Agent 12 · AI Product PRD Generator
How a constrained system prompt and structured delimiters turn a one-sentence idea into a stakeholder-ready PRD — every time.
The Problem
Writing a PRD is one of the highest-leverage activities in early product development — and one of the most consistently skipped. The blank page problem is real: most product ideas never get formally structured because the overhead of writing a proper document feels disproportionate to the early-stage uncertainty of the idea.
The result is misaligned teams, scopes that drift, and features built on assumptions that were never made explicit. The problem isn't that PMs don't know how to write PRDs — it's that the activation energy is too high when an idea is still forming.
Design Approach
Agent 12 is designed around a single constraint: one input, four outputs, always. The system prompt enforces a rigid four-section structure — Problem Statement, User Stories, Acceptance Criteria, and Success Metrics — using explicit delimiters that the frontend parses into rendered cards.
This is a deliberate departure from free-form generation. Unconstrained PRD output tends toward verbose, meandering prose that still requires significant human editing. By forcing the model into a strict schema, every generation is immediately usable — not a draft to be redrafted.
The four sections were chosen deliberately. Problem Statement grounds the feature in user reality. User Stories force role and motivation to be made explicit. Acceptance Criteria translate intent into testable conditions. Success Metrics prevent the classic failure mode of shipping a feature with no definition of success.
System Prompt Architecture
The system prompt stays under 300 tokens by doing three things: naming the persona, specifying the exact delimiter format, and stating the output rules tersely. No examples, no lengthy context — just structure and constraints.
Delimiter Pattern
Each PRD section is wrapped in ===SECTION=== / ===END=== blocks with explicit LABEL and CONTENT fields. The frontend splits on this pattern and renders each into a styled card — separating display logic from generation logic entirely.
Format Rules
User Stories follow a mandatory "As a / I want / So that" template. Acceptance Criteria must be testable conditions. Success Metrics must include measurable targets. These rules are stated in the prompt, not left to model discretion.
Assumption Handling
Rather than stalling on vague input, the model is instructed to proceed with assumptions and surface them after the four sections. This converts an ambiguous input into a concrete proposal the user can react to.
Token Budget
800 token output cap forces the model to be precise. A complete four-section PRD fits comfortably — the constraint discourages verbose generation without sacrificing coverage of all four required sections.
What I Learned
Structure beats flexibility for tool-like agents. The most useful version of this agent isn't the most creative one — it's the one whose output is predictably formatted and immediately actionable. Rigid delimiters in the system prompt are what make the frontend parse reliable.
Token constraints improve output quality. Counterintuitively, the 800-token cap produces better PRDs than unconstrained generation. It forces economy of language — acceptance criteria become sharper, success metrics become more specific.
Assumptions-first beats clarification-first. Testing revealed that asking a clarifying question before generating felt like friction. Generating with stated assumptions and inviting correction felt collaborative. Users preferred reacting to something concrete over answering abstract questions.
Typical Token Profile
A single PRD generation for a moderately detailed product idea typically produces: