Pre-Mortem a Plan With Multiple AIs: Find What Breaks Before You Commit

What is an AI pre-mortem?

A pre-mortem is a simple planning move: picture that your plan has already failed, then work backward to list every reason it could have gone wrong. It flips the usual optimism of planning into deliberate doubt, which is when weak points finally get named out loud. Running it across several AI models at once widens that list, because each model brings different training and different blind spots.

The routine has four steps: frame the plan, gather failure modes in parallel, rank them, then turn the top risks into fixes. You can do the whole thing in one sitting inside aiDex, moving from one mode to the next without leaving the chat. It pairs naturally with the rest of your multi-model workflow: decide first, then stress-test the decision.

Why run it across several models instead of one?

Because one model has one point of view, and a pre-mortem lives or dies on breadth. Ask a single model "what could go wrong" and you get a competent but narrow list, shaped by how that model was trained. Ask three models the same question and you get overlap on the obvious risks plus a long tail of items only one of them thought of. That tail is the whole reason to do this.

Different providers genuinely disagree about what is risky, and here the disagreement is a feature, not noise. Compare mode puts the answers side by side so you can see where they agree (probably real) and where only one flags something (worth a look). Use your own provider keys or the ones we manage, and pick the models you want. That way the panel reflects the models you actually trust.

How do I run a pre-mortem in aiDex, step by step?

Four steps, one per mode:

1. Frame it (Solo). Write the plan in a few lines, then add the pre-mortem prompt: "It is six months from now and this plan failed. List what went wrong and why." Attach the full plan as a document (DOCX, PDF, MD, or txt) so every model in the next step reads the same source, not your summary of it.
2. Gather failure modes (Compare). Send that prompt to several models at once and let each answer independently. Because they run in parallel, no model anchors on another's list, which keeps the failure modes diverse.
3. Rank the risks (Judge). Hand the combined lists to a Judge model. Ask it to merge duplicates, group related risks, rank each by likelihood and impact, and flag any risk that only one model raised. You get one ordered list instead of three overlapping ones.
4. Turn risks into fixes (Pipeline). Run the top risks through a Pipeline: Draft a mitigation for each, Critique it, Revise, and Polish. You leave with owner-ready fixes, not just a wall of worries.

For a plan still being argued over, open a Team chat and let the models react to each other in real time. For anything confidential, point the panel at local models with Ollama so the plan never leaves your machine.

What does the ranked risk list look like?

Here is the shape of the output (illustrative, not real model results):

The value is not the table itself, it is that three models argued their way to it. When two of them independently flag the same risk, you treat it as real; when only one does, you decide whether it knows something the others missed.

When should I run a pre-mortem, and when should I skip it?

Run one before any decision that is expensive or hard to reverse: a launch, a migration, a key hire, a signed contract. Those are the plans where an unspoken risk costs the most. Skip it for small, reversible calls, where the exercise costs more attention than the decision is worth.

A pre-mortem also works best downstream of a choice. Use a decision matrix to pick among options first, then pre-mortem the winner. If the plan is a strategy rather than a single choice, a strategy table sets it up and the pre-mortem pressure-tests it. To sanity-check which risks are real, you can even run a quick consensus pass over the ranked list.

Keep it honest

A multi-model pre-mortem widens your view; it does not guarantee you caught everything. Models can miss risks that are specific to your domain, your customers, or something genuinely new, and they can agree with each other and still be wrong. Treat the ranked list as a strong first draft, then hand the top items to people who know the territory. The point is to walk into the decision with your eyes open, not to outsource the judgment.