In short: Every interaction a judgment system records should serve an explicit outcome. But naming the conclusion in advance bends the reasoning to reach it. The craft is keeping the purpose fixed and the answer open.
Key points
- "Decide whether to renew this account" is intent; "find reasons to expand this account" is a destination.
- Language models construct plausible paths to wherever they are pointed, underweighting contrary evidence.
- Ask analytical questions in a fixed schema: signal, linked evidence, confidence, implication, what would change this.
- When advocacy is needed, make it a separate step and run the case against at the same time.
- Tag each recorded interaction with the decision it serves; untagged interactions belong elsewhere.
There is a tension at the heart of purposeful AI design, and most teams resolve it in the wrong direction.
On one side, interactions without a stated business outcome drift into busywork. Summaries pile up that serve no decision, and the record fills with noise. Intent is what makes a record worth keeping.
On the other side, a prompt that asks for an opportunity, a recommendation or a win points the model at a conclusion. Language models are very good at constructing a plausible path to wherever they are pointed. The reasoning quietly reshapes itself to arrive, and the evidence that would have argued otherwise is underweighted or left out.
Separate the outcome you serve from the answer you expect
The resolution is a distinction that is easy to state and surprisingly easy to lose in practice.
- "Decide whether to renew this account" is intent. "Find reasons to expand this account" is a destination.
- "Assess this counterparty's exposure to the sector downturn" is intent. "Identify the risks in this counterparty" presumes there are material ones.
- "Should we enter this market this year?" is intent. "Build the case for entering this market" is a destination.
The second version in each pair produces a more confident, more coherent, and less trustworthy answer.
Ask as a schema, not as a request
The practical tool is to ask every commercial or analytical question in a fixed shape rather than as an open request:
signal · evidence (linked) · confidence · implication · what would change this
The schema keeps the purpose in view because every field relates to the decision. It leaves the conclusion open because no field asks for one. The last field does the most work. Being required to say what would change the conclusion makes it much harder to produce a one-sided case without noticing.
When you genuinely want the case for something
Sometimes you do want advocacy: a pitch, a board paper, a response to a sceptic. Fine. Make it an explicit, separate step, and run the case against in the same schema at the same time. Record both. The decision record then shows that advocacy happened, which is very different from advocacy dressed as analysis.
Not every interaction needs to be recorded
Intent also tells you what to leave out. An interaction that serves no current decision does not belong in the judgment record. It may belong in a general knowledge store, but mixing the two dilutes precedent with noise. A lean record that maps cleanly to decisions is more useful than a complete one that maps to nothing.
In practice
- Audit your ten most-used prompts. Rewrite any that name a conclusion as one that names a decision.
- Adopt the five-field schema for analytical questions, with "what would change this" required.
- Make advocacy an explicit step, always paired with the case against.
- Tag every recorded interaction with the decision it serves. Untagged interactions go elsewhere.
What would prove this wrong. A blind comparison in which destination-framed prompts produce outputs no more one-sided, and no less accurate, than schema-framed ones on the same cases. If framing does not bend the reasoning, the discipline is ceremony.
Common questions
Why do AI prompts that ask for a recommendation give biased answers?
A prompt that asks for an opportunity or a win points the model at a conclusion, and the reasoning reshapes itself to arrive there while contrary evidence is underweighted or left out.
How should I frame analytical questions for an AI model?
Use a fixed schema: signal, evidence (linked), confidence, implication, and what would change this. It keeps the purpose in view without asking for a conclusion.
What if I actually want the AI to argue for something?
Make advocacy an explicit, separate step and run the case against in the same schema at the same time. Record both.