUDE.md │        │  api-*   │
  │  rules   │         │ guardrail│        │ contracts│
  └──────────┘         └──────────┘        └──────────┘
        ↓                    ↓                    ↓
  ┌────────────────────────────────────────────────┐
  │  stack docs + DEVELOPMENT-WORKFLOW + README    │
  │       + CHANGELOG + service-boundaries         │
  └────────────────────────────────────────────────┘

Default output: projects/{project-slug}/.

Why it exists

Most AI-generated PRDs fail in one of two ways:

Failure 1: sounds complete, hides decisions
Failure 2: five docs that contradict each other

This plugin pushes against both:

Defense	How
Hidden uncertainty	Forces explicit TBD { blocks_coding: yes/no }, Assumed { question, default, flip_cost }, Proposed { promote_when } labels
Doc drift	16 quality gates validate that PRD, endpoints, models, and RLS summaries agree
Workflow handwaving	Every workflow step must reference an exact endpoint or UI only
Stale runtime	Version Policy blocks EOL Node, Next.js, or framework defaults
Accidental activation	The skill only runs from /prd-generator, never from natural-language brainstorm

Install

In Claude Code, add the marketplace from feba-capital/prd-generator and install prd-generator. In Codex, the manifest is in .codex-plugin/plugin.json and the skill payload is the same skills/prd-generator/ directory.

See docs/skill-installation.md for the full path.

Use

/prd-generator

The workflow:

1. Picks language (English, Portuguese, Spanish, Chinese, or Other)
2. Confirms idea is brainstormed (or hands off to a brainstorm skill)
3. Offers an optional 7-question MVP scope check
4. Reads your briefing, asks 2 to 4 high-leverage questions in batches
5. Confirms summary before generating
6. Writes the package into projects/{slug}/
7. Runs 16 quality gates and reports pass/fail per gate
8. Returns an Implementation Readiness summary

Works with any stack. Two presets bundled (nextjs-supabase, yii2-mysql); other stacks get ad-hoc generation that you can promote into a permanent preset later.

What's in the box

Path	Purpose
commands/prd-generator.md	The slash command entry
skills/prd-generator/SKILL.md	The workflow source of truth
skills/prd-generator/interview/	Step 0 to Step 6 prompts
skills/prd-generator/templates/	Base templates copied into every package
skills/prd-generator/stack-presets/	Reusable per-stack docs
skills/prd-generator/scripts/validate-generated-docs.sh	The 16-gate validator
examples/wisercontent/	Frozen reference snapshot
tests/	Pass/fail fixtures and three regression scripts

Quality gates (the 16)

Group	Gate
Cross-doc	Workflow steps must point at endpoints that exist
Cross-doc	Reused endpoint anchors must carry distinct role/state qualifiers
Cross-doc	Endpoint access summaries must match RLS policy text
Cross-doc	Role claims must agree across PRD, endpoints, and RLS
Format	TBD / Assumed / Proposed must use classifier syntax
Format	No legacy TBD (reason: ...) in new output
Format	No em dash in prose
Format	No Lorem, example.com, foo bar, or TODO placeholders
Structure	All required sections present per file type
Structure	Implementation Readiness non-empty, with source citations
Structure	Future Versions and Launch Dependencies mutually exclusive
Structure	Validation Plan carries all four required fields
Structure	Scope Decisions items carry a non-empty reason
Hygiene	Runtime and framework versions current stable or pinned with reason
Hygiene	Non-admin UPDATE policies declare explicit transitions
Hygiene	Non-admin UPDATE policies enforce explicit columns

Develop on it

git clone https://github.com/feba-capital/prd-generator
cd prd-generator
bash tests/test-validate-generated-docs.sh
bash tests/test-sprint-1-regressions.sh
bash tests/test-sprint-2-regressions.sh

See CONTRIBUTING.md for the full flow, including how to add a new stack preset or a new language.

Authors

Created by Bill Madeira and Fabio Espindula at FEBACAPITAL.

Bill wrote the original PRD that became the seed of this plugin. Fabio reverse-engineered its structure into a reusable generator and built the validator, quality gates, stack presets, and multi-language workflow.

License

MIT. See LICENSE.