AI Guides › Playbooks
By Nigel Guy · 7 min read
The usual way to "go all in on AI" is to announce a grand system, subscribe to six tools, and start logging everything. It feels like progress because the dashboard is busy. A month later you have forty days of data, no decision it changed, and a new chore called "updating the system".
A note on the term first. The phrase "AI maxxing", and the idea of running your life like a startup while tracking every input like a founder, comes from a social-media post that we could not verify against a primary source, so we have not described what its author does. What follows is the durable part: a small, repeatable structure you can run with any AI assistant. It is built on one honest observation. A founder tracks a handful of inputs that drive an outcome, reviews them on a schedule, and changes something. Most of what gets called "tracking" skips the last step.
The rule: track only the inputs you are willing to change, review them on a fixed day, and let the AI do the summarising while you do the deciding.
The Ledger has five parts, done in order. Parts one to four are set-up and take an evening. Part five is the loop that runs every week.
Choose one outcome for the next 90 days, written as something you could check. "Be healthier" fails. "Run 5 km without stopping" or "finish the first draft of the course" passes. If you pick three outcomes, you have picked none.
An input is something you control that plausibly moves the outcome. An output is the result. Founders track both, but only inputs are actionable.
| Outcome | Good inputs (you control) | Poor inputs (vanity or outcome in disguise) |
|---|---|---|
| Run 5 km | Runs per week, longest run, sleep before run days | Steps all day, app "readiness" scores |
| Finish a draft | Writing sessions per week, words kept, days since last session | Hours "at the desk", number of notes saved |
| Save for a deposit | Amount moved on payday, number of unplanned spends | Net worth checked daily |
Limit yourself to three. Each must take under a minute to record. If recording an input is annoying, you will stop by week three, so choose a cheaper input.
Use a single place the AI can read. Both big assistants offer a workspace for this. Claude has Projects, where you can add project knowledge and set project instructions that apply to every chat in that project. ChatGPT has Projects too, grouping chats, files and project instructions, with settings under the three-dot menu of a project. At time of writing, both are available on free plans, with Claude's free tier capped at five projects. Check the current limits before you rely on them, because they change.
Keep the actual record boring: a note or spreadsheet with one row per week, a column per input, and a short "what I noticed" column. Upload or paste it into the project each week.
Write three rules now, while you are calm:
This is the loop. Paste your week's row into the project and use the prompt below. The AI summarises and challenges; you choose one change.
You are a calm, sceptical review partner helping me run a weekly review against one 90-day outcome. You are not a coach or a cheerleader, and you are not a medical, financial or legal adviser.
CONTEXT
- My outcome: [OUTCOME_WITH_CHECKABLE_MEASURE]
- Deadline: [DEADLINE_DATE]
- My three inputs and targets: [INPUT_1_AND_TARGET], [INPUT_2_AND_TARGET], [INPUT_3_AND_TARGET]
- My change trigger: [CHANGE_TRIGGER]
- My stop trigger: [STOP_TRIGGER]
- This week's record, plus the previous [NUMBER_OF_WEEKS] weeks: [PASTE_RECORD]
- My own notes on the week: [FREE_TEXT_NOTES]
TASK
1. Show each input as: this week, target, and the trend over the weeks supplied. Use only the numbers I pasted.
2. Say whether my change trigger or stop trigger has been hit. Quote the rows that show it.
3. Name the single most likely reason I missed or hit target, labelled as a hypothesis, not a fact.
4. Propose ONE change for next week, small enough to do in under 30 minutes, and say what result would tell me it worked.
5. List anything I tracked that did not influence any decision, as a candidate to drop.
RULES
- If any input, target or week is missing or unclear, ask me before answering. Do not guess or fill gaps.
- Do not invent statistics, benchmarks or "typical" figures. If a comparison would need outside data, say so.
- Do not give medical, mental-health, financial or legal advice; if the record suggests a health or money problem, tell me to speak to a qualified professional.
- Keep the whole reply under 300 words, in plain British English, no motivational language.
SELF-CHECK before replying: confirm every number you quote appears in my record, and that you proposed exactly one change.
Fill in the square-bracket items. If you have no earlier weeks yet, put "none" and the assistant will only report this week.
Imagine Priya, who wants to finish a first draft of a 30,000-word guide by 31 January. Her three inputs: writing sessions per week (target four), words kept per session, and days since the last session. She starts a project, pastes in a four-column table, and picks Sunday evening as review day. Her change trigger is two weeks under three sessions. In week three the review shows sessions at two for the second week running, and the trigger fires. The hypothesis the assistant offers is that her sessions are all scheduled after 8pm. The one change: move one session to a 7am slot. She keeps "days since last session" because it prompts her, and drops "words kept" after four reviews because it changed nothing. That is the Ledger doing its job: fewer inputs by week six, not more.
The trap is measuring for its own sake. Tracking feels like action, and an AI makes it cheap to produce charts and weekly reports, so the volume of tracking expands faster than the volume of change. Watch for these signs: you have more than three inputs, you review more than once a week, or you have revised the system more often than you have revised your behaviour.
A second trap is treating the assistant's explanation as a finding. It will always produce a plausible reason for a missed week. That is a hypothesis; test it with the one small change, not with another paragraph of analysis.
It will not choose your outcome, fix a bad plan, or replace a clinician, financial adviser or friend. It cannot tell you whether a goal is worth wanting. It also works on small data: a few weeks of three numbers can show a direction, not a proven cause.