The Rules of Engagement Guide

Anyone who works with AI on something that matters to them: a report due by five, an essay for Monday, a decision about a move, a plan for a trip, a letter to a grandchild. The guide is written for three settings in particular, because the rules look different in each. At work, the parts you hand off decide what you learn and what you can defend. At school, the effort is the point of the assignment. In later life, an assistant can do things for you or with you, and the difference decides what you keep.

It assumes no prior reading. The three rules rest on The Capability Fitness Framework (CH-F-001), which sets out why capability develops through practice and how help can support that practice or substitute for it, and on The Capability Model (CH-M-001), which names the specific ways delegating too early substitutes for the practice that builds each skill. The Rules of Engagement Model (CH-M-002) is the model behind the rules; this guide applies it. If you want the argument, follow those references. If you want to get better at working with AI this week, start here, and keep The Rules of Engagement Check (CH-I-001) beside the keyboard.

One boundary up front. This is not an argument for working with AI less. Help that removes a barrier, does work that carries no stakes for your growth, or restores what illness or injury has taken away is good help, and the right call is to take it. The failure this guide targets is narrower and quieter: help that substitutes for the effort through which a capability develops, on work where that capability still matters to you. The opposite failure is real too: keeping work you would be better off handing off, and calling it rigour.

The decision. Which parts of the work the AI does, and which you do. It is made before the work starts and again at every hand-off inside it.

What you are really deciding. Every task has two kinds of effort in it. Some is load-bearing: framing the problem, forming a first view, weighing the evidence, choosing between your own ideas, explaining the result in your own words. That effort is the practice, and it is the part a capability is built from. The rest is incidental: assembly, formatting, retrieval, the version of the thing you have done so often there is nothing left to learn from it. The retain call keeps the load-bearing part. The delegate call hands off the incidental part, without guilt. They are the same decision made well.

The questions.

1. What part of this is the thinking I am responsible for, or the practice I would be sorry to lose? That part you keep: at least the first attempt, in your own words, before you ask. (A rough instinct counts. Write it down before you open the thread.) If I hand it over a hundred times, what will I be left able to do?
2. What part is assembly, formatting, retrieval, or work I have done so many times there is nothing left to learn from it? That part goes to AI, without guilt.
3. Am I delegating by importance, or by convenience? Convenience hands off whatever is hardest right now, which is often the part that matters. Importance hands off what does not need you.

Two more, from the book, for when the first three do not settle it: Is this a part of the task I would want to be able to do next month without help? Is it the part I am here to learn, or the part I will be held accountable for?

The retain default. One rule inside the rule: your first attempt, in your own words, before you ask. Write the rough answer, the instinct, the one-line view, before you open the thread. It takes a minute and it changes everything that follows, because now there is something of yours on the page for the AI’s answer to be compared against. Without it you are not evaluating; you are adopting.

Signs it has gone wrong. The hard, important work goes to the machine because it is hard, and the easy work stays with you because it is easy. Finished work arrives that you could not explain or defend. A skill you used to have feels slow and unfamiliar on the day the tool is unavailable.

Your call, in words: I keep ______. AI does ______.

At work

A junior analyst has the quarterly numbers due to her manager by five. Before she types anything she has a choice about which parts of the report the machine will write: the tables and the first pass through the data, or the paragraph that explains why sales in one region fell. Left to convenience, the paragraph goes to the machine because it is hard and the tables stay with her because they are easy, which is exactly backwards. Explaining numbers is the job she was hired to learn. So she gives the assistant the tables, the formatting, and the first pass, writes two sentences of her own about the region before she asks for anything, and keeps the paragraph. The machine’s figures are in front of her and its first explanation is there to argue with. The paragraph comes out better than either would have written alone, and she can defend every line of it on Friday.

Her call: I keep the explanation. AI does the tables, the formatting, and the first pass.

