Becoming an AI Power User

You no longer have to surface that tacit knowledge by introspection alone. You can ask the AI to interview you about a task, propose the criteria it thinks you are applying, and show you the failure modes you had not named. Articulation stopped being a solo act of staring inward and became something you and the model do together. That shift is recent enough that the framework naming the skill predates it, which the research note records.^[1]

Power users climb along four describable levels

The climb has a recognizable shape, and it is worth seeing the whole staircase at once before walking up it. Power users tend to move through four levels, and each one is defined by what you can now say about the task. The list below names all four as one family so you can locate yourself on it, and each level carries the sign that tells you the next climb is ready.

• Manual prompt: you write each prompt from scratch and learn the shape of the work. Sign you are ready to climb: you keep writing structurally similar prompts for the same kind of task, and only the inputs change.
• Template: you pull the stable shape into a reusable structure with slots you fill each time. Sign you are ready to climb: the template holds quality across a range of inputs, and your time goes to the inputs and the review rather than to the structure.
• Workflow: you connect templates into a sequence with quality gates that pass work from one step to the next. Sign you are ready to climb: the workflow runs predictably, you know where it fails, and your time goes to supervision and exceptions rather than to running the steps.
• Supervised delegation: the workflow becomes a standing instruction the process runs while you stay accountable for the result. Sign there is still a level above: the delegation you designed becomes something you supervise, and the supervision patterns become candidates for their own saved prompts and standards.

Write each prompt by hand while you learn the shape of the work

At the first level you write each prompt from scratch, and that is where it should be. You are learning what the model needs and what good output looks like for this task. Every interaction is an experiment, so the improvement principles apply here directly, and on more than one dimension at once: you reject weak output and diagnose the failures, you play with wouldn't-it-be-cool variations and try things that might not work, and you save what holds up. With the morning brief, this is the two weeks of typing the request fresh each day, watching which versions you keep and which you delete, and noticing the odd experiment that turned out better than the request you meant to write.

The sign you are ready to climb is repetition with a stable shape. You notice you are writing structurally similar prompts for the same type of task. The variation lives in the input data, the specific context, and what you accept as good. The shape of the prompt has settled, which is exactly what it means to be able to describe it.

Extract the stable shape into a template you can reuse

At the second level you pull the stable shape into a template, which captures the structure, the context it needs, the quality criteria, and the output format. You fill in the variables for each use. This is the first place AI-assisted articulation earns its keep: hand the model three briefs you wrote on three different mornings and ask it to extract the template they share, including the criteria you were applying without spelling out. Your attention shifts from crafting each prompt to refining the template. Does it handle the odd day with no meetings? Does it hold quality when the calendar is packed?

The sign you are ready to climb is a template that travels. It works across a range of inputs, and your time now goes to the inputs and the quality review rather than to the structure of the template itself.

Connect steps into a workflow with gates between them

At the third level you connect templates into a workflow. A step that gathers your calendar and overnight changes feeds a drafting template. The draft passes through a checking step. The check triggers a revision pass when something is off. Your attention shifts from running each step to designing the sequence, the conditions that pass work from one step to the next, and the quality gates that stop a weak draft from reaching you as a finished brief.

The sign you are ready to climb is a workflow that runs predictably. You know where it fails, you have a quality gate at each handoff, and your time goes to supervision and the occasional exception rather than to running the steps yourself.

Hand the routine execution to a supervised process and design the delegation

At the fourth level the workflow becomes a standing instruction the AI follows. As of 2026 that looks like an agent the model runs on a schedule, one present form of supervised delegation where the process handles the routine execution and you stay accountable for the result. You design the instruction set, the guardrails, the conditions that send a decision back to you, and the criteria the output must clear. Your attention moves to the architecture: which parts the process handles, how the steps coordinate, where your own call still produces the better-aligned outcome, and how the whole arrangement keeps improving.

The form will change, and the move underneath it will not. The interface that runs these steps for you today is one snapshot. The chat box you type into now may become a voice you talk to, then glasses that carry world knowledge, then earbuds, then a brain-computer interface. What stays constant across every one of those is the climb: describe the work well enough to hand it off, then supervise it. There is always a level above, because the delegation you design today becomes the workflow you supervise tomorrow, and each new level reveals design problems that were invisible from below.

