name: natprd description: Activate this skill when the user wants to create, update, validate, or review a Product Requirement Document. Triggers on mentions of PRD, product requirements, user stories, acceptance criteria, or initiative documentation, or when the user describes a feature and asks for structured documentation. allowed-tools: Read, Write, Edit, Glob, Grep, Bash, WebFetch, AskUserQuestion

PRD Maker

An interactive, guided PRD generation skill for writing production-grade PRDs. Produces fully structured PRDs with user stories, Gherkin acceptance criteria, section-level rules, and optional sections — all in markdown format.

What This Skill Does

When activated, this skill:

Interviews the user section by section using focused questions
Validates each answer against the section rules before moving on
Writes the PRD incrementally, showing each section as it's completed
Enforces the PRD standard — no placeholders, no rule violations allowed
Outputs a complete prd.md file (default location: docs/prd.md)

The 12 core sections, 4 optional sections, scoring rubric, and per-section rules live in the supporting files referenced at the bottom of this document. Load them on demand when you need the full text — do not duplicate them here.

How to Activate

Simply say:

"I want a PRD for [feature/initiative name]"
"Help me write a PRD"
"Create a PRD for [description]"
"Review my PRD" — to validate an existing PRD file

Claude will detect the intent and load this skill automatically. In Claude Code it can also be invoked directly with /natprd.

User-Choice Presentation

Whenever an interview question offers a choice between 2–4 discrete, mutually-exclusive options, present it using the AskUserQuestion tool so the user can click rather than type. This applies across all three modes.

The tool caps at 4 listed options and automatically adds an "Other" escape, so you can cover up to 5 outcomes (4 explicit + Other) without listing them all.

Examples that MUST use buttons:

§2 Document Status — list the 4 most common (Draft / In Review / Approved / In Execution); "Other" covers Deprecated
§6 Hypothesis confidence (High / Medium / Low)
§8 MoSCoW priority (Must-have / Should-have / Could-have / Won't-have)
§7 Metric type (Leading / Lagging / Guardrail)
Any yes/no compliance flag in §0.3
"Any more user stories?" (Yes — add another / No — move on)
A diagram offer at §8, §9, or §11 (Yes, add it / No, skip)

Exceptions — keep as plain text:

Open-ended questions (the working name in §0.1, evidence narrative in §3, hypothesis prose in §6, baseline/target values in §7): no discrete option set.
Questions with >5 effective options (e.g., §0.2 initiative type with 8 options): keep as text, or split into two sequential button questions.
Multi-select (e.g., §0.3 compliance signals — multiple may apply): use AskUserQuestion with multiSelect: true, up to 4 grouped categories.

If a question doesn't naturally fit any of these patterns, default to text — never invent options to force a button UI.

Anti-Hallucination Rules

These rules apply at ALL times during Mode 1, Mode 2, and Mode 3. They are not optional.

When the user does not have information

If the user says they don't know a detail — or if no answer is provided — use the most specific [TBD] variant defined in the rules below. If no specific variant applies, write [TBD]. Do NOT invent a plausible-sounding value.

This applies without exception to:

Baselines and targets in §4 (Objectives / KRs) and §7 (Metrics)
Event names and metric mappings in §11 (Tracking)
Design artifact links in §9 (Solution)
Stakeholder names in §2, §10, §16
Evidence, sources, and citations in §3 (Background)
Dates, deadlines, and review schedules in §10 and §15
Risk owners, dependency confirmations, and compliance references
Dashboard or monitoring tool names and links in §10

Specific prohibitions

NEVER write a metric baseline or target that the user did not provide. Write [TBD] instead.
NEVER fabricate a design link, Figma URL, Confluence link, or Jira ticket. Write [No design link — status: Draft] instead.
NEVER invent a person's name as a reviewer, approver, DRI, or owner. Write [TBD — owner: ] or [TBD — approver: ] instead.
NEVER write "approximately", "roughly", or "around X%" to soften an invented number. If the user didn't give the number, write [TBD].
NEVER generate evidence, research citations, or data to support the background if the user has not provided it. Write unvalidated assertions as: [Team belief — unvalidated: ___].
NEVER fabricate an event name, tracking property, or data destination. If the user has not confirmed a tracking plan, mark every event row: [TBD — event name to be confirmed by data team].
NEVER pre-populate the data team sign-off checkbox as complete unless the user explicitly confirms it.
NEVER describe a UI or design in prose if no design link exists. Use: [Design pending — link to be added].

