You do not get better at AI by reading prompt templates. You get better by feeling the difference between a weak move and a strong one, on your own real work.
Every drill below is built to hand you one of those moments. Bring a real task, run one drill, ten minutes. The line at the bottom of each card tells you where the drill lives in your actual week.
Run one while the kettle boils. Each takes a single chat and something real from today. These drills train the habits the General Instructions template installs: get the template, then come here to feel it work.
Set a sixty second timer. One prompt, one shot: get a usable answer to something you actually need today. No retries.
What you feelThe clock strips out the throat clearing. What survives is the part your prompt actually needed, and it is almost always context, not politeness.
Real-work version: before you type a quick answer to a colleague, draft it with one timed prompt first and see which version is better.
Take a real situation you are facing. Ask for the five worst possible ways to handle it. Then flip it: now the five best.
What you feelThe bad answers map the edges of the problem. When the good ones arrive you can actually judge them, because you have seen the floor.
Real-work version: before a tricky email or a hard conversation, ask for the three worst ways to play it. Your real move gets sharper.
Ask for five facts about your niche. Hunt for the shakiest one and demand sources for it.
What you feelThe confident sentence and the true sentence read exactly the same. Your trust dial moves from “sounds right” to “show me”.
Real-work version: any AI fact you are about to repeat in a meeting or a document gets the same treatment. Source first, repeat second.
Bring a real task. These drills change how answers arrive, and what you do with them when they do.
Make the same request twice, in two chats. Once cold. Once with three lines first: who you are, what this is for, what good looks like.
What you feelSame model, minutes apart, two different classes of answer. Context stops being advice you have heard and becomes fuel you can pour.
Real-work version: keep a three line brief about you and your work in a note, and paste it at the top of any chat that matters.
Give it a fuzzy task and add one line: “ask me five questions before answering.”
What you feelIts questions expose what you had not decided yet. Half the value arrives before it writes a single word of the answer.
Real-work version: any brief you cannot write cleanly, open with “interview me first”.
Interrogate the AI on the one topic where you are the expert. Push into the fine detail until it slips.
What you feelYou find the exact boundary where fluent runs ahead of true. Now you know what a slip sounds like on topics where you could not have caught it.
Real-work version: carry the scepticism you earned on your strong subject into every answer on your weak ones.
One question, four constraint sets: explain it to a ten year old, in a tweet, as a table, to an investor.
What you feelThe substance barely changes; the framing does everything. The shape of the output was yours to set all along.
Real-work version: stop accepting the default shape. End real prompts with the form you want: as a table, in five bullets, one paragraph for the board.
Have it draft a real email you need to send. Before you may use it, you must push back three times. “Looks good” is banned.
What you feelThe third pushback is the hardest to find and lands the biggest improvement. Somewhere in the middle you stop accepting and start directing.
Real-work version: nothing drafted for you goes out under your name without at least one genuine pushback.
List five real tasks off your plate. Sort each one: a quick conversation, a written brief, or a full hand-off.
What you feelThe sorting is the skill. Most people run everything in one mode; you feel where each task actually belongs.
Real-work version: run the sort on your week every Monday. The tasks marked “brief” are usually the biggest wins hiding in plain sight.
Write down five things AI should always know about you, and three things it should never carry from one piece of work to the next.
What you feelContext turns into something you own and manage, not something that just happens to a chat.
Real-work version: the five become your standing instructions or project file. The three become your rule for when to open a fresh chat.
Give it two samples of your writing and have it write your next email in your voice. Name exactly what is off, then correct it.
What you feelYou cannot fix “sounds off” until you can say what is off. Naming your own voice is the actual drill.
Real-work version: keep the corrections. Three named voice rules turn every future draft from a forgery into a first draft.
Take one real task and fly it in legs: one instruction per message, and land each step before the next takes off.
What you feelOn anything with moving parts, short legs with checkpoints beat one giant prompt. The control comes back to your side of the desk.
Real-work version: fly your next report, deck or plan in legs, and clear each landing before the next take-off.
These change how you decide, not just how you ask.
Take a position you genuinely hold. Tell the AI to argue the other side, and tell it not to fold when you push back.
What you feelWith the agreeableness switched off for one chat, you discover which of your arguments survive contact and which were only ever unopposed.
Real-work version: before a real decision goes in front of other people, give it one round against a partner told not to fold.
Paste a finished answer into a fresh chat and say: “tear this apart.”
What you feelThe fresh chat has no stake in the answer and no memory of writing it. The same model that nodded along in one window dismantles the work in another.
Real-work version: anything important gets a second opinion from a chat that did not write it.
Same question, two chats, two framings: one leaning towards yes, one leaning towards no. You deliver the verdict.
What you feelYou watch the framing steer the answer in real time, and realise your own everyday questions were never neutral either.
Real-work version: for any big call, deliberately ask it both ways before you trust either answer.
Teach the AI a concept from your work. Then have it explain the concept back as if briefing a new hire. Grade the echo.
What you feelEvery gap in its echo is a gap in your explanation. Teaching it turns out to be a mirror.
Real-work version: before you delegate anything, to AI or to a person, run the echo test: explain, hear it back, grade it.
Take a finished AI draft. Ten seconds per section, verdict out loud: ship it, one line fix, or rethink. No lingering.
What you feelSpeed forces judgement. You already know which sections are weak; the drill removes the option of not deciding.
Real-work version: make this your standard review pass on any draft: verdict first, polish after.
Say: “I have opinions about this that I have not put into words yet. Interview me and pull them out of me.” Then just talk.
What you feelThe ideas were yours all along. The questions do the extracting, and the AI stops being an answer machine and becomes a midwife.
Real-work version: next time you need to write something up, get interviewed instead of staring at a blank page.
With your second brain loaded, ask: “based on everything you know about me, guess my new business idea.” Fair warning: this one only works if you have built a second brain. With no stored context there is nothing to read.
What you feelIf the guess lands close, your context is compounding. If it is miles off, you have just found exactly what your files do not yet say.
Real-work version: run it every quarter. The gap between its guess and your reality is your context backlog, and closing it is the whole game.
One drill, one real task, ten minutes, this week. The drills only work on real work: the moment they teach does not happen on example data. When one lands, write down what changed. That note is worth more than any template.
The moves behind the drills are taught in How to prompt → The tools you graduate to live in the Skill Toolbox → And the graduation drill needs a second brain →
© 2026 McCloskey.ai · Back to McCloskey.ai