AI Fluency.

The framework

1. 01

   Your question already contains the answer you're hoping for.

   Read your last prompt back as a stranger would. "Here is my plan, any thoughts?" announces ownership. "Is it fair to say the market is moving to X?" announces the conclusion. Even the framing carries it, because describing a decision you've already half-made in generous language tells the model exactly which way to lean. Write the question so that nothing in it reveals the answer you want: describe the plan as a competitor's, or as a stranger's submission you have been asked to assess. The rewrite is the whole intervention. Everything that comes back afterwards is downstream of what you gave away in the first line.

2. 02

   Commission the case against — by name, and with a job attached.

   "Any concerns?" invites a paragraph of balanced-sounding hedges. Give the model a role and a brief instead: the analyst hired by a board that wants this killed, the competitor who'd be delighted to see you commit, the auditor who has watched this fail four times. Ask for the three strongest arguments against, ranked, each with the evidence it would need to be true. This isn't the red teaming of Day 72, which puts real colleagues against a real plan and takes a week. It's a two-minute habit you attach to every ask, and it is the cheapest way there is to make the machine argue with you.

3. 03

   Ask it twice, from opposite sides.

   Run the same question two ways: once having said you favour the acquisition, once having said you oppose it. If the substance moves with your stated position, you've learned nothing about the acquisition and something significant about the instrument. Keep the two answers side by side rather than merging them, because the useful material is precisely what appears in one and not the other. Do this on any judgement expensive enough to be worth a fortnight of your best people, and do it before you have told anyone which way you are leaning, including the model and including yourself.

4. 04

   A machine can produce dissent, but it cannot insist on it.

   It will write the case against, and it will drop that case the moment you explain why you disagree: no cost, no relationship, no reputation on the line. Real dissent carries all three, which is what makes it stick. So use the machine to guarantee the objection has at least been articulated, then take the serious ones to a person who's willing to lose an argument with you and stay. Day 58 dealt with your own confirmation bias. This is that same failure, now with a fluent and infinitely patient accomplice.