Use this skill whenever the user wants to write, draft, create, or scaffold a PRD (Product Requirements Document), spec, product brief, one-pager, or "working backwards" PR/FAQ from an idea, rough notes, discovery inputs, or an existing partial draft that needs authoring. Trigger on phrases like "write a PRD", "draft a spec", "product doc for X", "one-pager", "PR/FAQ", "feature brief", or when the user describes a feature/product idea and asks for it to be turned into a written requirements document. Do NOT trigger for critique or approval-readiness review of an existing PRD (use prd-review), engineering design docs, RFCs, ADRs, prioritization, roadmaps, or technical architecture documents. The output of this skill is always a markdown file (or section thereof) that a real product team could ship from.
Installation
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Use this skill whenever the user wants to write, draft, create, or scaffold a PRD (Product Requirements Document), spec, product brief, one-pager, or "working backwards" PR/FAQ from an idea, rough notes, discovery inputs, or an existing partial draft that needs authoring. Trigger on phrases like "write a PRD", "draft a spec", "product doc for X", "one-pager", "PR/FAQ", "feature brief", or when the user describes a feature/product idea and asks for it to be turned into a written requirements document. Do NOT trigger for critique or approval-readiness review of an existing PRD (use prd-review), engineering design docs, RFCs, ADRs, prioritization, roadmaps, or technical architecture documents. The output of this skill is always a markdown file (or section thereof) that a real product team could ship from.
disable-model-invocation
false
PRD Author
A skill for producing spectacular Product Requirements Documents — the kind that
reduce ambiguity, force tradeoffs into the open, and make a team meaningfully smarter
about what they're building before a line of code is written.
This skill is opinionated. Generic "what is a PRD" advice is everywhere; what's rare
is a skill that consistently produces docs which survive scrutiny in a real product
review. That is the bar.
Contract
Use this as a write-capable product-document skill. It may draft or revise a
markdown PRD-class artifact when the user asks for a document. It must not
impersonate PRD approval review, engineering architecture, roadmap
prioritization, launch comms, or project management.
Hard gate: do not present a PRD as complete when the problem, target user,
decision to be made, or success metric is missing. Ask once with
AskUserQuestion when available; otherwise ask one consolidated question. If the
user wants forward progress anyway, continue with visible [TBD],
[ASSUMPTION], or [NEEDS RESEARCH] markers and return
DONE_WITH_CONCERNS.
Intake must classify the request before writing:
quick: one-pager or section draft for discovery or scoping.
standard: full PRD for a normal feature or product change.
deep: net-new, multi-quarter, regulated, high-risk, or cross-functional bet.
revision: revise an existing PRD when the user asks for direct edits.
Evidence requirement: inspect all provided source docs, analytics, research,
tickets, meeting notes, designs, and prior drafts before making claims. For
standard/deep PRDs, require at least two independent evidence classes for the
problem statement or explicitly label the gap. Never invent numbers, citations,
customer quotes, owners, dates, or legal/compliance positions.
1. Philosophy
A PRD is a tool for thinking, not an artifact for compliance. Its job is to:
Force the author to articulate the problem precisely enough that the solution
becomes negotiable.
Surface the hardest tradeoff explicitly, so the team makes it on purpose rather
than by accident.
Give success a falsifiable definition before anyone writes code.
Align stakeholders — engineering, design, data, GTM, legal — on the same understanding.
If a draft does not do all four, it is not done. It is a description, not a PRD.
The single most important sentence in a PRD is the problem statement.
Everything else is downstream of it. If the problem is fuzzy, the solution will be
fuzzy, the metrics will be fuzzy, and the team will ship something fuzzy.
Experiment / test plans that need product framing (not just statistical design)
Renaming, deprecating, or migrating a feature
Internal tools whose users are colleagues
One-pagers / mini-PRDs for discovery or scoping
Do not use it for:
Engineering design docs, RFCs, ADRs (architecture is downstream)
Pure project plans (Gantt charts, milestones without product framing)
Marketing briefs, launch comms, press releases as the primary artifact
Bug reports or task tickets
If the user request is ambiguous (e.g. "write a doc for our new feature"), default to
the PRD format and ask one clarifying question about the audience and decision the
doc must drive. See §4 (Interview).
3. The Three Sizes of PRD
Match the size of the doc to the size of the bet. Do not over-document a small thing,
and do not under-document a large one.
Size
When to use
Length
Required sections
One-Pager
Discovery, scoping, early alignment
< 1 page (~400 words)
TL;DR, Problem, Proposed Solution, Success Metric, Open Questions
Standard PRD
Most features and products
3–6 pages
All §6 sections, condensed
Deep PRD
Net-new product, multi-quarter bet, regulated/high-risk
8–15 pages
All §6 sections, full appendices
Never exceed 15 pages. If a PRD is longer than 15 pages, it has stopped being a PRD
and become a project plan in disguise. Split it.