When the user is vague

Push back with a clarifying probe before writing the section. Use the probes defined in prompts/interview-questions.md. Do not write the section until you have enough real information to avoid fabricating core content.

[TBD] handling in validation

[TBD] fields are treated as missing for scoring purposes. Each [TBD] is scored as if the field is absent — applying the same deductions as the rubric in prompts/validation-rules.md. [TBD] fields are reported as warnings in the validation report, not violations, but they do reduce the score accordingly.

Reference Research

The skill enriches PRDs with verifiable facts from sources the user explicitly references. Three source classes, each with a defined fallback:

| Source class | Examples | How to read it | |---|---|---| | Local file path | docs/research-2026-q1.md, prior PRD, exported CSV, post-mortem | Read tool directly | | Public URL | News article, regulatory body page, public research, vendor docs, public blog post | WebFetch tool | | Auth-walled URL | Confluence, Jira, internal wiki, private Notion, Figma | Ask user to paste the relevant excerpt — Claude cannot authenticate |

The skill NEVER goes searching for sources on its own (no WebSearch). If the user did not name a source, it does not exist for PRD purposes — write [TBD — needs source] instead.

Bundled registries (fetch-only lookup)

Two curated, versioned registries let the skill suggest authoritative sources without searching. They never replace user confirmation, and the skill still only WebFetches a URL the registry or the user provided — never one it discovered by searching. allowed-tools is unchanged: Bash runs the lookup scripts, WebFetch reads confirmed URLs. No WebSearch.

references/regulations.json — signal + geography → applicable regulations with official source URLs. Queried via python3 scripts/lookup_regulation.py --signals <list> --geo <code> during §0.3b. Carries a last_reviewed date; it is a starting point, not legal advice. The skill confirms candidates with the user, then optionally WebFetches each official_url to quote the exact obligation (verify-before-write + citation).
references/api-docs.md — official documentation URLs for common vendors. When a third-party API is named (§0.4 Q10) but no link is given, the skill suggests the official URL, the user confirms, and the skill WebFetches it to extract rate limits, quotas, and SLAs into §8/§9/§14 — with citations, and [TBD — confirm with vendor docs] when a value isn't stated.

Benchmarks follow the same fetch-only rule: user-pasted or WebFetched from a named URL, recorded in §3 Benchmarks with a mandatory comparability caveat and a retrieved date. Benchmarks are never bundled or invented.

Fetch flow

User mentions a path or URL during §0.6, §3, §4, §7, or any interview turn.
Identify the source class from the table above.
For local paths: Read the file directly.
For URLs: try WebFetch first. If it fails (auth wall, 404, paywall, JS-only rendering, timeout), tell the user "I couldn't fetch that — could you paste the relevant section?" and proceed with the pasted excerpt.
Whether fetched or pasted, the rest of the rules below apply identically.

Verify-before-write rule

For every fact that will land in the PRD with a [source: …] annotation, show the user the exact text first:

"I'll write this in §3 Background: '18.3% of Premium subscribers lapsed in Q1 2026 without an explicit cancel action' [source: docs/research-2026-q1.md, retrieved: 2026-05-15]. OK to write this, or do you want to adjust the quote?"

This is the user's check that the fetch / read actually returned what the model claims. One confirmation per quoted fact — for facts only, not for narrative paragraphs.

What to extract

Verifiable facts only:

Numbers (metrics, baselines, dates, sample sizes) — quote exactly.
Direct quotes (1–2 sentences) attributable to a named source.
Named people (researcher, owner, approver) — cite role and source.
Regulations or standards named in the source — cite the exact reference.
Benchmarks (industry / competitor figures) — quote the value with its date and a comparability caveat.
API / vendor constraints (rate limits, quotas, SLAs, data residency) — quote from the official docs.

How to attribute

Every research-derived fact carries an inline annotation. Format:

[source: docs/research-2026-q1.md, retrieved: 2026-05-15]
[source: https://www.bis.org/publ/bcbs239.pdf §12, retrieved: 2026-05-15]
[source: user-pasted excerpt from Confluence PRD-AUTH-V1 §3.2, retrieved: 2026-05-15]

Use today's date from currentDate. Do not invent a date. Do not invent a section anchor — only cite §N if the source actually has that heading.

