How to get real work out of AI — by treating it like a brilliant colleague you take the time to onboard.
A chat AI is astonishingly capable. But it has never met you, knows nothing about your work — and every conversation starts from a blank page.
That’s not a bad employee.
It’s an un-onboarded one.
Every conversation starts at an empty desk.
You already know how to get value from a brilliant newcomer. You don’t hand them a manual and vanish — you walk them round, explain what matters, show them how things are done here.
Instructions tell it what to do this minute. Onboarding helps it understand the place it walked into — so its help actually fits.
Brilliant. Fast. Tireless.
And brand new here.
A generic, grovelling email. It invents a reason for the delay, uses a tone that isn’t yours, and promises a date you never agreed to.
“Here’s the thread. Warm but precise, don’t over-apologise, don’t commit to a date — say we’ll confirm Friday.”
On-tone, references the real thread, holds the line. Sounds like you.
Same AI. The only thing that changed was context.
Everything it can see right now lands here. The desk only holds so much — when it’s full, the oldest papers slide off and it can’t see them anymore.
That’s why it “forgets” — not a flaw, just a full desk. The size of the desk is the one hard limit, the context window.
Before the desk overflows, it copies the essentials into a notebook it keeps. Next time, it opens the notebook first — and starts where you left off.
You’d never re-onboard a colleague every single morning. Same here — write it down once.
It doesn’t just answer — it can open your files, search, and carry out the steps. Like an intern who can actually open the filing cabinet, not just tell you what’s in it.
Write down how you want a task done, once. Then call it by name whenever you need it — and it’s done your way, every time.
“A 3-line summary, then every action with an owner and a date, decisions flagged separately.”
Rule of thumb: a recipe holds how to do something — not your facts and files. Those live in the handbook (next).
A chat window can only talk back. To give the AI a desk, hands and a notebook, you open it inside a workspace — an app that can see your files and do the work.
You still just type in plain English — no code. The names you’ll hear: Claude Code, Cowork, Codex. (The catch-all word is a harness — it just means “the place the AI works.”)
These are just plain-text notes it reads every time — written in markdown (ordinary text with a few tiny marks). In a workspace they’re small files; you’ll see the name CLAUDE.md. That’s the whole secret.
The handbook is what it reads every time. A wiki is what it looks things up in — your engagement’s facts: the client’s world, what you’ve found, what you’ve advised, as small linked notes with a contents page. A living map of the project.
project-os/ ├── CLAUDE.md # the handbook — read every time ├── README.md # what this project is ├── foundations/ # what you know │ ├── client/ # their world, stakeholders │ ├── brief/ # scope, goals, constraints │ └── domain/ # the subject matter ├── findings/ # what you’ve learned │ └── interviews/ # one note per conversation ├── deliverables/ # what you produce ├── playbooks/ # how you work — your skills └── log/ # decisions & learnings — filed back here
It reads the contents page, then follows the one branch it needs — like finding a spatula: kitchen → the drawer by the stove → there.
So the desk stays clear — it fetches the one note it needs, not the whole folder. And it writes back to log/ as it goes, so the project’s knowledge captures reality and compounds. You design the structure; the AI keeps it current.
People call this “dreaming.” Think of a school: nobody fixes the textbook mid-lesson — the day’s work happens, then the review happens after hours.
The golden rule holds: the AI proposes the changes — a human still approves them.
# for a heading, - for a bullet.Onboard the AI like a colleague and it compounds — a little more useful every week, because it finally carries your world.
Built on the Agent Onboarding Protocol (Emerald HQ). Draws on talks by Daisy Hollman & Lamis Mukta (Anthropic), Simon Willison (“give the AI what a new employee would need”), and the LLM-wiki idea from Andrej Karpathy & Jeff Gibbard.