4. Interview Before You Write
Before drafting, gather the following. If the user has not provided answers, ask
explicitly in a single consolidated question (not an interrogation):
Audience. Who is this PRD for? (Exec review? Eng team? Cross-functional kickoff?)
Decision. What decision must this doc drive? (Build / don't build? Scope? Sequencing?)
User. Who is the customer or user — specifically, not "everyone"?
Problem. What is the problem in their words, with evidence (data, quotes, tickets)?
Why now. What changed that makes this the right time?
Success. What numerical outcome would make this a clear win?
Constraints. Hard constraints — regulatory, technical, timeline, budget?
Out of scope. What are we explicitly not doing?
If the user provides only a feature idea ("let's add X"), reframe back to the problem
before writing. Never write a PRD that starts from the solution. The skill must
gently insist on the problem-first framing even when the user resists it.
If after one round of questions the user still cannot articulate the problem clearly,
draft a placeholder PRD with [NEEDS RESEARCH] flags in the relevant sections
rather than fabricating crispness.
5. Core Principles for Every Section
These apply everywhere in the document.
5.1 Specificity over completeness
A PRD with three crisply-defined requirements beats one with twenty hand-wavy ones.
If a sentence could be deleted without changing engineering's behavior, delete it.
5.2 Falsifiable success metrics
Every goal must be expressible as a number with a baseline, a target, and a timeframe.
❌ "Improve the checkout experience"
✅ "Reduce checkout abandonment from 34% to under 25% within 90 days of launch, measured on the existing cart_abandoned event"
5.3 Non-goals are not optional
A PRD without a Non-Goals section is broken. Scope is defined as much by what you
refuse to do as by what you commit to. Include at least three non-goals, and prefer
ones that are plausibly tempting (not strawmen).
5.4 Separate problem, solution, and implementation
Problem: what's wrong in the user's world
Solution: the user-visible behavior change that fixes it
Implementation: how engineering will build it (mostly not in the PRD)
When in doubt, the PRD describes what the user can do, not what the system does.
5.5 Tradeoffs explicit, not buried
Every meaningful product decision involves a tradeoff. Surface it. A Tradeoffs
subsection or callout is required for any PRD beyond the one-pager.
5.6 Evidence over assertion
Claims like "users want X" without a source are noise. Cite: support tickets, NPS
verbatims, analytics, sales calls, competitor analysis, user research. If there is
no evidence, mark it [ASSUMPTION] so reviewers can challenge it.
5.7 Reversibility test
For each requirement, ask: "If we shipped the opposite of this, would the doc
still make internal sense?" If yes, the requirement is not load-bearing — cut or
sharpen it.
6. The Anatomy of a Standard PRD
The required structure. Use these section headings verbatim. Order matters.
The whole PRD in a paragraph. Must answer: what are we building, for whom, why,
and how will we know it worked. If a busy exec read only this, would they know
enough to ask the right question? If not, rewrite.
6.2 Context
Why this, why now, what's the backdrop. Strategic fit (one sentence). Recent signals
(data point, quote, market shift). Keep it under half a page — context is not history.
6.3 Problem
The single most-revisited section. Structure:
Who experiences the problem
When they experience it (the trigger / moment)
What they're trying to do (the job-to-be-done)
Why the current state fails them
Evidence (cite at least two independent sources)
A great problem statement is one sentence the user themselves would nod at.
6.4 Goals & Non-Goals
Goals: 3–5 outcomes (not features), each tied to a metric.
Non-Goals: 3–5 things explicitly excluded, with one-line rationale each.
Primary persona (one — be willing to disappoint others)
Top 3 use cases as user stories: "As a [persona], when [trigger], I want to [job], so that [outcome]."
Anti-personas if helpful (who this is not for)
6.6 Proposed Solution
The user-visible behavior. Use storyboards, wireframes, or step-by-step narratives.
Describe what the user sees and does. Avoid implementation details; link out to the
eng design doc instead.
If multiple solution approaches are viable, pick one as the recommendation and put
the others in §6.9 (Alternatives Considered).
6.7 Requirements
The numbered, testable list. Each requirement:
Has an ID (R1, R2, ...) so reviewers and tickets can reference it
Is written as a user-observable behavior, not an implementation
Has a priority: Must / Should / Could (MoSCoW) or P0 / P1 / P2
Is falsifiable — a tester could decide pass/fail in <30 seconds
For analytics/data products especially, include explicit data requirements:
events to track, dimensions, retention, PII handling, downstream consumers.
