The AI Gym — Practise, Reflect and Prove the Learning Belongs to You
In Part I, we established the operating principle: Copilot means co-, not auto-.
AI should not automatically decide, create, believe or become a surrogate for every uncomfortable human experience.
It should help you:
- practise more intelligently;
- obtain feedback faster;
- identify mistakes earlier;
- understand your patterns;
- challenge your certainty;
- and convert information into wiser action.
The difference between cognitive leverage and cognitive decay lies in how you train with it.
First, teach AI who it is working with
Most people give AI a generic request and complain when the answer is generic.
Before expecting useful personalisation, give it the relevant field:
- what you do;
- your level of expertise;
- your current project;
- the audience;
- your values;
- your constraints;
- your recurring strengths;
- your recurring mistakes;
- what you have already tried;
- what “excellent” looks like in your context;
- and how much capacity you have right now.
This does not mean uploading every private detail of your life. Use judgement and protect sensitive information.
It means refusing to ask: “What is the best routine?”
When the real question is: “What is the best routine for a neurodivergent founder who travels six times a month, is currently recovering from illness, needs two uninterrupted writing blocks weekly and abandons complicated systems after ten days?”
Discernment sits at the intersection of:
- wiring: how you tend to perceive, decide and respond;
- life: the conditions you are operating inside;
- load: what your system can access today;
- direction: where you are trying to go;
- limits: what must currently be respected;
- potential: what can be developed rather than permanently accepted.
The best answer for you six months ago may be wrong now.
AI should work with the person you are today.
Tell it to ask you questions
This is one of the simplest upgrades available.
Most people treat AI like a vending machine:
Insert request.
Receive output.
Instead, make it interview you.
Try: “Before answering, ask me the questions you need to reduce wrong assumptions.”
Or: “Do not recommend anything until you understand my current load, constraints and desired outcome.”
Or: “What information is missing that would materially change your answer?”
This forces the process away from generic completion and towards collaborative diagnosis.
It also reveals how much of your original request was built on unspoken assumptions.
Triple-check, reread and recalibrate
Telling AI to “triple-check” is useful as a prompt for additional scrutiny.
It is not a magical accuracy certificate.
A model can check the same faulty reasoning three times and become three times more eloquent about being wrong.
A better process is:
Check 1: Internal consistency
“Reread your answer. Identify contradictions, unsupported leaps and assumptions presented as facts.”
Check 2: External verification
“Which claims require external verification? Use primary or official sources where possible.”
Then open those sources.
Confirm that they exist.
Confirm that they support the stated claim.
Check 3: Context calibration
“Which parts of this answer may not fit my particular context, current load or goals?”
Check 4: Adversarial review
“Assume your recommendation is wrong. Construct the strongest case against it.”
Check 5: Update
“Based on the weaknesses identified, revise only what needs revision.”
The goal is not to make AI sound more certain.
It is to make the whole thinking process more calibrated.
Use AI as a role-player
One of AI’s greatest advantages is the availability of low-cost, repeatable simulation.
You can train it to play:
- a sceptical client;
- a resistant board member;
- a defensive employee;
- a highly technical interviewer;
- an anxious patient;
- a difficult negotiator;
- a blunt editor;
- a customer who does not understand the product;
- a teenager who responds with one word;
- or the person whose disappointment you chronically fear.
Feed it the relevant, anonymised context.
Tell it not to be agreeable.
Specify the difficulty.
Ask it to vary the scenario.
Then practise.
Emerging evidence suggests that simulation works best when it includes structured correction rather than repetition alone. In a randomised study of 94 novice counsellors, those who practised with an AI-simulated client and received feedback improved their use of reflections and questions. Those who practised with the simulation alone did not show the same gains and performed worse on one measure of empathy. The study was released as preprint evidence, so it should not be treated as a final verdict, but it captures an important principle: repetitions do not automatically improve you if you keep rehearsing the same mistake.
Have fun faster
I was speaking with a Cambridge professor about how he was using AI to learn a new digital instrument.
People asked him why he did not drag out the process and learn all the difficult, painful parts in the traditional way.
He said there had to be a smarter way.
“Have fun faster,” I said.
He laughed. “That’s exactly the way to put it.”
You do not become a better pianist by playing the entire piece badly from beginning to end fifty times.
You isolate the bar.
Slow it down.
Identify the incorrect movement.
Practise the scale.
Correct the fingering.
Repeat the right version.
Then return it to the song.
A meta-analysis of musical achievement found a substantial association between task-relevant practice and performance, while also making clear that practice is not the only determinant of expertise.
AI can accelerate this process.
Ask it to:
- identify the three recurring errors;
- create drills targeting each error;
- increase difficulty gradually;
- generate variations;
- give immediate feedback;
- test retention later;
- and place the corrected skill back inside the whole performance.
The goal is not to avoid every difficult moment.
It is to stop spending six months repeatedly ingraining the wrong movement before somebody finally corrects it.
The quicker you identify what is wrong and practise what is right, the sooner you can enjoy the song.
Train the mistake out. Keep the fingerprint.
When I use AI on my writing, I do not want it to turn me into the average of every competent article online.
I might ask it to identify:
- repetitive openings;
- excessive qualification;
- unclear structure;
- weak evidence;
- unsupported assumptions;
- missing transitions;
- or sentences that carry five ideas when they should carry one.
Those are trainable errors. But I do not want it to bleach away:
- humour;
- unusual cadence;
- cultural reference;
- strong opinion;
- emotional truth;
- productive eccentricity;
- or the idea it would never have originated.
Train out the mistake. Keep the fingerprint.
