Run the Adversarial "What Didn't We Ask" Pass Over Your Elicitation

Workflow area: Elicitation & Discovery · Tier: Core · Hands-on time: ~30 minutes

The requirements that sink projects are rarely the ones written badly; they are the ones never written at all. Coverage gaps are invisible from inside the material because your notes only contain the questions you thought to ask. This is one of the few analysis tasks where a model's tendency to generate exhaustively is an asset rather than a risk: armed with a checklist frame like BABOK's elicitation guidance or the Volere NFR taxonomy, it will cheerfully enumerate everything a process like yours usually needs — and the delta between that enumeration and your notes is a list of questions, not answers. The output is never "missing requirements" to paste into the pack; it is an elicitation plan for the sessions you have not run yet.

When to use — and when not to

• Use it before declaring elicitation completethe pass is cheapest right before the findings pack is baselined, when a gap costs one more workshop instead of a change request.
• Use it before writing the BRDa gap surfaced during drafting derails the document; a gap surfaced the week before becomes an agenda item.
• Use it on inherited projectsjoining mid-stream, the pass tells you in 30 minutes where the previous analyst's blind spots probably are.
• Skip it as a hallucination licenseevery gap the model proposes is a hypothesis to investigate, never a requirement to add. If a "gap" gets written straight into the pack, the playbook has been misused.
• Skip it before you have real materialrunning it on two thin interviews returns a generic checklist; the pass needs a substantive corpus to find the specific holes in.
• Skip the checklist bloatif your project is a two-week config change, interrogating it against every NFR class produces noise; scale the frame to the risk.

The recipe

1. 1.

   Assemble what you have. Gather the elicitation notes, synthesis findings, draft requirements, and the stakeholder list with roles. Add your scope statement — the pass needs to know the boundary, or it will report the whole world as a gap.

2. 2.

   Run the adversarial review. Paste the corpus into Claude or ChatGPT with this prompt:

   Copy

   You are an adversarial requirements reviewer. Your job is to find
   what was NOT asked, not to critique what was.
   Project scope: [2-3 LINES]
   Stakeholders consulted so far: [LIST WITH ROLES]
   
   Review the elicitation material below against these lenses:
   1. STAKEHOLDER COVERAGE — which affected groups (users, approvers,
      auditors, support, downstream systems' owners, regulators) never
      appear in the notes?
   2. EXCEPTION PATHS — for each happy-path process described, which
      failure/exception/reversal cases were never discussed?
   3. NFR CLASSES — performance, security, audit, accessibility,
      retention, availability, usability, compliance: which have zero
      coverage in the material?
   4. LIFECYCLE STATES — create/read/update/delete/archive/purge for
      each major entity: which states are undiscussed?
   5. VOLUME & TIMING — peaks, seasonality, cutoffs, migrations
      never quantified.
   
   For each gap: name it, cite the closest thing the material DOES
   say (or state "no mention anywhere"), rate likely impact
   HIGH/MED/LOW for this scope, and write the specific question to
   ask and WHO to ask it to.
   Rules: do not invent answers to the gaps. Do not report gaps
   outside the stated scope. Output is questions, not requirements.
   
   Material:
   [PASTE NOTES / FINDINGS / DRAFTS]
   

3. 3.

   Triage the hits like a skeptic. Good output looks like this:

   Copy

   EXCEPTION PATHS — HIGH
   Gap: vendor onboarding rejection/rework path never discussed.
   Closest mention: S2 finance workshop describes approval steps
   only ("then it goes to the second approver", 22:04).
   Question: "What happens when the second approver rejects — who
   is notified, can the vendor resubmit, is prior data retained?"
   Ask: finance manager + ops lead together (they may disagree).
   
   STAKEHOLDER COVERAGE — MED
   Gap: internal audit never consulted; notes mention "audit
   finding" (S2, 08:17) but no auditor was in any session.
   

   Kill the false positives: gaps that are genuinely out of scope, things covered in documents the model did not see, and generic checklist echoes with no specific hook into your material. Expect to keep half — that half is gold.

4. 4.

   Convert survivors into an elicitation plan. Group the surviving questions by the person who can answer them, order by impact rating, and book the follow-ups — a 30-minute call per stakeholder group usually clears a dozen gaps. The gap list itself goes into the project log as evidence of coverage diligence.

5. 5.

   Re-run after the follow-ups. Feed the new session notes back through the same pass. When a run produces only LOW-impact or out-of-scope hits, you have a defensible claim that elicitation is complete — which is a sentence worth being able to say in a sign-off review.

Make it yours

• Load your real checklist as context: if your org uses BABOK task checklists, Volere templates, or a regulatory question set, paste it into the prompt as the authoritative lens — the pass gets sharper and speaks your governance language.
• Domain-tune the lenses: for a payments project add reconciliation, disputes and regulatory reporting as named lenses; for HR add consent and works-council constraints. Reusable per-domain prompts compound.
• Claude Project as a standing reviewer: keep the corpus in a Project and re-run the pass at each stage gate; the diff between runs shows the coverage debt burning down.
• Pair with pb-01 synthesis: the NOT SAID section of your synthesis output is a lightweight version of this pass — promote its entries into the full adversarial review rather than maintaining two lists.

Quality check

• Every gap cites the nearest misseach finding names what the material does say (or "no mention anywhere"), so a reviewer can verify the hole is real.
• No gap became a requirementthe output contains questions and owners, and nothing from the pass was written into the pack without a stakeholder answering first.
• False positives were culled with reasonsdismissed gaps carry a one-line reason (out of scope, covered in doc X), so the triage is auditable.
• High-impact gaps have booked follow-upsa named person, a date, and the specific question; a gap list without meetings is a guilt list.
• The scope boundary heldthe pass reported nothing outside the stated scope, and anything borderline was flagged to the sponsor rather than silently included.
• A clean re-run existsthe final pass after follow-ups is filed with the findings pack as evidence of elicitation coverage.