How to Build a PRD Skill in Claude
Every PRD starts the same way. A blank page, or a template you've already outgrown, or a Claude prompt where you re-explain your team's format for the fourth time this month. A PRD Skill built inside Claude that knows your team's format, tone, and bar fixes that. Here's how to build one.
Why a PRD Skill Beats a PRD Prompt
A prompt is a fresh start every time. You type out the format. You re-explain that your team leads with the customer problem, not the solution. You paste a recent example so the tone matches. By the time Claude is up to speed, you could have written half the doc yourself.
A Skill is memory. You write the instructions once, and Claude pulls them up next time you say "draft a PRD" or "write me a product spec." The format is right. The tone is right. The section order matches what your team expects. You're not re-teaching, you're just editing.
There's a compounding effect too. Every time you notice something off in the output, you sharpen the Skill. Maybe your team always cites a customer quote in the problem section, and Claude is making them up instead of leaving a placeholder. You add one line to SKILL.md. Now it never happens again.
What you're really building is a reusable spec for "what a great PRD looks like at our company." That's a useful thing to have written down, even setting Claude aside.
A quick note. Skills work in both Claude and Claude Code, and the SKILL.md format is identical. If you're searching for "claude code prd," the build process below works the same way. You just drop the Skill into your .claude/skills/ folder for Claude Code, or upload it as a zipped folder in Claude.ai.
What Your PRD Skill Should Know About Your Team
Before you open a text editor, write down the parts of your team's PRD culture that live in everyone's head but aren't formally documented anywhere. This is the work most people skip, and it's why most AI-drafted PRDs feel generic.
A few things to capture:
Format conventions. What sections does every PRD have, in what order? Problem statement, user stories, success metrics, requirements, non-goals, open questions? Do you use bullets or prose? Is there a TL;DR at the top, or does the reader have to earn it?
Quality bar. What makes a PRD get shipped vs. sent back? Is it specificity in the success metrics? A clear non-goal? A named customer quote in the problem statement? Write down the three things your best PRDs do that the mediocre ones miss.
Cultural cues. Some teams open with the metric. Others open with a customer story. Others open with the strategic bet. Yours has a default, even if no one has said it out loud. Name it.
Definition of done. When is a PRD ready for review? Filled-in success metrics? Engineering sign-off on feasibility? A linked Figma? Skills get sharper when "done" is concrete.
This piece matters more than the technical setup. A Claude PRD template that knows your format but not your bar will produce something that looks right and reads wrong.
The SKILL.md File for a PRD Generator
A Skill is a folder. Inside that folder is one required file: SKILL.md. The folder lives in your .claude/skills/ directory for Claude Code, or gets uploaded as a zipped folder to Claude.ai. Here's a working example to copy and adapt.
The YAML frontmatter at the top is the trigger. The description is what Claude reads first to decide whether to load the Skill at all, so it needs to be specific about when to use it. Vague descriptions either fire constantly when they shouldn't, or never fire when they should.
The body is the actual instruction set. Keep it tight. Long, hedging instructions confuse the model. Short, declarative rules work better.
Adding a Template and Example for Consistent Output
The Skill folder should hold two more files alongside SKILL.md: TEMPLATE.md and EXAMPLE.md.
TEMPLATE.md is the empty scaffold. Section headers, standard subsections, any boilerplate. Claude fills it in. Why this matters: Claude follows structure much better when it has a structure to fill than when it generates one from scratch. Output gets dramatically more consistent the moment you add a template.
EXAMPLE.md is one real PRD your team actually shipped, with any sensitive details swapped out. This is the quality anchor. The Skill instructions say "match the voice and depth of EXAMPLE.md," and Claude uses it as a reference point for what a finished doc should feel like.
Picking the right anchor matters. Don't pick your most ambitious PRD. Pick the one that represents the median your team aims for, the one a new PM would learn the most from. If you anchor too high, every output will feel overwrought. Anchor on the steady-state quality bar instead.
Testing and Iterating Your PRD Skill
Before sharing the Skill with anyone, test it on three real product problems your team has worked on. Run each one and look for two things.
First, did the trigger fire? Type "draft a PRD for X" and watch whether Claude actually invokes the Skill, or whether it ignores it and freelances a generic doc. If invocation is unreliable, the description in your YAML frontmatter is too vague. Tighten it with the specific phrases your team uses ("write a spec," "do a PRD for," "turn this brief into a PRD").
Second, where does the output drift from your team's bar? Maybe it's inventing customer quotes instead of leaving placeholders. Maybe the success metrics are still too soft. Maybe the TL;DR keeps creeping past 80 words. Each of these is one new line in SKILL.md.
This is also where understanding why the feedback under your PRD matters becomes important. A PRD is only as good as the customer signal it's built on. A Skill that drafts a beautiful doc from stale or thin input is still drafting the wrong thing well.
Update the Skill whenever your team's conventions shift. New section in your PRD template? Add it. New rule about non-goals? Add it. The Skill is a living spec.
Where to go from here
A PRD Skill is the first half of the workflow. The second half is making sure the customer input feeding it is real, current, and specific. The natural next step is building a user research Claude Skill that pulls real customer language into your draft instead of letting Claude guess at it.
Frequently asked questions about building a PRD Skill in Claude
What is a PRD Skill in Claude?
A PRD Skill is a folder containing a SKILL.md file with instructions Claude follows when drafting product requirements documents. The instructions encode your team's format, tone, and quality bar. Once installed, Claude uses the Skill automatically every time you ask for a PRD, so you stop re-explaining the format on every draft.
How is a PRD Skill different from a PRD prompt?
A prompt is a one-time instruction you type into the chat that disappears when the session ends. A PRD Skill is a reusable instruction set Claude loads automatically when your request matches its description. The Skill persists across every conversation, can pull in supporting files like a template and an example, and gets sharper every time you refine it.
What files do you need to build a Claude PRD template?
Three files inside one folder. SKILL.md holds the instructions and the trigger description. TEMPLATE.md is the empty scaffold with your section headers. EXAMPLE.md is one real PRD your team actually shipped, used as the quality anchor. The Skill instructions reference TEMPLATE.md for structure and EXAMPLE.md for tone.
Does the PRD Skill work in both Claude and Claude Code?
Yes. The SKILL.md format is identical between Claude (the web and desktop app) and Claude Code (the CLI). For Claude Code, drop the Skill folder into your .claude/skills/ directory. For Claude.ai, zip the folder and upload it. Searches for "claude code prd" and "claude prd template" lead to the same build process.
Why does my Claude PRD Skill keep making up customer quotes?
Two fixes. First, add a rule to SKILL.md saying that if no customer quote is available, Claude should leave a clearly marked placeholder like [QUOTE NEEDED] instead of inventing one. Second, point the Skill at a real customer feedback source in the discovery step. A PRD Skill connected to real customer language outperforms one drafting from Claude's guesses, every time.
