AI Guides › Playbooks
By Nigel Guy · 7 min read
You watch a tutorial, nod along, copy the prompt, get a good result, and feel you have learned something. Then a week later you open a blank chat and cannot rebuild any of it. Smooth practice is pleasant because the work is being done by the tutorial, or by the model, rather than by you. That ease is exactly why so little sticks.
The rule: practise one narrow skill, and make each session force you to produce, recall, vary or explain something before you are allowed to look at the answer.
Cognitive psychology has studied this for decades, and four findings are well documented. None is specific to AI; they apply to any skill you want to keep.
| Habit | Finding it rests on | What it means for AI practice |
|---|---|---|
| Attempt before you look | The generation effect: Slamecka and Graf (1978) found that words people generated themselves were remembered better than the same words simply read. | Write your own prompt before reading anyone else's. |
| Rebuild from memory | The testing effect: Roediger and Karpicke (2006) found that after a delay, students who had been tested on a passage retained more than students who re-studied it. Restudying won only on a test five minutes later. | Close everything and recreate last week's best prompt cold. |
| Space the sessions | Spacing: Cepeda and colleagues (2006) reviewed 839 comparisons across 317 experiments and found the best gap between sessions grows with how long you need to remember. | Several short sessions over weeks beat one long weekend. |
| Vary the conditions | Varied practice, one of the "desirable difficulties" described by Robert and Elizabeth Bjork: it feels harder and helps you apply a skill in new situations. | Same skill, different audience, length and constraint each time. |
The research on any one of these has caveats, and the studies mostly used word lists and prose passages, not prompts. Treat them as a sound reason to design practice this way, not as a guarantee of a particular result.
The ledger is a single plain file named reps-log, paired with a four-week timetable. Sessions run for roughly 20 to 30 minutes, and each closes by adding three short lines to the file:
Before the first session, choose the one skill the whole month is about. Name it in a short phrase of no more than four words, something concrete enough that you could tell a friend what you were practising. A tight example is "drafting outreach to strangers", "interpreting campaign numbers" or "building a single assistant". Labels such as "AI in general" or "prompting" are far too wide. If the skill will not narrow down, you will drift between topics each week, and that is just the tutorial trap again.
Each day, take one real task. Write your prompt with no searching for ten minutes, run it, and save the output. Only then ask for a comparison.
You are an expert in [MY_SKILL] and a patient coach.
Below are a prompt I wrote and the output it produced.
PROMPT: [MY_PROMPT]
OUTPUT: [PASTE_OUTPUT]
Goal: show me how you would have prompted for this task, so I can see the gap.
Steps:
1. Write your own version of the prompt for the same task.
2. List each meaningful difference between your version and mine.
3. For each difference, say what it changes in the output.
4. Name the one difference that matters most.
Constraints: do not praise my prompt unless the praise is specific. If the task description is too thin to judge, ask me one question first rather than guessing.
Before answering, check that every difference you list is one you can actually justify.
Fill in your skill, your prompt and the output.
Nothing open: no tabs, no notes. In a blank chat, reconstruct last week's best prompt. Compare afterwards, never before.
You are a coach reviewing a prompt I rebuilt from memory.
ORIGINAL (from my notes): [ORIGINAL_PROMPT]
FROM MEMORY: [MY_REBUILD]
Goal: find out what I forgot and whether it mattered.
Steps:
1. List every element in the original that is missing or changed in my rebuild.
2. For each, judge: did the omission change the output quality, or was it cosmetic?
3. Tell me which two elements I should be able to recall without help.
Output as a table: element, kept or lost, impact (high, low, none), reason.
If you cannot judge impact without running the prompt, say so instead of guessing.
Twice this week, skip the rebuild and do recall only:
Act as a quizmaster on [MY_SKILL]. Ask one question at a time, starting easy and getting harder. Wait for my answer. Before moving to the next question, tell me what I got wrong and why. After eight questions, summarise my weak spots in a short list. Do not reveal answers in advance.
Use the same skill in a different setting every day: another audience, another length, another constraint. Midweek, remove a crutch, such as half the words or no examples. Once, run a stress test:
You are a sceptical reviewer. My current approach to [MY_SKILL] is: [DESCRIBE_APPROACH].
Give me five realistic situations where this approach would fail, ordered from least to most difficult. For each, say what would go wrong and why. Then set me the hardest one as an exercise and wait for my attempt before commenting.
Keep the situations realistic for someone doing [MY_CONTEXT]. If my description of the approach is too vague to test, ask me to expand it first.
Days one to three: apply the skill to something with a real consequence, such as a client email actually sent or a report someone will read. Day four: teach it to another person, aloud or in writing. I have not cited a named study for this step; it is simply a cheap way to expose gaps in your own understanding. Day five: review the whole log.
You are an honest performance reviewer. I have practised [MY_SKILL] for four weeks. My practice log (around 20 sessions) is below.
LOG: [PASTE_REPS_LOG]
Goal: tell me what actually changed and what to do next.
Compare my earliest entries with my latest, then give me:
1. What has specifically improved, with entries as evidence.
2. A mistake that recurs across the log.
3. The single habit that seems to have produced the biggest gain.
4. A focus for the next 30 days.
Quote the log when you make a claim. If the log is too thin to support a conclusion, say that plainly. Be direct, not flattering.
Before answering, check each claim against the log.
Hypothetical: Priya runs a small bakery and wants to get good at writing supplier emails. Her skill sentence is "writing supplier emails". In week one she drafts a request for a delivery change cold, then asks for the expert version, and notes that she never stated a deadline. In week two she rebuilds the prompt from memory and finds she left out the tone instruction, which did matter. In week three she writes the same email for a friendly supplier, a late one and a new one. In week four she sends two real emails and explains her approach to her business partner. Her log shows the same recurring slip: she forgets the deadline. That is next month's focus.