[CURRENT PROMPT]: Summarize this meeting transcript. [ACTUAL OUTPUT]: The team discussed the Q3 product launch, reviewed marketing plans, and talked about budget concerns. Several action items were assigned to team members. The group agreed to follow up next week. [DESIRED OUTPUT]: A summary I can paste into our project tracker: every decision made, every action item with an owner and a deadline, and open questions listed separately. **Diagnosis:** 1. **Missing:** a format spec (the tracker needs sections, not prose), a definition of what counts as a decision versus a discussion topic, an instruction to extract owners and deadlines, and a rule for missing data — the current prompt gives the model nothing to be wrong about, so it retreats to safe generalities. 2. **Too much of:** nothing — the prompt is under-specified rather than over-written. "Summarize" is the one vague verb doing all the work, and it invites narration instead of extraction. 3. **Techniques:** structured output with fixed headings; extraction framing ("extract" instead of "summarize"); an explicit null-handling rule so the model does not invent owners or dates. **Improved prompt:** Extract structured notes from this meeting transcript for a project tracker. Transcript: """ [paste transcript] """ Output exactly three sections: ## Decisions - One bullet per decision actually made (not topics discussed): what was decided and who confirmed it. ## Action items | Task | Owner | Deadline | If an owner or deadline was not stated, write UNASSIGNED / NO DATE. Do not infer. ## Open questions - Questions raised but not resolved. Rules: no narrative summary, no "the team discussed..." phrasing. If a section is empty, write "None recorded." **Change log:** - Replaced "summarize" with "extract structured notes" — names the real task. - Added a fixed three-section format so output pastes into the tracker unedited. - Separated decisions from discussion topics — the weak output blurred them. - Added null-handling (UNASSIGNED / NO DATE) to stop the model inventing owners or deadlines. **Expected improvement:** the output becomes deterministic in shape, so the tracker paste needs zero reformatting; extraction framing forces the model to point at transcript content instead of generalizing over it; the null rule removes the main fabrication risk.