AI PRD Generator
Fill in 6 sections. Get a structured AI PRD as markdown — ready to paste into Notion, Linear, or any doc tool. Built for AI features specifically: includes eval rubric, cost model, and edge-case slots that standard SaaS PRDs skip.
2-5 words. The thing you are building.
Who is hurting, what hurts, how much. Numbers if you have them.
One sentence. Their role and what they care about.
"If we build X, then [user] will [outcome] because [reason]."
What the AI does. Model, inputs, outputs, confidence thresholds.
How you will know it is working. Eval suite + production sampling.
0-100. What pass rate gates the launch.
From your model + token estimate. Eg. 0.07 cents = $0.0007/req.
Cost per request × requests per user per month.
Cap so a power user or jailbreak can't blow the bill.
Multi-issue inputs, non-English, prompt injection, empty input, abusive content.
What happens when the model is uncertain or the API fails.
If the launch is broken, how do you turn it off in 60 seconds.
# [Feature name] > AI Product Requirement Document. Generated by [ShipSet AI PRD Generator](https://shipset.app/tools/ai-prd-generator). ## 1. Problem _Describe who is hurting, what hurts, and how much (with numbers when possible)._ **Primary user:** _Who you are building for._ ## 2. Hypothesis _If we build X, then [user] will [outcome] because [reason]._ ## 3. Solution _What the AI does. Model, inputs, outputs, confidence thresholds._ ## 4. Eval rubric (the section most PMs skip) **Target pass rate:** 90% on the 50-row eval suite. **Eval suite shape:** - Happy path: 20 rows of typical, well-formed inputs. - Edge cases: 15 rows of rare-but-expected inputs (multi-language, very short, very long, ambiguous). - Adversarial: 10 rows designed to break the feature (prompt injection, conflicting instructions, off-topic). - Empty / malformed: 5 rows testing graceful failure (empty, whitespace, non-text, profanity). **Scoring:** binary pass/fail per row, with optional 1-5 gradient for generation tasks. **Rerun cadence:** every prompt change, every model version change, every two weeks regardless. **Success metric beyond eval:** _How you will know it is working in production._ ## 5. Cost model | Variable | Value | |---|---| | Cost per request | ~$[per request] cents | | Cost per user per month (p50) | ~$[per user/month] | | Cost per user per month (p95 estimate) | ~$[2x p50] | | Hard rate limit per user per day | 50 requests | **Implications:** - Soft cap warning to user at 80% of daily limit. - Hard cap returns "limit reached, try again tomorrow" message. - Variance > 3x p50 triggers a per-user investigation in production logs. ## 6. Edge cases _List at least 5. Adversarial inputs go here._ ## 7. Graceful degradation _What the feature does when the model is uncertain, the API fails, or output is suspect._ ## 8. Rollback plan _How you turn this off in 60 seconds if launch breaks._ --- _Generated with the [ShipSet AI PRD Generator](https://shipset.app/tools/ai-prd-generator). This template format is taught in ShipSet Lesson 16. Free, no login required._
Why standard PRD templates break for AI features
A SaaS PRD has deterministic acceptance criteria ("returns the user's last 10 invoices"). An AI feature has probabilistic ones ("correctly tags 92% of support tickets"). The template needs three sections that SaaS PRDs do not have:
- · Eval rubric: how you will score outputs, with sample rows and target pass rate.
- · Cost model: per-request, per-user-per-month, p95 variance, and the hard rate limit.
- · Graceful degradation: what the feature does when the model is uncertain or wrong.
This generator includes all three. The output is a real working PRD, not a slide.
Want to learn how to use this in production?
ShipSet teaches AI PRD writing across Lessons 15-22, with critique from Captain (an AI mentor) on every section. By Day 22 you have a full AI PRD for the feature you will ship by Day 90. The PRD is one of 10 portfolio artifacts at the end of the program.