A premortem is the opposite of a postmortem. Instead of asking "what could go wrong?" — which tends to produce polite, hedged answers from any LLM — you tell the model the plan already failed at a future horizon and ask it to explain why. The frame switches the model into narrative mode and the failures get specific, honest, and useful.
Anything where the cost of being wrong is real: a customer-facing AI workflow about to ship, a pricing change, a hire, a policy rollout. If the worst that happens is "regenerate the output", a Stress Test is enough.
Steps
Open Constraint Studio
Go to /constraints and click "New constraint set". At the top, switch the mode toggle from "Rules-first" to "Premortem".
Fill the three context fields
What is it (one sentence). Who is it for / who does it affect. What does success look like in 6 months. The premortem is only as good as these three answers.
Run the premortem
We fast-forward to your horizon, assume it failed, and generate 4–9 specific failure scenarios — each with a story, a hidden assumption, and an early-warning signal you can actually observe.
Convert scenarios to constraints
Every scenario produces at least one action-verb constraint ("Never quote a price without checking the live table"). Accept, edit, or discard each one — accepted constraints save into your library as a reusable Context Card.
Use the hidden assumption
The synthesis surfaces the single biggest unquestioned belief behind your plan. That is usually where the real value of the premortem lives — read it last and read it twice.
The Stress Test critiques the wording of a prompt — tensions between blocks, missing context, places the model will guess. The Premortem assumes the prompt already shipped and the output caused a real problem, then asks why. Different frame, different output, different constraints.
The premortem method is model-agnostic. The same SKILL.md is published at /.well-known/agent-skills/premortem/SKILL.md so external agents (Claude, ChatGPT, Gemini, custom MCP clients) can load it and produce output Keypra's Constraint Studio renders identically.