A step-by-step way to work
Pick the right spec type for the job
Tell AI which kind of spec you need and who reads it — the format differs for a functional spec vs. a data spec.
Give it the real constraints
Supply the system facts it can't know; otherwise it invents plausible ones.
Draft, then translate
Get the detailed spec, then ask it to produce a stakeholder-friendly summary or a more detailed breakdown as needed.
Cross-check against criteria
Ask it to find inconsistencies between the spec and your acceptance criteria or PRD.
Verify with engineering
The spec isn't done until the people who'll build it confirm it matches the system.
Real prompts, and what good output looks like
Turn this high-level intent into a functional spec a developer could work from: 'Let users undo their last action.'
Functional spec — Undo last action: • An 'Undo' control appears after any reversible action (specify which actions are reversible). • Clicking Undo reverts the single most recent reversible action and shows confirmation. • Undo is available for [time window / until next action — decide]. • Non-reversible actions (e.g. permanent deletes) do not show Undo. • Open question: is multi-level undo in scope, or single-step only?
I've flagged the decisions you need to make — reversible-action list, time window, multi-level.
A prompt you can reuse
Fill in the highlighted parts and paste it into your AI tool of choice. Edit the output — it's a starting point, not a finished answer.
Help me draft a [spec type: functional / data / technical-adjacent] specification for [reader: a developer / a stakeholder]. The feature/intent: [describe what's being specified] System facts you should use (don't invent others): [real constraints, existing behaviour, data available] Please: 1. Draft the spec in the appropriate format for that type and reader. 2. Mark every assumption and every decision I still need to make as a bracketed open item. 3. Flag anything that would need engineering to confirm against the real system.
What AI gets wrong here
Invented system behaviour
AI describes how it imagines your system works, which may be wrong.
Wrong spec type for the audience
A developer spec and a stakeholder spec aren't interchangeable; AI may blur them.
False completeness
A tidy spec can look finished while hiding undecided questions.
AI accelerates the writing of a spec, not the knowing of your system. It's strong at structure and at translating between levels of detail, but every confident claim about how things work needs your and your engineers' verification. Use it to draft and translate; reserve 'this matches reality' for the people who know the reality.