What NOT to do

NEVER paraphrase a source — quote the number or sentence directly.
NEVER synthesize across sources without showing the user the synthesis and getting explicit confirmation. "Both reports suggest X" is a synthesis; cite each report separately.
NEVER extrapolate a number ("12% in 2024, so by 2026 it's ~25%"). Write the measured value with its date; mark the rest [TBD].
NEVER fabricate the contents of a WebFetch result. If the fetch failed or returned empty, say so and ask the user to paste — do not invent text the page "would have" contained.
NEVER follow links inside a fetched page to fetch more pages. One fetch per user-provided URL.
NEVER use WebSearch to look up facts the user did not source. The skill has no WebSearch capability for this reason.

When research returns nothing relevant

If the source contains nothing useful for the current section, that is the result. Keep the [TBD]. Do not pad the section to make the source feel useful.

Interaction with anti-hallucination rules

If a source contradicts the user's interview answer, flag the contradiction — do not silently choose.
If a source names a regulation not on the confirmed §0.3b list, ask the user before adding it to §3 Regulatory Context.
Research enriches what the user provided; it does not override user-confirmed answers.

Diagrams (Mermaid)

The skill can add Mermaid diagrams inline in the PRD — a flowchart of a user flow, or a sequence diagram of an event's path — placed next to the prose or table it depicts. Two types only: flowchart and sequenceDiagram. No new tool is needed (allowed-tools is unchanged); Mermaid is plain text written into the markdown.

Two rules govern every diagram, and both are non-negotiable:

Derive from confirmed content only. A diagram is a re-rendering of what the user already confirmed in that section — never a source of new facts. Every node, edge, actor, and message must trace to something the user said. No invented steps, branches, systems, or actors. This is the Anti-Hallucination contract applied to diagrams. If a step is unknown, label it [TBD]; do not fabricate it.
Propose, then confirm. Diagrams are never auto-written. Offer one at the natural point (§8, §9, §11), draft it from confirmed content, show the raw ```mermaid block plus a one-line description of what it depicts, and write it only after the user approves — the same verify-before-write mechanic used for sourced facts.

Full type list, Mermaid conventions, and worked examples are in prompts/diagram-rules.md. Load that file whenever a diagram is offered or requested. Diagrams are an advisory enhancement: they do not change the 100-point validation score.

Version and Date Auto-Update Rules

These rules apply automatically whenever Claude writes or saves the PRD file. No manual entry is needed.

Last Updated

Always set Last Updated to today's date in YYYY-MM-DD format.
Apply on every write: new PRD creation, section edit, status change, or validation fix.
Use the current date from your context (currentDate). Do not invent a date.

Version

Use the vX.Y format. Read the existing Version field from the PRD file before writing. Parse vX.Y into X and Y as integers and apply the matching rule:

| Trigger | Rule | Example | |---|---|---| | New PRD created (Mode 1) | Set to v0.1 | — | | Any section added or edited (no status change) | Increment Y by 1 | v0.2 → v0.3 | | Status change: Draft → In Review | Increment Y by 1 | v0.3 → v0.4 | | Status change: In Review → Approved | Increment X by 1, reset Y to 0 | v0.4 → v1.0 | | Status change: Approved → In Execution | Increment X by 1, reset Y to 0 | v1.0 → v2.0 | | Status change: any → Deprecated | Increment Y by 1 | v2.1 → v2.2 |

Precedence: When a status change and a content edit occur in the same save, apply the status-change rule only. Do not double-increment.

Fallback: If the existing Version field is absent or unreadable, treat it as v0.0 before applying the rule (first increment → v0.1).

Output Path

Default output: docs/prd.md.

If the user specified an explicit file path in their request (e.g., "save it to prds/2026-Q2-auth.md"), honor that path instead. The same rule applies to Mode 2 (read path) and Mode 3 (update path): if the user names a file, use it; otherwise default to docs/prd.md.

When the user requests a one-page stakeholder summary, write it to docs/prd-summary.md (or alongside the PRD if it lives at a custom path: same directory, -summary suffix).

Workflow

Mode 1: Generate (default)

Used when user wants to create a new PRD from scratch.

Step 1 — Intake
  Run the full §0 Intake from prompts/interview-questions.md.
  Do NOT begin the section-by-section interview until all §0 questions are complete.
  Use §0 answers to determine which optional sections to include BEFORE proceeding to Step 2.
  Optional section decisions are final after intake — do not re-ask screening questions later.