At school

A tenth-grader has a lab report due and an assistant that will write the whole thing from his data in thirty seconds. The report exists so that he learns to turn measurements into a claim and defend it; the formatting, the reference list, and the tidy graph exist so that the teacher can read it. He writes his own one-paragraph conclusion first, badly, then asks the assistant to find the holes in his reasoning, not to write a better one. The assistant does the graph and the references. The claim, and the argument for it, stay his, and so does the practice the course was built around. The same move works for an essay: the thesis and the outline are his first; the assistant is a tutor he can argue with, never a ghostwriter.

His call: I keep the conclusion and the argument. AI does the graph and the references.

In later life

A retired engineer in his seventies is writing the story of his family for his grandchildren. The assistant can transcribe his recordings, build a timeline from his notes, check the dates against the records, and tidy his paragraphs. He lets it do all of that. The memories, the judgments about what mattered and why, and the voice are his, because they are the point of the thing and because they are what his grandchildren will hear. When the assistant offers to polish a passage, he reads the result, keeps the fixed commas, and puts his own sentences back.

His call: I keep the memories and the voice. AI does the transcription, the timeline, and the fact-checking.

The decision. When the AI acts on its own, and when you take over. It applies whenever the AI is doing something rather than answering something: drafting ahead of you, running a sequence of steps, sending, booking, buying, posting, or acting in your name. It is the newest of the three rules, because AI that acts is newer than AI that answers, and it is the one that matters most as assistants become agents.

What you are really deciding. Where to place yourself in a process you are not performing. Delegate or Retain divides the work. Let it run or Step in governs your presence inside the part you handed off. A task can be fully delegated and still need you at one point in it.

The questions.

7. Where is the point of no return? Sent, published, spent, committed, told. Step in before it, not after.
8. What would tell me it has gone off course, and will I see it in time? If you cannot see it, do not let it run on its own. Has my attention drifted only because nothing has gone wrong yet?
9. Could I still do this part myself if I had to? If the answer is no and the part matters to you, step in sometimes on purpose. The rep is what keeps the capability.

And, from the book: What will the machine do without me, and in what order? Where will I look in on it? Which step carries the judgment I will have to answer for? If it goes wrong halfway, will I notice, and could I take over?

Oversight decays. The analyst who checks every AI summary in month one and none in month six has not made a new decision; the old one has quietly stopped being made. This rule has to be remade, on a schedule if necessary, because nothing in the process will remind you. Nothing going wrong yet is not evidence that you would notice.

The deliberate rep. Where the part you handed off is one you want to stay able to do, step in sometimes on purpose: do the step yourself, or a sample of it, at intervals, not because the AI is failing but because the rep keeps the capability. Pilots call it manual flying hours. A capability that is never exercised is not kept.

Signs it has gone wrong. A task runs to the end and no one can say who decided what along the way. You stop checking in the sixth month what you checked every time in the first. When the tool fails, you no longer remember how to do the step yourself.

Your call, in words: I let it run until ______. I step in at ______.

At work

At four o’clock the assistant offers to pull the final figures, assemble the slides, and send them to the analyst’s manager on its own. The point of no return is “send”. So she lets it pull and assemble, which is routine, and she takes the last step herself: she reads the deck once, as her manager will, and she presses send. Once a quarter she builds the deck by hand, start to finish, so that the day the tool is down is an ordinary day. Six months from now, the question that matters is whether she still reads the deck before it goes.

Her call: I let it run through assembly. I step in before anything leaves my desk.

At school

A student runs a group project through an assistant that researches, outlines, drafts, and formats citations in one pass. The point of no return is “submit”, and before it there are two steps that carry the judgment she will have to answer for: whether the sources are real and say what the draft claims, and whether the argument is one she can defend when the teacher asks. She lets the pipeline run, stops it before submission, opens every source, and rewrites the argument in her own words. For one assignment a term she does the whole thing by hand, on purpose, which is how she knows the pipeline is a convenience and not a crutch.

Her call: I let it run through the first draft. I step in at the sources and the argument, and before submit.

