# PRD Template: How to Write a PRD an Agent Can Build From > How to write a PRD a coding agent can build from: the sections that matter, requirement words that remove guessing, and a PRD template you can copy. --- [Skip to main content](#main-content)Guide # How to Write a PRD Your Coding Agent Can Build From (Template) How to write a PRD a coding agent can build from: the sections that matter, requirement words that remove guessing, and a PRD template you can copy. **Matej Stefanik**Nov 19, 2026 · 5 min read [LinkedIn](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fclosedloop.sh%2Fblog%2Fhow-to-write-a-prd)Copy link On this page - [How to write a PRD when an agent builds from it](#how-to-write-a-prd-when-an-agent-builds-from-it) - [Requirement words that remove guessing](#requirement-words-that-remove-guessing) - [The PRD template](#the-prd-template) - [Five PRD mistakes that agents make worse](#five-prd-mistakes-that-agents-make-worse) - [Where ClosedLoop AI fits](#where-closedloop-ai-fits) [Product Management](https://closedloop.sh/tag/product+management)[AI Coding](https://closedloop.sh/tag/AI+coding)[Product Builders](https://closedloop.sh/tag/product+builders)Key takeaways - The first reader of a PRD is increasingly a coding agent, and agents fill every gap with a confident guess - Precise requirement words (MUST, SHOULD, MAY, MUST NOT) remove most of that guessing, the same way they do in internet standards - An agent-ready PRD adds three things most PRDs skip: examples and edge cases, an explicit out-of-scope list, and links to the customer evidence - Keep the problem in the customer's words, so the agent and the reviewer can both check the change against the need Learning how to write a PRD used to mean learning how to write for engineers: people who would read your product requirements document, push back on the vague parts and ask what you meant. More and more, the first reader is a coding agent. It reads every word, never pushes back and fills every gap with a confident guess, which is why the PRD template you use matters more than ever. The internet faced a version of this problem decades ago. In March 1997, Scott Bradner published RFC 2119, a document about three pages long, that did one thing: it defined exactly what the words MUST, SHOULD and MAY mean when they appear in an internet standard. It became one of the most referenced documents in internet engineering, because two companies reading the same vague spec will build two incompatible products. A coding agent is the most literal reader any PRD has ever had. Write for it the way standards bodies write for strangers. ## How to write a PRD when an agent builds from it A **product requirements document**(PRD) describes the problem, who has it, what success looks like and what the solution must do, without dictating every line of code. Its job has not changed. Its audience has. Human engineers skim a PRD, then fill gaps from conversations, context and common sense. An agent has none of that unless you give it. When a requirement says "support bulk export", a person asks how many rows, which format and what happens on failure. An agent picks an answer and writes clean, tested code around it, which is worse than no code, because it looks finished. A PRD written for a meeting "Improve the export experience" "Support bulk export" "Should be fast" The context lives in people's heads A PRD an agent can build from "Month-end exports take finance teams a whole day" "MUST export up to 50,000 rows as CSV in the background" "MUST show progress; SHOULD finish within two minutes" The context lives in the document, with links to the evidence ## Requirement words that remove guessing Borrow the RFC 2119 vocabulary. Agents follow it well, and reviewers can check against it line by line. Word Meaning Example MUST Required. The change is wrong without it. The export MUST run without blocking the page. MUST NOT Forbidden. The change MUST NOT alter the existing CSV column order. SHOULD The default. Deviate only with a stated reason. The export SHOULD finish within two minutes for 50,000 rows. MAY Optional, at the builder's discretion. The export MAY offer an XLSX format. The MUST NOT line is the most valuable one in an agent-ready PRD. Agents are eager. Given a goal, they will happily refactor a neighboring module, rename a field or "improve" something nobody asked them to touch. A short list of things that must not change saves more review