6.8 Success Metrics
The exam the launch will be graded against. Include:
North-star metric: the one number we will be judged on
Input metrics: 2–3 leading indicators we can move week to week
Guardrail metrics: what must not get worse (latency, error rate, revenue, CSAT)
Measurement plan: how each is instrumented, where dashboards live, who reviews
If the launch is going through experimentation (A/B, holdout), specify the design
here at a high level: unit of randomization, MDE, expected sample, decision rule.
6.9 Tradeoffs & Alternatives Considered
For each meaningful decision, a short row:
Decision: [the choice]
Alternatives: [what else was considered]
Why this: [the reason we chose it]
What we give up: [the cost of choosing it]
Reviewers will ask "why not X?" — answer the top 3 in the doc, in advance.
6.10 Risks & Mitigations
A small table:
Risk
Likelihood
Impact
Mitigation
Owner
Include at least one risk in each of: technical, user, business, regulatory/legal.
Be honest. A PRD with no real risks is a PRD that has not been thought through.
6.11 Rollout Plan
Internal alpha → beta → GA stages
Feature-flag / cohort strategy
Communication plan (internal, customer, support enablement)
Rollback criteria — how and when we'd reverse the change
6.12 Open Questions
Numbered, with an owner and a target resolution date. Open questions are not weakness
— they are the active edge of the doc. A PRD with zero open questions late in
discovery is suspicious; one with thirty is unfinished.
6.13 Appendix
Research links and verbatims
Competitive analysis
Earlier explorations
Glossary (if domain-specific)
Decision log (entries with date, decision, rationale, who decided)
7. The PR/FAQ Variant ("Working Backwards")
When the user asks for a PR/FAQ, Amazon-style, replace §6.1–6.6 with:
Press Release (1 page)
Headline (the customer benefit, not the technology)
Subhead (who and what)
Summary paragraph (location, date, value prop)
Problem paragraph
Solution paragraph
Internal quote (leader)
Customer quote (representative)
Call to action / availability
External FAQ — what a customer would ask
Internal FAQ — what the team would ask (this is where the hard tradeoffs live)
Then keep §6.7–6.13 as-is in an appendix-style section called "Implementation
Details" so eng/data partners have what they need.
The PR/FAQ is the right form when the product is net-new or customer-facing
with positioning risk. It is overkill for internal tooling or incremental features.
Process (§8): Writing Workflow
Follow this order. Do not jump ahead.
Intake and mode — parse $ARGUMENTS, classify quick / standard /
deep / revision, identify the audience, decision, and artifact path.
Evidence pass — read the provided materials and record what supports the
problem, user, metrics, constraints, and risks. Mark missing evidence before
drafting conclusions.
Frame — fill in metadata, write a placeholder TL;DR (you'll rewrite it last).
Problem first — write §6.3 Problem with evidence. Stop. Show it. Iterate
until the user nods. If you skip this step, the rest will be wrong.
Goals & Non-Goals — define success and explicitly cut scope. Often the
hardest 30 minutes of the whole doc; do not rush.
Users & Use Cases — concrete personas and stories.
Solution & Requirements — now you may describe the build.
Metrics, Tradeoffs, Risks, Rollout — make the bet legible.
Decision gates — stop for the user on format/size, scope cuts, unresolved
metric choice, regulated/legal risk, or cross-functional ownership conflicts.
Open Questions — list every uncertainty surfaced during writing.
Rewrite the TL;DR last — it should now write itself.
Self-review against §10 checklist, then cut 20%. Almost every PRD is
20% too long on the first pass. Be ruthless.
9. Tone & Format Conventions
Voice: plain, direct, declarative. Avoid hedging ("we might consider possibly
exploring..."). If the team has not decided, say "Open question: ...".
Tense: present tense for the proposed end state ("The user sees..."), future
tense only for rollout milestones.
Acronyms: define on first use, even internal ones. Future readers will thank you.
Markdown: use H2 for top-level sections, H3 for subsections. Tables for
structured data (requirements, risks, metrics). Avoid deep nesting.
Links: prefer linking to research, dashboards, and design files over
embedding screenshots. Embeds rot.
Length signal: if a section is longer than ~400 words, split it or move
detail to the appendix.
No emoji in the doc body. They date the artifact and read as unserious.
No buzzwords: "synergy", "10x", "delight", "best-in-class", "world-class",
"seamless", "leverage" (as a verb). If a word would survive being deleted, delete it.
10. Quality Bar — Self-Review Checklist
Before declaring a PRD done, the skill must verify all of the following. If any
fail, fix and re-check.
TL;DR test: can a smart outsider read only the TL;DR and ask a useful
follow-up question? (Not "what does this mean?" — that's a fail.)
Problem test: is the problem statement one the user would agree with,
in their words, with cited evidence?