In later life

A woman in her late seventies lets an assistant handle her pharmacy refills and reminders, and steps in herself whenever a dose or a prescription changes: the one point where a mistake would matter and where her own understanding of her treatment is worth keeping. The refills are substitution she should welcome; nobody gets more capable by phoning the pharmacy. The moment of change is where she wants to be present, and she keeps that understanding in use by explaining her own treatment to her daughter now and then, in her own words.

Her call: I let it run on the routine refills. I step in whenever anything changes.

The rules are the same everywhere. What changes from setting to setting is which part of the work is load-bearing, what the common slip looks like, and who else is in the room.

Two notes on the settings. At work, the rules are the practical answer to a hard pattern in the labour market: the work through which professionals used to build expertise is the work AI does first. A person who knows which part of the job is theirs can hand off the rest and still build the expertise the senior analyst has. At school, the rules are not an anti-cheating device; they are how a student works with AI to learn more rather than less, which a ban cannot teach. In later life, the guide makes no claims about ageing minds. It says only that an assistant that does things with you keeps your understanding in use, and that the moments of change are the ones to be present for.

Run the check before a task that matters. The Rules of Engagement Check (CH-I-001) is the nine questions in this guide on one page, with a “your call” line for each rule. Two minutes before a task that matters, not before every task; if it takes longer, you are using it on something that did not need it. The first three questions come before you open the assistant, the next three when the work comes back, and the last three whenever AI is doing something on its own.

Write one line afterward. What is one thing I will keep doing myself, on purpose, and one thing I will hand to AI without guilt? Put it where you will see it next time. Over weeks these lines are a record of your calls, and the record is what shows the pattern moving.

Watch the pattern, not a score. There is no score, and the rules produce none. What is worth tracking is whether your calls change: what you now keep, how often you check, where you step in. A shift toward deliberate calls should show in weeks. That is what getting better at working with AI looks like from the outside. Whether it also makes you more capable over months is a different claim, and it is the one Capably Human’s research program exists to test.

Keep the deliberate rep. For anything you want to stay able to do, do it by hand sometimes, on purpose, on a schedule you set. The rep is what keeps the capability; nothing in the tool will remind you.

Let the rules move you both ways. Expect the list of things you hand off to grow as your calls get more deliberate. If it only ever shrinks, you are practising distrust, not judgment.

With Humably. Humably runs this practice with you on your own work: it asks for your view first, helps you test what the AI gave you, makes the three calls explicit at the end of a session, and keeps the record over time so you can see your way of working with AI changing. And it carries the check itself, as a card you can run before any task. Its promise is that you get better at working with AI, on your own work, with a better result now. Whether capability follows is what the app is designed to build and what Capably Human’s trial program tests, and the findings will be published whichever way they come out.

It is not an argument for working with AI less; its examples hand off more than they keep. It is not a list of which tasks should be delegated; that depends on the person, the stakes, and what they want to stay able to do, and the guide says only how to decide. It is not an assessment, a score, or a measure of anyone’s worth. And it is not settled science. The three rules are a working model (CH-M-002), the worked examples are composites, and the claim that practising the calls builds capability over time is a hypothesis that Capably Human’s trial program will test and report on, whichever way the results come out. Until then, the honest claim is the one this guide makes: a person who makes the three calls deliberately is better at working with AI than one who makes them by default.

1.0 — October 26, 2026. First edition, published on capablyhuman.com with the model (CH-M-002). The morning’s draft was edited the same day for the claims ladder (a shift “should show” in weeks; the closing claim narrowed to the practice), the in-app check, the check’s v0.2 wording, and two trims.

0.1 — October 26, 2026. First draft, written after CH-I-001 (published October 26) and CH-M-002 (draft) per the October 26 decision that the instrument publishes first. Three rules, each with the decision, the nine questions from the check plus the book’s, the signs it has gone wrong, and worked examples at work, at school, and in later life; the test in both directions; a settings table; eight common situations and the calls; how to practise; three questions for managers, teachers, and parents. The examples are consistent with Heavy Lifting, Chapter 14 and Appendix F. Later the same day the quoted questions were brought to CH-I-001 v0.2 (questions 1, 2 and 8). To be revised against the model and against first-user feedback from the check and the Humably Alpha.