--- feature_id: FTR-35 title: Prompt Vetting Checklist category: Research optional: true depends_on: FTR-28 source: Not from Project_Kickoff_Seed_Prompt_v2.3.md — supplied directly by the user, drafted from a real gap found in production use on the Sluix AEPM pipeline project (see that project's U3_Content_Research_Spawner.md SOP, `Prompt Vetting Checklist` section, for the full worked version). No corresponding item in Bootstrap_Function_Registry.md. See coverage_report.md Section 10. registry_refs: [] --- # Prompt Vetting Checklist Depends on: [FTR-28 — Multi-AI Research Synthesis System](../items/multi-ai-research-synthesis.md) ## What this is A structured, five-question review a research prompt goes through before it's considered genuinely ready to submit to a multi-AI research run. FTR-28 covers what happens once a prompt is good — save it consistently, submit it to multiple platforms, consolidate results into one authoritative master document. It has no step for checking whether the prompt is actually good before that point. This feature is that missing step: a deliberate second read against five fixed structural questions, distinct from a general "does this look done" skim. ## Why this exists A prompt can look complete — full sections, plausible objectives, reasonable length — and still have real, non-obvious gaps that only surface from reading it against specific structural questions, not from a surface skim. In one real case, a prompt spanning three disciplines (legal, electrical, governance) had gone unreviewed past a first read. Walked against the five checks below, it turned out: no expert role was specified for a topic spanning three genuinely different disciplines; a practical follow-up question implied by the core research question was never captured in any objective; one objective had silently compressed a stable fact and a volatile fact into a single bullet, which reliably produces a shallow pass on both; and nothing flagged which facts would age fastest, risking a stable statute and a rapidly-changing incentive program being treated with the same one-time-stated confidence. None of these would surface from a completeness skim. They only surface from checking the prompt against named structural questions. ## Assess first - Confirm FTR-28 (Multi-AI Research Synthesis System) is installed — this feature is a quality gate inside that pipeline, not a standalone tool. - Read the research prompt in full before applying any of the five checks — this is a deliberate second read, not a skim. ## Ask, only if needed - If it's not clear which prompt is meant: "Which research prompt would you like vetted?" - Nothing else — the five checks below don't require user input to apply; they're a reading discipline the AI applies directly to the prompt text. ## The five checks 1. **Role & framing** — does the prompt state what kind of expert the research AI should be, matched to what the topic actually spans? Multi-discipline topics need this stated explicitly, or one discipline crowds out the others. 2. **Goal coverage** — read the core research question, then the objective list, and check whether the objectives actually cover everything the question implies, not just its most obvious parts. 3. **Objective granularity** — for each objective, is it genuinely one research question, or two different threads compressed together? Compression is especially risky when one half is stable and the other is volatile. 4. **Cross-reference correctness** — where the prompt references sibling or prior research, does it name the specific source, state what's already covered at what depth, and say clearly where this prompt's job picks up — avoiding both duplication and gaps? 5. **Staleness/currency flagging** — are the fastest-aging facts in the prompt explicitly flagged with a "verify current as of / re-check before reuse" instruction, distinct from the prompt's stable, slow-moving facts? ## Write Walk the prompt against all five checks, in order. For each one, state plainly whether it passes as-is or what's missing. Where something is missing, propose the specific addition or edit rather than just flagging the gap abstractly — a vague "objective 3 might need work" is not a completed check; a concrete rewrite of objective 3 is. A prompt passes when someone has actually walked it through all five and can state plainly what — if anything — changed as a result. Passing with zero changes is a legitimate, recordable outcome; skipping the check silently is not. If the project has a way to record this (a tracker row, a status field, a note in the prompt file itself), record that the prompt was vetted and what came of it — but this feature doesn't require or install any particular tracking mechanism of its own; it works with whatever the project already has, including nothing more than a note in conversation. ## Trigger On request only — "vet this prompt," "double-check number six," "run the checklist on [topic]" — or as an explicit step before marking a prompt ready when the person asks for that specifically. Not run automatically on every prompt in a batch scan; the five checks are a deliberate second read, not a background linter. ## Confirm Confirm: the Prompt Vetting Checklist is installed as a gate inside the FTR-28 pipeline — say "vet this prompt" or similar any time before submitting a research prompt for multi-platform research. Want to install another, or is that everything for now?