They keep widening what they can do by working at the expanding edge
Chapter Progress: Early DraftPractice on the tasks just past your skill
The stall has a recognizable shape, and you have probably lived it. Early on, your summaries from AI were rough, you noticed it, and you adjusted: you added a line about what to emphasize, you said how long you wanted them, you gave an example. They got better fast. Then they got 'good enough,' and something quiet happened. You stopped reading them against what a better summary would look like. You stopped noticing the gap, because the gap stopped bothering you. From the outside it looks like you mastered the task; from the inside, you simply stopped measuring it.
re-opens the loop on purpose. The psychologist Anders Ericsson, who spent his career studying how experts become expert, described it as working on tasks just beyond your current ability, getting immediate feedback, and adjusting. In AI work, a simple version is to set yourself a concrete challenge: take something that normally costs you ten prompts and an hour of back-and-forth, and try to reach it in one well-built prompt. The value is in what the attempt shows you. Pushing the edge of what you can do reveals exactly where your current approach is weak, which is the information a plateau starves you of.
One running example: you want AI to give you a useful morning rundown of your day so you can plan it. As of 2026 this looks like typing into a chat model such as Claude Fable, and the interface will keep changing under it, to voice, then to glasses, then to brain-computer interfaces, then to whatever comes after. The example shows how two people approach the same goal. A person who is not improving runs it the same way every day. Each morning they type 'give me a summary of my day,' accept whatever comes back, and ask again tomorrow. They stay busy and learn nothing the routine carries forward. A person working at their edge asks for the version just past what they can reliably get today: the rundown that names their actual first meeting and flags the one thing they are dreading, then watches where it breaks. Maybe it invents a cheerful generic schedule instead of reading the real calendar. That break is the feedback. It names the exact thing the next instruction, or a standing piece of context that connects the calendar, has to fix.
The adjustment depends on knowing your own intent well enough to put it into words a model can use. You have to know what you wanted, surface what you expected, and turn that into context or an example the model can act on. This discipline, externalizing the in your head so a system can use it, sits underneath everything in this chapter. The clearer you can make what you were after, the more of it you can hand off as instruction. Diagnosing the gap itself is often work you can delegate: when a result disappoints, you can ask the model to read its own output as a symptom rather than a verdict, the way an experienced doctor does not stop at 'the patient feels unwell' but asks what the symptom points to. Ask it what it would have needed to do better. Did it lack context? Did it have no example of what good looks like? Was the prompt too thin for how large the task was? An evaluation you set up once can flag those gaps for you on every future run.
The skill you are building outlasts every interface. The output of any one session fades. The skill of focusing your attention on what you want, and putting it into words or examples a system can use, carries across every tool you will ever touch, and across every change to how those tools work. The chat box becomes voice, then glasses, then a brain-computer interface, then whatever follows. The discipline of knowing and saying clearly what you want stays the same.
Play to find uses you would not have planned
Inspecting the work sharpens what you already do, and play finds what you have not tried. Power users run small experiments and side projects, and they keep asking the open question that opens new ground: what if this were possible? Wouldn't it be cool if AI could do this? They let the question reach past what they already know works.
These experiments stay concrete and low-stakes. Have two AI tools pass work back and forth so you stop copying and pasting between them by hand; as of 2026 that might look like one model handing a task to another, and the pairing will change while the move stays the same. Point one piece of your setup at a problem only you have. Ask the model to brainstorm twenty uses for AI in your week, then try the three that surprise you. The model is often a better source of 'have you considered' than a blank page is, so power users use it to widen their own imagination, not only to carry out a plan they already have.
Play and inspection feed each other. A loose experiment often surfaces a skill worth sharpening on purpose, and a sharpened skill makes the next experiment more ambitious. They keep handing work back and forth, and whatever proves out is worth encoding so the system carries it forward.
Encode what you master, then aim higher
Each reliable gain brings the next one into view. After you get AI to reliably draft a client update from a few lines of context, two new things sit within reach that were out of reach a week ago. You could have it pull the underlying numbers itself, or have it critique its own draft against your standards. Each thing you can now do reliably puts a slightly more ambitious thing inside your reach.
The move that keeps that edge expanding is to evolve the system: take what you just did by hand and encode it so the system does it next time. You look at a task you have now done well a few times, step back to see its repeatable shape, and turn that shape into a reusable part: a saved prompt, a standing piece of context, a written standard, a rule, a skill, an instruction that connects a data source, a workflow, a specification the model follows on its own. Then you supervise that part from one level up, which frees your attention for the next problem. This is abstraction jumping, and it is the engine of the whole book: you keep climbing from single prompts, to the context you supply, to the intent you express, to harnesses that run on their own, handing more off at each level. The next chapter, on building the system that builds the work, develops this orientation, and a later chapter on climbing to higher levels of abstraction shows how far it goes.
Hold these forces together and the reason power users keep improving comes into focus. Inspecting the work raises the quality of what they get. Imagination and play widen what they attempt. Evolving the system turns each mastered level into the floor for the next. One caution keeps this from tipping into busywork: use enough of it to protect the work, not so much that the practice becomes the work. On a throwaway task, a good-enough output and a quick move on is the right call. The aim is not to keep yourself in the loop for its own sake. When a result matters, the move is to judge whether AI can reach the quality you need and, if it can, to push the system toward that quality rather than doing the work by hand. The aim is the best aligned result from the best available intelligence, yours or the model's, and to keep improving the judgment wherever it is weaker, on either side.
This way of working does not expire with any one model. It held for chat, it holds for Claude Fable in 2026, and it will hold through voice, glasses, brain-computer interfaces, and whatever artificial superintelligence becomes, because the human part is durable: people will always push the systems in front of them to do more, imagine more, create more, and make the world a little better. Keeping those systems aligned with what we actually want is hard and serious work, and the researchers carrying it deserve real credit. For everyone else, the through-line is simpler: take on a project, solve a problem you care about, and build something good, and let your way of working keep widening what good you can do.