Template / Product planning / Introductory

A Better Brief for Any Web-Enabled AI Feature

Fill this in before anyone writes retrieval code. Most disagreements about model quality turn out to be unanswered questions on this page. It is a planning aid, not a specification standard.

For Product managers, founders, and engineering leads

01The job

The user's question, written the way a user would actually type it
The decision or action the answer supports
What the user does today instead, and how long it takes them
What counts as a good answer, described concretely enough to grade

02Sources and scope

Named source types, in priority order, with any domains explicitly in or out
Whether an open search is permitted or the source set is fixed
Maximum acceptable age per claim type (pricing, policy, documentation, commentary)
Sources deliberately excluded, and the reason

03Output

Shape of the result: prose, structured record, table, or digest
Citation level required: source list, per-claim link, or per-claim excerpt
Length or field limits the consuming surface imposes
What the feature should say when it cannot answer

04Quality and escalation

Minimum number of independent sources for a material claim
Behaviour when sources conflict
Behaviour when evidence is thin: refuse, answer with a caveat, or escalate to a human
Who reviews outputs before they are shown, if anyone, and on what sample

05Constraints

Latency budget per request and cost ceiling per answer
Rate limits, terms, and access rules that apply to the named sources
Personal data that must never be retained, even incidentally
Logging and retention policy for retrieved content and excerpts

06Before build starts

Ten example questions written down, with the answer you would accept for each
One of them walked through by hand, end to end, with real URLs
The failure behaviour agreed in writing by product and engineering
A named owner for source-set changes after launch

productplanningretrievalscope