From tasks to thinking
The Meridian exec team can now write a proper brief - Dana gets her themed summaries, the CMO has her campaign template. But the highest-value use of AI for a senior leader is rarely another task. It is a thinking partner that never tires, never worries about your reaction, and holds more patterns than any advisor - if you know how to stop it from simply telling you that you are right. Dana has a real decision on the table this month, and we will use it.
Beyond tasks: thinking together 8 min live
Andrew Ng's non-technical prompting course devotes a whole module to this shift: from AI as a doer of tasks to AI as a partner in reasoning. Three moves cover most of what a leader needs - and each one has a trick that separates a useful session from a pleasant waste of ten minutes.
LiveMove 1 · the wide brainstorm - ask for 20, not 33 min▶
Ask for "some ideas" and you get the five most obvious ones - the same five your team already said in the meeting. The fix is volume plus variety rules:
- Demand quantity. "Give me 20 options" forces it past the obvious. Ideas 1-5 are the meeting; ideas 12-20 are where the interesting ones hide.
- Ask for the unusual explicitly. "Include at least 5 that a conventional retailer would never consider." It will not volunteer weird - you have to license it.
- Force distinct angles. "No two options from the same category" or "vary who pays, who benefits, and what we would stop doing." Variety rules prevent 20 flavours of the same idea.
The CMO's loyalty brainstorm. Asked for "ideas to improve our loyalty programme" she got points, tiers, and birthday vouchers - a 2015 conference talk. Asked for "20 options, at least 5 unconventional, no two from the same category, and for each say what we would stop doing to fund it" she got two ideas that made the next exec agenda. Same tool, ten minutes apart.
LiveMove 2 · the tough reviewer - point it at your own draft3 min▶
Most people use AI to write drafts. The more senior you are, the more value runs the other way: you write the draft, and AI critiques it before a human ever sees it. The trick is casting - do not ask "is this good?", assign a hostile role:
- "You are a skeptical board member who has seen ten transformation plans fail. Find the 3 weakest claims in this paper and say why each would make you vote no."
- "You are our toughest competitor's strategy head. Where would you attack this plan?"
- "Read this as the store managers will read it. What lands badly, and what question will every one of them ask that I have not answered?"
A tough review costs you nothing but pride, arrives in a minute, and no career is harmed in the making. The board sees draft four, not draft one.
Self-studyMove 3 · the premortem - failure, explained in advance3 min read▶
The premortem is a decision-science classic: instead of asking "what could go wrong?" (which invites polite shrugs), you declare the failure already happened and ask for the story. "It is 12 months later and this initiative failed - write the post-mortem." The tense change unlocks specifics, because explaining a fact is easier than imagining a risk.
AI is unreasonably good at premortems. It has absorbed thousands of retail case studies, failed rollouts, and post-mortems, and it has no stake in your project being a good idea. It will name the awkward failure modes - the ones involving your own org chart - that nobody says in the room. The live exercise today is exactly this, on a real decision. You will see the template in the workshop section.
The sycophancy trap 7 min live
Here is the uncomfortable fact behind all three moves: by default, models tend to agree with you. They are trained on human feedback, and humans reward agreeable answers. Show it your plan and it will find the good in it - every time, regardless of the plan. For a junior analyst that is annoying. For an executive whose staff already hesitate to disagree, it is a second yes-man, at scale.
LiveEscaping the loop - the sentences that buy honesty4 min▶
You cannot make a model less agreeable by wishing. You make it honest by making disagreement the assignment:
| Instead of... | Say... |
|---|---|
| "What do you think of my plan?" | "Argue against this plan. Steelman the case for NOT doing it." |
| "Is this analysis right?" | "What would a skeptical CFO say about this analysis? Be specific about the numbers." |
| "Any feedback on the draft?" | "Give me the 3 weakest points and the strongest objection I have not answered. No compliments." |
| "Do you agree?" | "Under what conditions would this be wrong? What evidence would change the recommendation?" |
Self-studyWhy models flatter - and why it is partly your fault2 min read▶
Sycophancy is not a bug someone forgot to fix. Models are tuned on human preferences, and humans consistently rate agreeable, encouraging answers higher - so agreement gets trained in. Worse, your phrasing leaks your preference: "I'm thinking we should go ahead with the pilot - thoughts?" tells the model which answer will please you, and it obliges. Ng's course flags this as THE trap for non-technical users precisely because senior people are the least likely to get pushback from humans either.