  MANDATORY: If any compliance signal is confirmed in §0.3 (payments, eKYC, PII, regulated data,
  credit, etc.), run §0.3b Regulation Identification before proceeding. Identify the specific
  applicable regulations and get user confirmation. The confirmed regulation list carries forward
  into §3 (Regulatory Context), §5 (Scope), and §13 (Risks & Mitigations). This step cannot
  be skipped — at minimum, produce a TBD list with a compliance team owner.

Step 2 — Guided Interview
  Work through each section in order.
  Ask focused questions per section (see prompts/interview-questions.md).
  Validate each answer against section rules before proceeding.
  Show the completed section output before moving to the next.
  At §9 and §11, after the content is confirmed, offer a Mermaid diagram per
  prompts/diagram-rules.md (propose → show source → write only on approval).

Step 3 — Requirements Deep Dive
  For §8, collect all user stories one at a time.
  For each story: role → action → benefit → scenarios.
  Ask "Any more stories?" after each one until the user signals done.
  After a story's scenarios are confirmed, you may offer a flowchart of that
  story's flow per prompts/diagram-rules.md — derived from the Gherkin only.

Step 4 — Optional Sections
  Optional sections were determined at intake. Present the confirmed list to the user:
  "Based on what you told me, I'll include: [list]. Does that sound right?"
  Generate each confirmed section in order: §13, §14, §15, §16.
  Do not re-run optional section screening — the decision was made at intake.
  If the user wants to add or remove a section, adjust accordingly.

Step 5 — Output
  Apply Version and Date Auto-Update Rules: set Version to v0.1, set Last Updated to today's date.
  Assemble and write the complete PRD to the output path (see "Output Path" above).
  Run validation: invoke `python3 scripts/validate.py <path>` for the deterministic baseline,
  then layer the semantic checks from prompts/validation-rules.md on top.
  If python3 or the script is unavailable, skip it and score from prompts/validation-rules.md alone.
  Report validation score and any warnings.
  Offer a summary of what was generated.

Mode 2: Review

Used when the user says "review my PRD" or provides an existing PRD file.

Step 1 — Read the existing PRD file (path from user, or default docs/prd.md).
Step 2 — Run validation: invoke `python3 scripts/validate.py <path>` first,
  then add semantic checks from prompts/validation-rules.md.
  If python3 or the script is unavailable, score from prompts/validation-rules.md alone.
Step 3 — Report: score, issues found, sections that need work.
Step 4 — Offer to fix specific sections interactively.
  After each fix is written to the file: apply Version and Date Auto-Update Rules
  (content-edit rule, or status-change rule if the Status field was modified).

Mode 3: Update

Used when the user wants to revise a specific section of an existing PRD.

Step 1 — Identify the target section and the PRD file path.
Step 2 — Show the current content of that section.
Step 3 — Ask focused questions to gather new/updated content.
  "Add a flow/sequence diagram to §N" is a supported update — draft it from the section's
  confirmed content per prompts/diagram-rules.md, show the source, write only on approval.
Step 4 — Rewrite the section and update the file. Apply Version and Date Auto-Update Rules
  before saving (content-edit rule, or status-change rule if the Status field was modified).
Step 5 — Re-run validation on the updated section.

File References

| File | Purpose | When to load | |---|---|---| | SKILL.md | This file — always loaded | Always | | prompts/interview-questions.md | Questions to ask per section during generation | Mode 1 (each section), Mode 3 | | prompts/section-rules.md | Full rules for every section — used for validation | Modes 1, 2, 3 when validating | | prompts/validation-rules.md | Scoring rubric and validation report format | Modes 1, 2 (after writing/reading PRD) | | prompts/diagram-rules.md | Mermaid diagram spec (types, rules, examples) | When a diagram is offered (§8/§9/§11) or requested | | templates/prd-template.md | The blank PRD template | Mode 1 when assembling final output | | templates/prd-summary-template.md | One-page stakeholder summary template | When user requests a summary | | scripts/validate.py | Deterministic baseline validator (run via Bash) | Modes 1, 2 — before producing report | | scripts/lookup_regulation.py | Registry-backed regulation lookup (run via Bash) | Mode 1 §0.3b when compliance signals exist | | references/regulations.json | Versioned regulation registry (signal+geo → regs + official URLs) | Loaded by the lookup script; not read into context directly | | references/api-docs.md | Official vendor documentation URLs | Mode 1 when a third-party API is named (§0.4 Q10) |

See README.md for installation, the full feature list, and the FAQ.

License

BSD 3-Clause — free to use and redistribute, with attribution. Do not use the author's name to endorse derived works.

Rédacteur de PRD

Recommandé pour