The Courier level 1 carries text between a person and a model tell asks once, pastes the answer, and moves on AI as a better search engine. Everything travels by copy-paste, so every trip loses context — the result is only as good as whatever happened to be on the clipboard. The two rungs other ladders spend here (never tried / tried and quit) are not levels of use; they are the same place off the ladder. next When an answer disappoints, don't re-roll it — say what was wrong with it. The first follow-up is the whole difference between this level and the next.
The answer tracks your phrasing, and the model can only see the words you typed — so the first thing worth improving is how you ask.
The Asker level 2 shapes the question until the answer holds tell gives a role, an example and a format — and iterates instead of re-rolling Role, examples, output contracts, chain-of-thought, multi-turn refinement. The same model, several times the usable output. This is the entirety of what the internet calls prompt engineering — one level, not the four that both reference ladders carve out of it. next Stop describing the material and start attaching it: the actual file, the actual error, the actual house style.
However well you ask, it still knows nothing about your code, your data or your standards — so the next thing to improve is what you feed it.
The Supplier level 3 feeds the model the facts it cannot know tell arrives with the file, the spec and the error already gathered — and keeps reusable material instead of re-explaining The level both reference ladders skip. Knowing the window is scarce, this person edits rather than stuffs: the relevant excerpt instead of the whole document, a standing rules file instead of repeating house style every session. next Stop being the transport. Give it a tool that reads the file itself.
Knowing everything, it can still only talk: every conclusion comes back for you to carry out by hand — so the next step is handing over the tools.
The Delegator level 4 hands over tools, and reviews conclusions tell reads a diff instead of copying code, and can say what the agent is allowed to touch Tool calls, permissions, a sandbox: it reads files and runs commands on its own. The job changes from carrying information to reviewing results — which is a promotion in leverage and a demotion in certainty, because now something happens whether or not you were watching. next Answer the question this level cannot: how do you know it worked? Write the check before the fix.
It acts now, and you have no reliable way to tell a correct result from a plausible one — so the next step is deciding who gets to say “done”.
The Verifier level 5 the dividing line lets a machine decide when it is done tell asks for a failing test before a fix, and treats a green gate — not a confident summary — as the finish line The line between AI being usable and AI being trustworthy, and the level neither reference ladder has. Tests, gates and reproducible checks decide completion; failures flow back into the next attempt instead of ending it. Everything below this line produces output; from here up it produces output you can defend. next Write down the judgement you have been applying by hand — the rules, the gates, the checks — so it survives you being asleep.
One window, one attention span, and all the judgement living in your head — so the last step is turning that judgement into something a team, and a fleet of agents, can run.
The Legislator level 6 turns judgement into institutions tell ships rules files, skills and gates that other people — and other agents — run without asking Rules in the repo, reusable skills, gates in CI, several agents with a verifier that is never the producer. You stop managing tasks and start managing the system that decides whether output is good. The ceiling of this level is still being written: coordination cost, and where trust ends. next The tools keep levelling; judgement, taste and the ordering of what matters do not. That is the only part of this ladder no release notes will hand you.
every ceiling is the next stage's foundation · click a stage to read why it exists
