Recommendation Adversarial Stress-Tester

A stress-test exists to name who will attack the recommendation, and with what, before the steering committee. It is not a rewrite. The moment the prompt says "make this more palatable," "soften for the CFO," or "give me a safer version," you have hired a ghostwriter. The advice gets quieter. The meeting does not.

The output is a cast list: objection, who raises it, the evidence they would cite, and whether answering it would change the recommendation or only the implementation design. A partner still decides which to answer, which to carry, and which are noise. The slides stay as written until that decision is made.

This is not a bid red-team. A proposal red-team reviewer scores a draft against a published RFP sheet: criterion IDs, weights, mandatory gates. Mixing those prompts is a common failure. You get evaluator-speak ("lack of innovation") on a client recommendation, or political gossip on a scored bid. Keep the jobs apart.

General assistants do this work: Microsoft Copilot on the Word or PowerPoint pack, ChatGPT or Claude with the draft plus a constraint memo pasted in. There is no special recommendations product to buy. If the option is still a matrix, stop. Finish the strategic option comparison synthesizer and pick a preferred path. Stress-test the recommendation you are actually taking into the room.

Without named stakeholders and stated constraints, the model fabricates a steering committee that sounds senior and is not yours. It invents a Chief Transformation Officer or a digital sponsor because those titles appear in generic change-management writing. Partners then waste a cycle answering people who will not be in the room.

Feed the model, in this order:

1. The draft as it will be shown (slides, memo, or both). A three-bullet summary hides the objection in a footnote, a phasing chart, or a missing owner.
2. Constraints the client has already stated: budget freeze, works-council consultation, a prior failed program, a board red line. Label each as stated by the client, not inferred by the team.
3. Named people with known positions: title, what they have already said, and where you heard it. Positions from a multi-stakeholder interview transcript analyzer belong here only if you can still attribute them without breaking what you promised interviewees.
4. Implementation risks the team already believes are real (close calendar, dual-run, data migration). Keep these apart from political constraints. The fix for each is different.
5. The ask: objections mapped to who would raise them. No rewrites. No new stakeholders. Mark uncertainty when you did not name someone.

If you only have roles ("finance," "ops") and not names, say that, and tell the model to flag invented individuals. A finding that names a person you never mentioned is a hallucination. Discard it.

A client organizational readiness classifier rates whether the organization can execute: capacity, skill gaps, competing programs. This pass rehearses how named people will fight the recommendation in the room. Do not collapse those jobs.

A first rehearsal shows the difference between an objection you should answer and advice the model is trying to water down. Take a mid-size manufacturer. The team is recommending that three regional finance centers become one shared-services hub, with a staged migration and a dual close for two cycles. The preferred location is already on the slide.

Known constraints in the prompt:

• The EU site requires works-council consultation before any role moves.
• The CFO sponsored an offshoring that missed a quarter-end close. She has said she will not "risk the close again for a location story."
• A regional managing director's bonus still includes local finance headcount. He argues that service levels die when the team is not on site.
• The board previously rejected a "lift and shift" proposal. The phrase is radioactive even though this pack is not that proposal.

Named people: CFO, regional MD, CHRO, CIO, sponsor COO. Ask, for each, the opening objection, the evidence they would cite, whether it attacks the recommendation or the implementation, and the smallest answer that would not change the hub decision. No rewrites. No new names.

CFO, close risk. She will open with the missed close, not with cost. Dual-run, a named cutover owner, and a kill criterion if the first dual close fails stay inside the recommendation. They are not a reason to keep three centers.

Regional MD, headcount and service. Answering by leaving phantom roles in each region is a rewrite dressed up as phasing. Carry the objection: measure service levels, name the residual on-site roles you actually need, and do not keep a shadow finance team to protect a bonus metric. The bonus issue itself does not go in the client pack.

CHRO, consultation clock. The works-council timeline can slip the migration. That is sequencing, not location. Put consultation on the critical path. Do not move the hub to avoid a process you will have to run anyway.

The same pass invented a Chief Digital Officer who "will ask why this is not a platform play." Nobody in the interviews holds that title. Kill the line. It also offered a "phased regional autonomy model" that left transaction processing in each country. That is the model watering down the advice because the MD's objection was loud. Reject the rewrite. Changing the recommendation is a partner decision after the rehearsal, not a paragraph the tool slipped in. If the pack is silent on which ERP instance the hub posts into, the CIO will spend the slot on architecture. Fill that hole. Do not reopen the option comparison.

A finding earns a slot in the pre-read only if a partner can decide, in one sentence, what to do with it.

• Answer. The objection is fair, and the smallest fix is implementation design, evidence, or sequencing. Add the dual-run. Name the owner. Put consultation on the plan. Quote the figure the CFO will ask for, sourced from the implementation business case generator if the attack is really about payback rather than politics.
• Carry. The objection is real, often political, and answering it would change the recommendation into something you no longer believe. Walk in ready to hear it. Do not pre-compromise the slide.
• Discard. Invented stakeholders, generic catalogues ("change fatigue," "lack of vision"), and evaluator language copied from bid red-teams. If it cannot point to a named person you fed in, or to a constraint the client stated, it is color at best.

Cap the list. Ask for the objections most likely to derail the slot, mapped to people, not an exhaustive risk register. If the model starts producing replacement slides, stop it. A reviewer that rewrites becomes a co-author, and the rehearsal turns into an argument with the tool instead of with the room.

Write politics in a restricted note. Do not put them in the leave-behind. Bonus schemes, rivalries, and who lost last time are useful to the partner. They are poison in an appendix that travels.

Keep two artifacts. The meeting rehearsal is shareable with the core team: objections, who raises them, answer or carry, and any implementation hole to fill. The political note is sponsor-only or partner-only: no client logo on the file name, no forwarding, and never pasted into a prompt on an unmanaged ChatGPT or Claude account. Use the firm's approved tenant. If you cannot, do not put the politics in. Readiness belongs in the rehearsal when it talks about capacity and skills. Political feasibility belongs in the restricted note.

Run the pass while dual-run, consultation, and owners can still move. A night-before run produces anxiety, not a better pack. The pass is doing its job when the partner walks in knowing which three people will object, has answers for two, has decided to carry the third, and the recommendation on the pack is the same one the team believed the day before.