Knowing when to climb is a question of what you can put into words

Handing off too early is the first failure mode. You package a task before you can describe its failure modes, and the system inherits your blind spots alongside your judgment. The output comes back confident, consistent, and consistently wrong in ways you did not anticipate: a morning brief that cheerfully leads with a low-stakes reminder every day because nobody ever told it which changes count as urgent. The fix is the same each time. Go back down a level, do the work by hand until the failure mode shows itself, ask the AI to help you name the missing criterion, update your standards, and rebuild the handoff with the gap closed.

Staying manual long after you could climb is the opposite failure. You keep doing work a process could handle because the by-hand version feels safe and controlled. This caps the scope of problems you can take on, because your attention is spent on execution that could be freed for design. The diagnostic question is plain: could this task be done well by AI, and if so, how do I push the system toward that quality of output, rather than keep running it by hand out of habit? Where your own call still produces the better-aligned result, keep it; where the system could match or beat it once it is described, hand it off.

So the gauge is accountability, not a ceiling. A task is ready to climb when you can describe its inputs, its outputs, its criteria, its failure modes, and the points where it should escalate to you. That description is what keeps you accountable for output a process now produces. You do not have to arrive at the description by yourself. Articulation is a skill AI increasingly helps you build, and the trajectory runs toward models that anticipate intent from less, drawing out what you mean the way a skilled interviewer surfaces what a person knows but has not said. Where you cannot yet put a task into words, that is the next thing to work on with the model's help, not a wall that ends the climb.

Each level you can describe becomes the foundation for the next

The other principles teach you to improve at a given level; abstraction climbing moves you to a higher one. Raising your standards lifts the bar each output must clear. Improving the conditions sharpens the prompts and workflows you run. Curiosity and play find new capabilities, the wouldn't-it-be- cool experiments that turn into your next project. Supplying the right context loads the distinctions a task turns on. Verifying in proportion to stakes scales your quality checks. Saving what you learn carries your standards and prompts across sessions so each one starts sharper. Abstraction climbing then moves you to a new level where all of those run again, at higher altitude, on harder problems, with broader scope. None of them sits above the rest; they work at once, which is why the edge keeps expanding instead of settling.

This is one reason the work compounds upward. Each level you master and can describe produces the judgment, standards, and saved instructions that become the foundation for the level above. The power user who started by writing better prompts is now supervising delegated work. The power user who started by rejecting weak output is now calibrating the checks that reject it at scale. The altitude changed and the closing move stayed the same: encode what you can now describe into a prompt, standard, skill, or workflow, then evolve that system as the work and the models change. The system is there to protect the work and free your attention for the next problem; keep it light enough that it stays a tool for the work rather than the work itself.

Take the brief one more time, now as a do-versus-don't. Do not hand-build it forever, retyping the same adjustments by voice or by text every time the model changes, and do not package it before you can say what a good version does, which only freezes your early guesses into the system. Instead, climb. Once you can describe what your setup should do, what it reads, and where it tends to get things wrong, with the AI helping you put that into words, package it as a standing instruction the model maintains and move your attention to supervising it: checking that it still fits your day, adjusting the brief when your life changes, and deciding which new capabilities are worth folding in. The library and assistant mechanics that hold all of this together are the subject of the reusable-AI-assets chapter; here the point is the move itself. You are no longer the person building the setup by hand. You are the person who supervises the system that builds it, and who can still describe what good looks like when something drifts.

Altitude is a horizon, not an entry requirement. A beginner can run one honest turn of the loop today, describe one small thing well, and hand off one task, and the rest of the climb stays open for whenever they are ready to take it.

The move outlasts the model. Today you climb with Claude Fable in 2026 typed into a chat box. The same move carries through voice, glasses, earbuds, a brain-computer interface, and on toward systems far more capable than any of us, because the durable truth underneath it does not depend on the interface: people keep pushing their tools to do more, imagine more, and create more, so they can take on projects, solve problems, and make the world a little better. That is also why the people working on alignment matter, so that the more capable these systems become, the more reliably they help us build what we actually intend. You are learning the part that survives every upgrade: knowing what you want clearly enough to ask for it, and climbing as soon as you can describe the next thing well.