Read the same story twice, with two different models
A workflow that runs the same question through more than one provider and then has a third step look at where the two disagree.
The problem
One model's answer looks equally confident whether it is well grounded or not. Two models reading the same text will agree on what the text says and diverge exactly where the reading is doing work — which is the part worth checking.
The workflow
-
1
Show the Business Impact set per step
The first reading. Set this step's model explicitly to the provider you use most.
-
2
Tell Me Why This Is Dumb set per step, a different provider
The second reading, on a different provider, of the same source — an adversarial pass rather than a rephrasing of the first.
-
3
Look at both and say which claims survived. Where the two passes disagree is the output that matters.
What you put in
Anything where being wrong would be expensive: a funding decision, a vendor claim, a piece of analysis you are about to act on.
What you get back
The last step answers under these headings. They are read from Challenge the Story itself, so this list changes when the template does — it is not a transcript of a run.
- What claims are made and what evidence supports them?
- Are there logical fallacies or biases present?
- What important questions remain unanswered?
- What alternative perspectives should be considered?
Setting it up
- Open the workflow builder and start a new workflow.
- Add the three templates above as steps.
- On each of the first two steps, set the model explicitly rather than leaving it on the default — two steps on the same model is one reading run twice.
- Use provider keys from two different providers, each with its own spend cap.
- Save it, and keep it for the decisions that justify paying twice.
If it does not work
Both steps used the same model.
A step with no model of its own falls back to the workflow's. Set the model on each of the first two steps; if only one provider key is valid, only that provider's models are offered.
It costs twice as much.
It does — that is the trade. The cheap-model-first recipe is the other side of it, for the cases where the second opinion is not worth paying for.