AI Fluency.

The framework

1. 01

   One page, one hour, once a quarter.

   Monthly is too often for this. Your tooling moves weekly, but your task map and your scarce assets don't shift in four weeks, so run it on its own quarterly clock, outside the monthly operating review you already run. The page has three movements and they go in this order: calibrate, which is what the machine can and can't do on your work; direct, which is how you brief it, ground it, argue with it and check it; rebuild, which is what you've redrawn around it and what that has cost. Fixed order, same page every time, because the value sits in comparability and not in any single quarter's figures. Read the whole page before acting on any line of it, because the reflex this module set out to correct is the one that jumps at the first alarming number. If a reader can't get through it in ten minutes, you've built a dashboard, and dashboards get admired instead of used.

2. 02

   The first third: calibrate — the task map, the tests, the tax.

   Start where the module started. Day 31's leverage audit and your task map are one list now, so re-sort it: which recurring tasks have moved from own to augment, which have moved inside the frontier since last quarter, which sit stubbornly outside it, and which you've never tested but have quietly assumed. Say where the reclaimed hours went, in a sentence. Then the counts: tests you ran, not tools you tried; load-bearing asks you sent twice in separate sessions and what the two runs disagreed about; entries in the wrong-answer log and the shape of failure they add up to. Then the harder line: the judgements you promised to keep doing by hand, each marked held or lapsed. A skill you stop practising leaves without announcing itself, and this is the only line on the page that catches it before a client does. Then the other side of that line: which steps a person must still sign before the work moves, which of those checks you let go this quarter, and what has caught the error since.

3. 03

   Direct is eight lines, and each has a right answer.

   Is the brief written down — context, goal, constraints, example — or reassembled from memory by whoever's prompting that morning? Are your standards recorded as acceptance criteria and failure conditions, or does everyone carry a slightly different version of done? What are you grounding answers against, how old is the oldest document in that library, and how many answers went out with no source document named at all? How many decisions did you put twice, from opposite stated positions, and what changed because of the one argument you couldn't answer? Is the judgement you wrote a rule for still being applied as written, and how many exceptions reached the record? What did your held-out ten score this quarter, against which model and which version of the prompt, and is that better or merely different? How many claims left your desk on the machine's word alone, and what did the worst of them cost you? Then the register: every permission an agent holds to act rather than suggest, who granted it, what its most damaging reversible action would be, and what has changed since.

4. 04

   End in changes to the system, then file the page.

   Deming's red bead demonstration made the point by rigging it: willing workers put in front of a process that guaranteed defects, then exhorted to try harder. Above the changes, record what you actually redrew rather than merely sped up, and the verdict of any bake-off you ran, with the date you will put the same cases through it again. Then write your changes as changes to the system — a standard added, a check moved earlier, a permission narrowed, a case added to the held-out set — each with one owner and one date, and cap the list at three. Finish with the two economic lines: what the tooling cost set against what it produced, and one sentence on the scarce asset you funded. Then file the page, because next quarter's honesty depends on this quarter's version still existing.