time than any other section. ## The PRD template Copy this into your repository or your docs. It works for a two-hour fix and a two-week feature alike; small changes get short sections. ``` `TITLE PROBLEM What is going wrong, in the customer's words. Links to the calls, tickets or chats it came from. WHO IS AFFECTED Which customers or roles, how many, and the revenue behind them. OUTCOME The metric this should move, its baseline and the target. REQUIREMENTS MUST: MUST NOT: SHOULD: MAY: ACCEPTANCE CRITERIA Given , when , then . EXAMPLES AND EDGE CASES A normal case, an empty case, a huge case, a failure case. OUT OF SCOPE Related things this change leaves alone. CONSTRAINTS Performance, compatibility, security, data handling. OPEN QUESTIONS Anything the builder must ask before guessing.` ``` Two sections deserve a note. Examples and edge cases do more for an agent than paragraphs of description; one concrete failure case ("the export hits a deleted account halfway through") prevents a whole class of bugs. Open questions give the agent permission to stop and ask, which most agents will do when the document invites it. ## Five PRD mistakes that agents make worse - **A solution with no problem.**The agent builds the solution perfectly and nobody can tell whether it solved anything. - **Vague verbs.**"Support", "handle" and "improve" mean whatever the agent decides they mean. - **No out-of-scope list.**Agents expand to fill the space you leave them. - **No examples.**Descriptions are interpreted. Examples are matched. - **No evidence.**Without the customer's words, neither the agent nor the reviewer can check the change against the actual need. The last one matters most for the people who now write PRDs and build from them in the same afternoon. We covered that role in our piece on the [technical product manager](https://closedloop.sh/blog/technical-product-manager), and the risk of fast building without context in [Claude Code can build anything](https://closedloop.sh/blog/claude-code-agentic-coding-needs-product-intelligence). The next step beyond that is spec-driven development. ## Where ClosedLoop AI fits The hardest sections of an agent-ready PRD are the first two: the problem in the customer's words, and who is affected. ClosedLoop AI has already read the calls, tickets and chats where customers described the problem, so those two sections fill themselves from evidence instead of memory, quotes and revenue included. It can draft the business case, the PRD and the technical spec from the same evidence, for you to tighten and approve. Through [MCP](https://closedloop.sh/mcp-server), the evidence also reaches Claude Code or Cursor directly. The agent reading your PRD can pull up the customers behind it and what they said, which is the context a human engineer would have asked for. [See ClosedLoop AI inside your coding agent](https://closedloop.sh/mcp-server)or [book a demo](https://closedloop.sh/demo). Matej Stefanik CPO and co-founder of ClosedLoop AI. 20 years building and bringing products to market, from development and management to marketing. [More about ClosedLoop AI →](https://closedloop.sh/) ### Get insights like this in your inbox Product insights delivered weekly. No spam. Unsubscribe anytime. Leave this field blank Subscribe Related ## More you might find useful [Opinion Oct 29, 2026 ### Product Engineer vs Product Manager vs Product Builder: Who Owns the Outcome? What a product engineer does, how the role differs from a product manager, and why both are merging into the product builder, who owns one m… 11 min read →](https://closedloop.sh/blog/product-engineer-vs-product-manager-vs-product-builder)[Opinion Nov 17, 2026 ### Technical Product Manager: The Role AI Is Rewriting What a technical product manager does, how the role differs from a product manager, and why AI coding agents make TPMs the people closest to… 5 min read →](https://closedloop.sh/blog/technical-product-manager)[Case study Nov 6, 2026 ### How Mews Turned Its R&D Org Into Product Builders Mews moved its entire R&D organization to product builders in a single day. Here is why they did it, how it works day to day and what they l… 10 min read →](https://closedloop.sh/blog/how-mews-turned-rd-into-product-builders)[Browse all articles →](https://closedloop.sh/blog) --- ## More Information - Website: https://closedloop.sh - Documentation: /docs - Pricing: https://closedloop.sh/pricing - Contact: https://closedloop.sh/contact