Two habits blunt it. First, hide your position when you want a clean read: present the options as A and B without revealing your favourite. Second, make pushback structural: end big asks with "list the strongest arguments against your own recommendation." The model does both happily - it just will not start.
Knowing vs searching vs reasoning 7 min live
A thought partner is only as good as what it is thinking WITH. Modern AI chat apps have four distinct gears, and executives who mix them up get burned: trained knowledge with a cutoff date, live web search, deep research that reads dozens of sources for you, and reasoning models that plan before answering. Knowing which gear you are in is a 30-second check that prevents most embarrassments.
LiveWhich gear, when - the reach-for table4 min▶
| You need... | Reach for | Watch out |
|---|---|---|
| Concepts, frameworks, drafting, timeless how-to | Trained knowledge | Anything dated may be stale - it cannot know past its cutoff, and it may not warn you |
| Today's facts: prices, people in roles, this week's news | Web search | Skim the sources it cites; search grounds the answer, it does not guarantee the reading |
| A market scan, vendor comparison, competitor landscape | Deep research | Takes minutes, not seconds; check the source list before you trust the synthesis |
| A hard trade-off, a plan with many moving parts | Reasoning model | State the goal and the bar, not the method - micro-managing its steps makes it worse |
Dana's gear check. Planning the fulfilment decision, Dana used all four in one afternoon: trained knowledge to explain ship-from-store models, web search for what competitors announced this quarter, deep research for a cited report on rollout economics, and a reasoning model for the staffing trade-off across 20 stores. Four gears, one decision, no confusion about which answer came from where.
Self-studyThe reasoning-model habit, one level deeper2 min read▶
Session 1 planted the habit: with reasoning models, state the goal and the bar, do not dictate method. Here is the why. These models are trained to plan internally - they draft an approach, check it, and revise before you see a word. Prescribing your own step-by-step method overrides a planner that is usually better at planning than your prompt is. Every major vendor now says the same thing in their guidance.
What replaces the method? The bar. "A good answer will compare at least three options, quantify the staffing impact, and tell me what would have to be true for the cheapest option to win." You define done; it finds the route. If that sounds like how you delegate to your best director, that is the point - and the full technical story lives in Builder session 4.
The premortem: one real decision, stress-tested ★ 15 min · work it live
You brought one real decision. Dana brought hers: whether Meridian should launch ship-from-store fulfilment in 20 stores this autumn. We run her premortem together, then you run yours. This is the tough-reviewer move and the sycophancy antidote in one exercise.
Pick a live decision. Real and undecided - a pilot, a hire, a vendor, a reorg. The premortem only bites when you still have time to act on what it finds.
Write the one-paragraph brief. What you are deciding, the scale, the timeline, what success looks like. Session 2 discipline: the premortem is only as sharp as the context you give it.
Run the premortem prompt. Use the template above. Read the five findings and mark the one that made you wince - the wince is data.
Add the skeptical CFO pass. Follow up with: "Now, what would a skeptical CFO say about the financial case? Attack the numbers specifically." Different role, different failure modes - roles are lenses.
Harvest one action. Pick the most plausible early warning sign and give it an owner and a date this week. A premortem that changes nothing was a séance, not a meeting.
Three things before session 4 ◐ 25 min total
- Run one premortem on a real decision using the template - and harvest one owned action from it. Bring the finding that surprised you most.
- Catch the flattery once. Show AI a plan and ask "thoughts?" - then in a fresh chat, ask it to argue against the same plan. Compare the two answers side by side. That gap is the sycophancy tax you have been paying.
- Try one deep research task - a market scan or competitor question you actually care about. Before reading the report, read its source list. That habit is the bridge to session 4: judging what comes back.
What this session covers
Leader-level coverage of the sources below - working ideas in plain language, honestly flagged where the depth lives in the builder track. Official certificates and videos stay with the vendors; links are in the course map.
Three questions before you go 🎯 ◐ 90 seconds
1 · Every time you share a plan, the AI tells you it is strong and well-considered. The best fix is to...
Models are tuned toward agreeable answers, and your phrasing leaks your preference. Ratings get inflated and model-switching does not remove the training. Making disagreement the assignment does.
2 · You need a competitor's current market position and this quarter's announcements. You should...
Trained knowledge is frozen at a cutoff date, and no phrasing can update it. Current facts need the search or research gear - and the sources are there to be opened, not admired.
3 · You hand a reasoning model a hard staffing trade-off. The best brief...
Reasoning models plan internally, and every major vendor now warns that dictating the method degrades them. You define done - the bar - and it finds the route. Delegation, again.