Treat AI like a brilliant new hire with no memory and too much confidence. Give it context, demand receipts, verify before you trust, and write down what it learns. Every habit below is one of those four moves made concrete.
Two kinds of habit, two ways to teach them
Habits the labs already build
Name them. Don't re-teach them. The deep practice happens in the lab; the card is the quick reference participants keep.
The second chair
Reinforced · Lab 3 (recurs 5 & 6)Make the AI attack its own work before you do. Have it switch roles into a skeptic — board member, funder, reviewer — and try to break the draft. The catches are the point.
AI produces plausible and wrong with equal confidence. A brief that read as finished held 21 errors; the only reason they surfaced was asking the model to tear into its own work first. Plausible is the danger zone.
Participants meet this as "the second chair" in Lab 3, then again as a hostile funder in Lab 5 and a standing critique agent in Lab 6.The receipts rule
Reinforced · Labs 3 & 5Every claim gets a source line, and the AI names what it could not check. The second half is the part nobody else teaches — the honest list of gaps is where the risk hides.
The model sounds identical whether it knows or it's guessing. A claim-to-source line, plus a stated list of what it couldn't confirm, is the only way to see the difference before someone else does.
This is the source-accountability checklist saved into every Project in Lab 3, and "every figure traces to a source row" in Lab 5.Done means shown
Reinforced · every lab (explicit in Lab 4)Never accept "it's done." Ask to see it. The changed text, the actual number, the result of running it. A claim of completion is not evidence of completion.
A setup guide once "passed testing" and was still broken. A site "deployed successfully" while the live page showed the old content. Running without errors is not the same as being correct.
This is the verification ritual inside every lab — spot-check five rows, reconcile the totals, calculator-test the finding, most visibly in Lab 4.Make the correction stick
Reinforced · Labs 2 & 6When you fix the AI, fix it in the system, not just the chat. Write the correction into the Project's instructions so it applies every time — otherwise the same mistake returns next week.
Fixing a wrong answer in the moment fixes it once. Writing the rule into the model's standing instructions fixes it for good. A prompt is a thing you remember to do; a saved instruction is a thing that just happens.
This is the org brain built in Lab 2 and the prompt-becomes-a-standing-agent move in Lab 6.The genuine extras
None of these live in a lab, so they never compete for the session's focus. Drop one into an opener, a closer, or a 90-second aside.
The hype filter
Extra · recurring 3-minute openerTell the AI who you are first, then ask what's signal. Context turns a generic summary into a "here's the one thing worth your time" answer. Great as a weekly "AI noise vs. signal" warm-up.
Most AI content is mostly hype. Run the filter without context and you get a generic recap. Run it with context and you get a decision: worth your time, or not.
Options with consequences
Extra · 90-second aside, any labDon't ask for an answer. Ask for options, their costs, and a recommendation. This turns the AI from an oracle you have to trust into decision support you can steer.
A single answer hides the trade-offs and asks for blind trust. Options-plus-consequences put the judgment back in your hands, where it belongs, and it's a pure prompting habit anyone can adopt in week one.
The handoff note
Extra · fits the buddy-pair weekEnd a working session by having the AI write a note to whoever picks it up next. Solves the number-one complaint — "it forgot everything" — with zero tooling, in any chat.
AI starts every session blank. A handoff note carries the thread across the gap. It maps straight onto the buddy-pair model — pairs hand the work to each other between sessions, and to themselves next week.
The gap audit
Extra · prep habit, pairs with Lab 3 / 5Before a meeting, ask the AI what you don't know yet — not what you do. Surfacing the unanswered questions while there's still time to answer them is the whole value.
Summarizing what you already know is easy and low-value. Naming what you don't know, before the room, is where the leverage is. The "mark it open, never guess" clause is what keeps a gap from hardening into a false fact.