What I teach, written down once and for all.

The intern with amnesia

Plain chat. You ask, you get something useful, you close the tab, it is all gone. Many live for months inside one endless conversation, afraid to close the only memory they have. The model is brilliant and passive: it forgets you every session. Prompting better will not save you, because you will retype the prompt tomorrow.

The assistant that remembers

The most important shift on the whole ladder. Projects, Gems: you write the prompt once — role, objective, context, source files, brand voice, examples — and every conversation in that project loads it by itself. You stop opening with four hundred words and start with one line. This is also where you load trusted expertise, frameworks and methods, so it applies every time.

Build ONE project, for the work you do most. Do not skip to the spectacular stuff.

The coworker with file access

AI leaves the tab. You point it at a folder: it reads and writes your files, updates your spreadsheets, reorganises, creates. You stop uploading and copy-pasting. Seen at a client's: after a call, three CRM deals created, to-dos set, notes filed, where an hour of cleanup used to go.

The orchestrator

You stop talking to an assistant and build a team with a router on top. An orchestrator takes the request and routes it to specialists, each with its own context and role. The whole team is a folder tree of plain-text files, and work moves by commands, not clicks.

A warning at this level: every agent loads the context you defined. If those layers are coherent, the team amplifies coherence; if they are not, it amplifies incoherence faster than any single chat. The architecture is not the point. What the architecture protects is the point.

The factory

The team runs without you. You set the work, you sleep, you review on waking and ship what holds. Two caveats, both serious: tokens cost real money, a team running overnight can burn hundreds; and taking the human out of the loop means amplifying incoherence at full speed with nobody watching. Most people do not need this level.

Don't fire the question straight away

The flow has three beats, and everyone skips the third. One: offer two or three expertises the machine should take on before starting, and let it choose. Two: let it ask you its context questions, one at a time, until it has none left. Three: bring your critical thinking and your creativity to bear on the answer. Never accept the first draft. Never.

The first draft is a rough the machine hands you as a finished piece. It is the one thing it genuinely does badly.

Score the output out of ten, and give the example that scores ten

“What you just delivered is a four out of ten. Here is an example I consider a ten. What is missing from yours?” The score alone is not enough: the reference example carries the information. This loop yields more than any other, and it has a valuable side effect: it forces you to state your own standard, which most people have never had to do explicitly.

Always ask what is missing to reach ten. A criticism without the way out is just a mood.

One subject, one conversation

The golden rule, and the most broken. Beyond one subject the model circles its own material and repeats what it has just produced. On a file of two thousand pieces to describe, that means one conversation per piece. It sounds absurd and it is exactly what makes the output usable: each conversation starts from clean context instead of dragging the last one along.

The tell is unmistakable: the same turns of phrase come back, the same images, the same angles. Close it, open a new one.

Encode the rule the moment you find it

As soon as a way of working proves out, have it saved to memory, explicitly: “remember that you must always use dates in the French format”. A method that exists only in the head of whoever found it is not a method, it is a dependency. Encoded, it travels across people, projects and weeks.

It is the move that turns individual usage into a company asset. The only one, in fact.

Make it write its own prompts

Load a guide of prompting patterns, describe your need, and ask it to pick the right pattern then write the final prompt, ready to copy. Test it in a fresh conversation, watch the gap, iterate on the prompt and not on the output. You stop guessing what works: you have it formulated by the thing that knows how it is built.

True for images too: have the text model write the prompt, run it in the generator, iterate on the gap.

Document your voice by example, not by adjective

“Professional but warm” means nothing and produces exactly what you fear. Give three texts you approve and three you reject, and ask for the enforceable rules that separate them. You get a list you can reread, correct and hold up against an output. It is the only kind of tone-of-voice charter that stands up to a machine.

One approved example is worth ten adjectives. One rejected example is worth twenty.