Falsifiability test: does every goal have a number, baseline, target, and
timeframe? Could a stranger declare success or failure 90 days from launch
without ambiguity?
Non-goals test: are there at least three non-goals, and are at least
two of them plausibly tempting?
Reversal test: if you flipped any single requirement, would the doc
still make sense? (If yes, that requirement is filler — cut or sharpen.)
Tradeoff test: is the hardest tradeoff named explicitly, with what we
give up by choosing this path?
Evidence test: is every "users want X" claim cited or labeled
[ASSUMPTION]?
Risk test: is there at least one uncomfortable risk listed (i.e., one
that, if surfaced in review, would prompt a real conversation)?
Rollback test: is it clear how this gets reversed if it goes wrong?
Length test: every section ≤ ~400 words; total ≤ 15 pages; passed the
"cut 20%" pass.
Owner test: every open question and risk has a named owner.
5-year test: would this doc still be readable and informative to someone
onboarding to the area three years from now?
Anti-Patterns (§11)
These are the ways PRDs fail in practice. Catch them in your own drafts.
Solution masquerading as problem. "Users need a dashboard." No — users have
a job they're failing to do; a dashboard might be the answer. Reframe.
The everyone persona. "Our users are anyone who uses email." If the persona
fits everyone, the product fits no one.
Missing non-goals. Every absent non-goal becomes scope creep later.
Feature soup. A long bullet list of features with no prioritization or
justification. Convert to numbered, prioritized requirements.
Implementation leak. Database schemas, API endpoints, framework choices in
the PRD. Move to the eng design doc.
No tradeoffs named. A PRD where every choice was "obvious" hasn't been
thought hard enough.
Optimistic-only risk section. "Risk: low adoption. Mitigation: marketing."
That's not a mitigation, that's a wish.
Stale TL;DR. The TL;DR was written first, then the doc evolved, and now
they disagree. Always rewrite the TL;DR last.
Doc-as-decision-record without a decision. A 12-page PRD that ends with
"we'll figure out the metric later" is not a PRD.
Output Format (§12)
When invoked, this skill produces a single markdown file. Conventions:
Filename:prd-<kebab-case-name>.md unless the user specifies otherwise.
Location: save to the user's preferred directory; if none, default to the
current working directory as ./prd-<kebab-case-name>.md. When revising an
existing PRD, write beside the source unless the user names another path.
Do not paste the entire PRD into chat unless explicitly asked — point the
user to the file and offer a brief summary plus the top 3 open questions.
Section anchors: use stable H2 headings (verbatim from §6) so other tools
and skills can address sections by slug.
Status field: always include and set to Draft on first write.
Do not invent specifics. Names, numbers, dates, citations, customer quotes
— if not provided or verifiable, use [TBD] or [ASSUMPTION]. Fabricated
specifics are the single fastest way to get a PRD rejected by a real team.
Completion status: end the handback with one of:
DONE (artifact saved and checklist passes), DONE_WITH_CONCERNS (usable
artifact with visible assumptions or unresolved low/medium-risk gaps),
NEEDS_CONTEXT (blocked on user/source input before a responsible draft),
or BLOCKED (cannot proceed because required source access, policy, or
tool capability is unavailable).
13. Iteration & Reviews
PRDs are living documents. When the user asks to revise:
Preserve the decision log at the bottom of the appendix. Every revision
that changes a goal, non-goal, or requirement gets a dated entry: what
changed, why, who decided.
Bump the Last updated field.
If the change is material (scope, metrics, target users), move status back to
Draft or In Review and flag the affected sections with a [CHANGED YYYY-MM-DD]
marker for the next round of review.
When the user asks to review an existing PRD for approval readiness, route to
prd-review. Stay in this skill only when the user asks for direct PRD revision,
rewriting, or section authoring.
14. Worked Examples
For calibration, see references/examples.md — two before/after pairs (a weak vs.
strong problem statement, and a vague vs. falsifiable goal) that show the
specificity, evidence, and measurability this skill demands. Read them when a
draft's problem statement or goals feel thin, and hold your output to that bar.
15. When To Push Back
This skill is allowed — and expected — to push back on the user when:
They ask for a PRD but have not articulated a problem. Ask for it.
They ask for a "quick PRD" for something that is genuinely a multi-quarter bet.
Recommend the deeper format and explain the cost of underdocumenting.
They want to skip non-goals or success metrics. Refuse politely; offer
[TBD — needs decision] placeholders so the gaps are visible rather than hidden.
The "users" are described generically. Ask who specifically, and what evidence
supports it.
A skill that produces only what is asked for produces mediocre PRDs. A skill that
produces what the team actually needs produces spectacular ones. Err toward the
latter, with a light touch.
