prd-to-skill

Turn any PRD into an AI coding assistant instruction file — in one command.

prd-to-skill converts Product Requirements Documents (PDF and Word) into ready-to-use instruction files for Claude Code, Cursor, OpenAI Codex CLI, GitHub Copilot, Windsurf, and Aider. One command, correct format, right directory — no manual copy-paste.

It uses AI to intelligently extract only the implementation-relevant parts of your PRD (requirements, architecture, business rules, acceptance criteria) and discards project management noise (timelines, stakeholders, budgets). The output is a structured, actionable instruction file that your AI coding tool can use immediately.

Need granular control? The split command breaks one PRD into multiple focused skill files — one per topic — so you can load only the skills relevant to your current task.

Why prd-to-skill?

• Save hours of manual work — stop copy-pasting PRD sections into markdown files by hand
• Works with every major AI coding tool — one CLI covers Claude Code skills, Cursor rules, Codex AGENTS.md, Copilot instructions, Windsurf rules, and Aider conventions
• AI-powered extraction — doesn't just dump raw text; intelligently transforms requirements into actionable instructions with code examples, decision criteria, and verification checklists
• Split into focused skills — generate multiple targeted skill files from one PRD (tech stack, business rules, architecture, etc.) and load only what's relevant
• Bring your own LLM — works with OpenAI, Anthropic, Google Gemini, and Mistral with zero SDK dependencies
• Smart edge case handling — detects truncated documents and incomplete LLM output, warns you, and suggests fixes
• Zero config — just set an API key and run npx prd-to-skill ./prd.pdf

Quick Start

# Set any supported API key
export OPENAI_API_KEY="sk-..."

# Generate a single Claude Code skill (default)
npx prd-to-skill ./my-feature-prd.pdf

# Split into multiple focused skills (tech stack, business rules, architecture, etc.)
npx prd-to-skill split ./my-feature-prd.pdf

# Generate a Cursor rule
npx prd-to-skill ./my-feature-prd.pdf --target cursor

# Generate an AGENTS.md for OpenAI Codex CLI
npx prd-to-skill ./my-feature-prd.pdf --target codex

# Generate for every tool at once
for t in claude cursor codex copilot windsurf aider; do
  npx prd-to-skill ./my-feature-prd.pdf --target $t
done

Supported AI Coding Tools

| Tool | Flag | Output Path | Format | | -------------- | --------------------------- | --------------------------------------------- | --------------------------- | | Claude Code | --target claude (default) | .claude/commands/<name>.md | Markdown + YAML frontmatter | | Cursor | --target cursor | .cursor/rules/<name>.mdc | MDC + YAML frontmatter | | OpenAI Codex | --target codex | ./AGENTS.md | Plain Markdown | | GitHub Copilot | --target copilot | .github/instructions/<name>.instructions.md | Markdown + YAML frontmatter | | Windsurf | --target windsurf | .windsurf/rules/<name>.md | Plain Markdown | | Aider | --target aider | ./CONVENTIONS.md | Plain Markdown |

Output directories are created automatically if they don't exist.

Supported LLM Providers

The tool auto-detects your provider from whichever API key is set (checked in this order):

| Provider | Env Var | Default Model | | --------- | ------------------- | ---------------------------- | | Anthropic | ANTHROPIC_API_KEY | claude-sonnet-4-6-20250217 | | OpenAI | OPENAI_API_KEY | gpt-5.4 | | Google | GOOGLE_API_KEY | gemini-2.5-flash | | Mistral | MISTRAL_API_KEY | mistral-large-latest |

No SDK dependencies — all LLM calls use lightweight raw HTTP adapters, keeping npx installs fast.

Run prd-to-skill help <provider> for setup instructions and links to official docs for any provider.

Usage

prd-to-skill <file> [options]

Options

| Flag | Description | Default | | -------------------------- | ------------------------------------------------------------------------ | --------------------- | | -t, --target <target> | Target tool: claude, cursor, codex, copilot, windsurf, aider | claude | | -p, --provider <name> | LLM provider: openai, anthropic, google, mistral | Auto-detected | | -m, --model <model> | Model name | Provider default | | -o, --output <path> | Output file path (overrides default) | Target-specific | | -n, --name <name> | Skill/rule name | Derived from filename | | -d, --description <text> | Description for frontmatter | Generated by LLM | | --max-tokens <number> | Max output tokens | 4096 | | -v, --verbose | Show extraction and API details | |

Commands

| Command | Description | | ------------------------------ | ------------------------------------------------------- | | prd-to-skill split <file> | Split PRD into multiple focused skill files | | prd-to-skill help | Detailed help with all targets, providers, and examples | | prd-to-skill help <provider> | Setup guide with docs link for a specific provider |

Examples

# Convert a PDF PRD into a Claude Code skill using Anthropic
prd-to-skill ./prd.pdf --provider anthropic

# Convert a Word doc into a Cursor rule
prd-to-skill ./prd.docx --target cursor --name auth-feature

# Use Google Gemini with a custom model
prd-to-skill ./prd.pdf --provider google --model gemini-2.5-pro

# Custom output path
prd-to-skill ./prd.pdf --output .claude/commands/auth.md

# Increase token limit for large PRDs
prd-to-skill ./prd.pdf --max-tokens 8192

Splitting a PRD into Multiple Skills

The split command analyzes your PRD with AI and automatically determines the right set of focused skill files to generate — one per topic area.

