The Six Levels
Which level for which job, on one page you can keep by your desk.
Start at just-ask. Climb only when the job forces you to: iteration (loop), a multi-step run on live tools (goal), outside facts (deep research), a hard decision (deep thinking), or working code (ultracode). Length, density, or a stressful deadline are not reasons to climb. Each card below leads with the tell and when to skip it, because the boundaries are where people mis-pick, not the definitions.
just-ask
One clear request, one good answer, done.
You could get a usable answer in a single reply. Nothing to check against the world, no draft to polish, no decision riding on it.
"Summarise this PDF." "Rewrite this plainer."
The first answer will not be good enough and needs cycles (loop), or the answer lives in sources you have to go find (deep research).
loop
Draft, grade it against a written bar, improve, repeat until it passes.
You can name the finished thing and the bar it must clear, and you know draft one will miss it.
"Ten cold emails that don't read as spam." "A one-page pricing brief."
There is no artifact to polish and the real output is a decision (deep thinking), or it has to run as software (ultracode).
goal
You name the outcome, the AI plans the steps and runs them with connected tools.
A multi-step job across live tools, no single document to grade, and a connector exists that can act.
"Triage my inbox and draft replies to anything urgent."
There is one document clearing a bar (loop), or no connector exists to take the action (it leaves at the gate).
deep research
Go out to many outside sources, cross-check them, come back with a cited report.
The answer lives in the world, not in your head or your files, and you need it sourced.
"Research the UK market for electric bikes."
The call is a judgement about your own situation (deep thinking), or the facts are already in your files (just-ask or loop).
deep thinking
Reason a hard, open question all the way through, a real thinking budget and both sides argued, because being right beats being fast.
The output is a decision or a diagnosis, not a document. Nothing to polish, and a quick confident answer could cost you.
"Should I take this job?" "Why does this keep failing?"
The answer needs outside facts you have to gather first (deep research), or there is a draft to improve against a bar (loop).
ultracode
Build real software or a non-trivial automation, then have a separate pass review it.
The deliverable is code that has to actually run, not a sketch or a description of one.
"Build a tool that scrapes and compares live competitor prices."
The deliverable is words or a document, not running code (loop).
Where is the eval? Inside loop and ultracode.
You will not find "eval" as a seventh level, and that is on purpose. An eval is not a kind of job you route to. It is the check baked into loop and ultracode: a separate reviewer pass, a maker and a checker that are never the same breath, the way an exam is marked by someone who did not sit it.
That check is the loop. Score the work against the bar, send it back if it fails, hand it over only once it passes. It is what makes the heavy levels trustworthy instead of the agent grading its own homework, and the more checkable the bar (a test, a number, a rubric), the better it works.
Before any level: does the job even belong here?
Empty input ("help") gets one question back: what are you trying to get done? And three kinds of job leave at the gate:
A real-world action with no connector (book a flight, make a payment). If a connector can do it, like Gmail or Calendar, it is a goal instead. A visual-design or taste job (a logo, a brand mark, a polished image) is a Claude Design or human-designer job. A promise about what another person will do ("guarantee a reply", "make this go viral") is something no level can promise.
Each time: name the limit, point at the right place, or reframe to the part the machine controls (for a reply, a tight loop against a checkable bar).
Two jobs in one sentence? Run a sequence.
Some jobs are two kinds of work wearing one sentence: gather then write, decide then build. The tell is an "X then Y" shape where X and Y need different levels. Run them in order, each with its own bar, the first feeding the second. For example "research our top three rivals then write a one-pager" is deep research, then a loop. Do not cram two jobs into one level, and do not over-split: a plain summary is one job.
Tip: print this (Ctrl or Cmd + P) for a desk-side reference or a workshop handout.
This cheat sheet is for learning to pick the level yourself. If you would rather brain-dump and have it sorted, use The Translator → It picks the level, sets the job up, and tells you why. Two doors when you get there: brain-dump if you roughly know what you want, or let it interview you first when the job is fuzzy or matters. To see the levels as rungs from one prompt to a team of agents, visit the Leverage Ladder →
McCloskey.ai · Resources. Built for people who want to use AI well without writing code.