AI Guides › Playbooks
By Nigel Guy · 7 min read
Most people answer the AI question with an identity. One camp refuses on principle and calls it standards. The other hands over every task it can and calls it leverage. Both feel decisive, and both skip the one thing that matters: deciding, task by task, who does what and how you will know the result is good.
The rule: Decide per task, not per identity. For every piece of work, name what you keep, what AI drafts, and what check you run before anything leaves your hands.
The brief for this guide describes a viral Dan Koe essay that splits people into three groups: Purists, AI Maximalists, and a third group. I could not locate that essay, so I cannot confirm its wording, its definitions or what it says the third group does. The labels below are therefore my own working definitions, not quotes or a summary of his argument.
What I could verify is adjacent. In a June 2026 letter on his future/proof newsletter, Koe argues that adopting an "anti-AI" identity feels like action while changing no behaviour. In a November 2025 letter, he argues that most people treat AI like a slot machine, typing one line and hoping, rather than briefing it in detail. Those two points are real and are the durable core of this guide. The mechanism, the ledger, is mine.
The Purist default. Refuse, or use it so rarely that you never learn what it can and cannot do. The appeal is real: you protect your own thinking and avoid a lot of mediocre output. The cost is that your judgement about AI stays theoretical. You cannot tell a bad answer from a good one in a tool you have not used, and you cannot advise colleagues or clients who are using it anyway.
The Maximalist default. Delegate everything the tool will accept and ship what comes back. The appeal is speed. The cost is that the model fills every gap in your instructions with the average of what it has seen, so your work drifts towards everyone else's. Nobody on the receiving end is told that nobody checked it.
The third position is not a midpoint. It is a different question. Purists and Maximalists both ask "should I use AI?" The third position asks "for this task, what is my part, what is the tool's part, and what is the test?"
A ledger is one row per recurring task. You keep it somewhere you will actually open, such as a note or spreadsheet. Five columns.
| Column | What you write | Example row (hypothetical) |
|---|---|---|
| Task | The recurring job, in a few words | Monthly client update email |
| I keep | The part only you can do | Deciding what the client needs to hear; the final send |
| AI drafts | The part you hand over | First-pass structure from my bullet notes |
| Brief | What the tool must be told every time | Audience, tone, length, three facts, what to leave out |
| Check | The test before it leaves you | Every figure matches my source; I would say each sentence aloud |
Hypothetical: Priya runs a small bookkeeping practice and writes a monthly update to each of twelve clients. Her Purist habit was writing each from scratch, which took an afternoon. Her brief experience of the Maximalist habit was pasting "write a client update" into a chatbot, which produced something generic that she then felt too embarrassed to send.
Her ledger row says she keeps the decision about what each client needs to hear. The AI drafts structure and wording from her bullet notes for that client. The brief names the client's sector, her usual tone, a length cap and a rule against any figure she has not supplied. The check is that she compares every figure to the source spreadsheet and deletes any sentence she would not say to the client's face.
Nothing about this is dramatic. That is the point. The tool is used inside a process she owns, so she can see where it helps and where it does not.
Fill in the bracketed parts, then paste it into the AI tool you use.
You are a careful workflow coach helping me decide how to split one recurring task between me and an AI assistant.
Task: [TASK_NAME]
How I do it today: [CURRENT_STEPS]
What goes wrong or wastes time: [PAIN_POINTS]
What could go wrong if the output is wrong: [STAKES_AND_WHO_IS_AFFECTED]
My standard for "good": [QUALITY_BAR]
Do this in order:
1. If any of the five inputs above is missing or vague, ask me for it before continuing. Do not guess.
2. Propose what I should keep, with a one-line reason for each item.
3. Propose which narrow parts the AI could draft.
4. Write a reusable brief I can paste each time, with [SQUARE_BRACKET_PLACEHOLDERS] for the inputs that change.
5. Propose three checks I can run in under ten minutes without the AI's help.
Format: a five-row table (Task, I keep, AI drafts, Brief, Check), then the brief in its own block.
Constraints: do not invent facts about my business. Mark any assumption as an assumption. If a task looks too high-stakes to delegate any part of, say so plainly.
Before answering, check that every check you propose is something I can actually do, and that nothing in "AI drafts" is something I said I must keep.