prd-to-skill split ./my-feature-prd.pdf

What it generates

Given a PRD, the LLM identifies the relevant categories (e.g. 3-7 depending on the document) and generates a separate skill file for each. Common categories include:

| Category | What it covers | | ------------------------- | ------------------------------------------------------------------ | | tech-stack | Frameworks, libraries, dependencies, version constraints | | business-rules | Domain logic, validation rules, edge cases | | architecture | File structure, component organization, naming conventions | | functional-requirements | Core features, user flows, expected behaviors, acceptance criteria | | ui-ux | Layout, interaction patterns, responsive behavior | | api-contracts | Endpoints, request/response shapes, authentication | | data-models | Schemas, data structures, relationships |

The categories are determined per-PRD — only topics with sufficient content are generated.

Output files

All files are written to the same directory as the single-file command (e.g. .claude/commands/ for Claude Code):

.claude/commands/
  my-feature-tech-stack.md
  my-feature-business-rules.md
  my-feature-architecture.md
  my-feature-functional-requirements.md
  my-feature-index.md          ← index listing all skills with descriptions

Split command options

prd-to-skill split <file> [options]

| Flag | Description | Default | | ----------------------- | -------------------------------------------------------- | --------------------- | | -t, --target <target> | Target tool: claude, cursor, copilot, windsurf | claude | | -p, --provider <name> | LLM provider: openai, anthropic, google, mistral | Auto-detected | | -m, --model <model> | Model name | Provider default | | -n, --name <name> | Base name for generated files | Derived from filename | | --max-tokens <number> | Max output tokens per skill file | 4096 | | -v, --verbose | Show extraction and API details | |

Note: codex and aider targets are not supported for split because they use fixed output filenames (AGENTS.md / CONVENTIONS.md). Use the default command for those targets.

Split examples

# Split a PRD into focused Claude Code skills
prd-to-skill split ./prd.pdf

# Split into Cursor rules
prd-to-skill split ./prd.pdf --target cursor

# Use a specific provider and model
prd-to-skill split ./prd.pdf --provider anthropic --model claude-opus-4-6

# Custom base name for output files
prd-to-skill split ./prd.pdf --name auth-feature

How It Works

Default command

┌──────────────┐     ┌──────────────┐     ┌──────────────┐     ┌──────────────┐
│  PDF / DOCX  │────▶│   Extract    │────▶│   LLM Call   │────▶│  Write File  │
│   (input)    │     │    Text      │     │  (any model) │     │ (target fmt) │
└──────────────┘     └──────────────┘     └──────────────┘     └──────────────┘

1. Extracts text from your PDF or Word document
2. Detects issues — warns if the document appears truncated or incomplete
3. Filters intelligently — keeps requirements, architecture, business rules, and acceptance criteria; strips timelines, stakeholders, and budgets
4. Generates with AI — sends to your chosen LLM with a target-specific prompt that produces the correct file format
5. Validates output — detects if the LLM output was cut off and suggests fixes (increase --max-tokens, use a larger model, or split the PRD)
6. Writes the file to the right directory for your chosen tool

Split command

┌──────────────┐     ┌──────────────┐     ┌──────────────┐     ┌─────────────────────┐     ┌───────────────┐
│  PDF / DOCX  │────▶│   Extract    │────▶│  Plan Call   │────▶│  Parallel LLM Calls │────▶│  Write Files  │
│   (input)    │     │    Text      │     │  (identify   │     │  (one per category) │     │  + Index File │
└──────────────┘     └──────────────┘     │  categories) │     └─────────────────────┘     └───────────────┘
                                          └──────────────┘

1. Extracts text from your PDF or Word document
2. Plans categories — makes one LLM call to analyze the PRD and identify 3-7 relevant skill categories
3. Generates skills in parallel — sends one LLM call per category, each focused on only that topic
4. Writes skill files — one file per category, all in the target tool's directory
5. Writes an index file — a summary listing every generated skill with descriptions and links

Common Use Cases

• Starting a new feature — convert the PRD to a Claude Code skill so Claude has full context when implementing
• Onboarding AI to a project — generate Cursor rules or Copilot instructions from existing specs so the AI follows your project's requirements
• Multi-tool teams — generate the same PRD as instructions for every AI tool your team uses with a single loop
• Keeping AI context up to date — re-run whenever the PRD changes to regenerate the instruction file

Troubleshooting

Output is incomplete

The LLM hit its token limit. Try:

prd-to-skill ./prd.pdf --max-tokens 8192

Or use a model with a larger output window, or split your PRD into smaller documents.

Document appears truncated

The tool detected that your PDF/DOCX ends mid-sentence. Check that the source document is complete. The tool will still generate output from what's available and list what appears to be missing.

No API key found

Set one of the supported environment variables:

export ANTHROPIC_API_KEY="sk-ant-..."  # or
export OPENAI_API_KEY="sk-..."         # or
export GOOGLE_API_KEY="AI..."          # or
export MISTRAL_API_KEY="..."

Run prd-to-skill help <provider> for step-by-step setup.

Contributing

Contributions are welcome! Please open an issue or pull request on GitHub.

License

MIT