Generic polish is not the same as excellent writing.
Sometimes generic polish is merely the embalming of a living voice.
Use the Devil’s Advocate properly
Asking AI to “play Devil’s Advocate” is useful — but only if the opposition is genuinely strong.
Do not ask it to provide three token objections and then reassure you that your original position was excellent.
Try:
“Assume an intelligent, informed critic believes my argument is dangerous. Build their strongest case.”
Then:
“Which criticism exposes a genuine weakness, and which reflects a values disagreement?”
Then:
“What evidence would make me revise my position?”
This helps identify anti-patterns:
- situations where your usual strength becomes a liability;
- repeated assumptions that no longer fit;
- contexts in which your preferred strategy reliably fails;
- the point where confidence becomes rigidity;
- the point where empathy becomes enabling;
- the point where research becomes avoidance;
- the point where reflection becomes rumination.
The aim is not to destabilise every belief.
It is to know which beliefs can survive contact with an intelligent opponent.
Co-reflect without building an overthinking machine
AI can be an excellent reflection mirror.
After a meeting, performance or difficult conversation, give it:
- what happened;
- what you intended;
- what you actually said;
- what you noticed in your body;
- how the other person responded;
- what happened afterwards;
- and what you are currently telling yourself it means.
Then ask:
- “Separate observation from interpretation.”
- “Which facts am I treating as proof of a story?”
- “Give me two plausible explanations that contradict mine.”
- “What was within my control?”
- “What should I practise before the next attempt?”
- “What is the smallest useful action now?”
But place a stopping rule around reflection.
AI has infinite conversational runway.
Your nervous system does not.
If you keep asking for a twenty-sixth interpretation, you are no longer co-reflecting.
You are using computation to industrialise rumination.
A useful protocol is:
- Extract the lesson.
- Choose one action.
- Name what remains uncertain.
- Stop.
- Return to life.
Wisdom requires space after analysis.
Once in a while, do it manually
Write the piece without AI.
Draft the message.
Recall the main points before opening your notes.
Navigate without GPS.
Make the estimate.
Hold the conversation without a generated script.
Work through the problem before requesting the solution.
This is not because manual effort is morally purer.
It is because unassisted performance tells you what has actually transferred.
Call it a cognitive fire drill.
If the skill disappears when the tool closes, you may have rented performance rather than built capability.
A useful sequence is:
- Attempt independently.
- Use AI for feedback or extension.
- Study the difference.
- Repeat the improved version.
- Test yourself without AI later.
AI assistance can then become scaffolding that is gradually removed rather than a permanent exoskeleton you are terrified to lose.
Use GoogleDocs as a thinking laboratory
Do not let AI silently replace your work.
Preserve the layers.
Layer 1: Your original
Write what you think before AI touches it.
Layer 2: AI’s suggestions
Paste them separately or use comments and tracked changes.
Layer 3: Your reviewed version
Highlight:
- green: genuinely improves the work;
- yellow: useful, but must be adapted;
- red: incorrect, generic or not-you;
- comment: why you accepted or rejected the suggestion.
Then review periodically:
- Which mistakes recur?
- Which corrections have I internalised?
- Where is AI consistently better than I am?
- Where does it flatten my voice?
- Which advice do I accept merely because it sounds polished?
- Can I now produce the improved version unaided?
This converts editing into metacognition.
You do not simply improve the document.
You learn how you think.
AI is becoming the new wellness-and-innovation programme
Organisations are at risk of repeating an old pattern.
Buy the tool.
Run one generic webinar.
Announce that the company is innovative.
Leave everyone alone with it.
Measure logins.
Blame employees when the outcomes are inconsistent.
This is the AI equivalent of installing a meditation app while preserving seven back-to-back meetings.
Real AI capability requires:
- use-case mapping;
- an understanding of the jagged frontier;
- privacy and confidentiality rules;
- source standards;
- role-specific drills;
- permission to report when AI makes work worse;
- human review at meaningful risk points;
- manual baselines;
- and time to reflect on what is being learned versus merely produced.
Otherwise, AI becomes a new source of work slop:
more output;
more checking;
more correction;
more polished material nobody needed.
High performers especially do not need a generic list of fifty prompts.
They need to know:
- where AI multiplies their real strengths;
- where it conceals a skill gap;
- where it saves high-value time;
- where it creates rework;
- and where their judgement must remain firmly in command.
The final test: are you becoming wiser?
AI can make you faster.
It can help you remember more, practise more, see more alternatives and access personalised explanations.
But wisdom asks different questions:
- What deserves my attention?
- What should remain human?
- What am I avoiding by automating this?
- Is this answer true, useful and appropriate now?
- What will this choice compound?
- Does this strengthen my relationship with reality—or merely make my preferred story more articulate?
- Am I using AI to become more capable of life, or to avoid the friction of living it?
You are right to worry about AI.
But fear without experimentation incinerates mental energy and teaches you nothing.
AI will make some people more cognitively passive.
It will make others faster, better rehearsed, more reflective and more capable of seeing beyond their first interpretation.
The difference will not simply be intelligence at baseline.
It will be how they practise.
Do not ask AI to spare you from thinking. Ask it to give you better thinking to do.
Use it to have fun faster.
Practise more often.
Expose mistakes earlier.
Ask better questions.
Challenge your own certainty.
Then close the laptop occasionally — and prove that the learning belongs to you.
Want to know more? Book a Performance Call with us to find out how we can help your organisation.
Book your Performance Call.
Register for our free monthly 20:20 webinar
Clearer views on the neuroscience of high performance.