Front / Level notes

Six levels, one next move each

A level is neither seniority nor job grade. It is “how far you understand the layer of mechanism between you and AI”. Each level below states four things: what it feels like, what you should know, where people usually go wrong, and the single next move.

L1 Listen 0–15

You know \"AI is impressive\" but can't say which stretch it is impressive at; you have used a chat box, never let it finish one complete piece of work for you.

What you should know
  • The rough outline of what a large model is and what it can do
  • What the word “prompt” actually means
Where people usually go wrong
  • Treating AI as a smarter search engine
  • Assuming bigger models are more reliable

Single next move: pick one repetitive task you do every week, hand it over end to end once, and write down “what counts as done well”.

L2 Wade 16–34

You can get decent answers and have started trying tools; but output quality feels like a lottery — you can't say why this run was good and that one bad.

What you should know
  • What context, hallucination and knowledge cutoff really mean
  • In one good output, which part came from the prompt
Where people usually go wrong
  • Assuming “written in detail” equals “said clearly”
  • Switching models whenever something goes wrong

Single next move: take one task, rewrite the instruction with “criteria + counter-examples”, and compare the two outputs.

L3 Handle 35–52

You can wire tools together so one small thing really runs by itself; you know it will fail, and roughly where in the chain.

What you should know
  • How retrieval, tool calls and context windows fit together
  • Why cost and latency decide the design
Where people usually go wrong
  • Only “does it run”, never “can it be reproduced”
  • Treating one success as stable capability

Single next move:Add three things to your automation: criteria, logs, failure retry.

L4 Make 53–70

You build tools for your own work instead of using off-the-shelf ones; you are making trade-offs — which model, which step stays human, how much to spend.

What you should know
  • How to build an eval set and set a baseline
  • In a multi-step task, which step must stay human
Where people usually go wrong
  • Assuming “I know it” equals “my team can run it”
  • Convincing yourself with one case, never doing regression

Single next move: write your own way of working into a criteria checklist someone else can follow, and hand it to a colleague to run once.

L5 Orchestrate 71–86

You are designing “who does what, how it is verified, who owns the miss”; what you manage is not prompts but tasks, permissions, definitions, and reviews.

What you should know
  • The relation between Agent headcount and the task ledger
  • How permissions, audit and rollback land inside the process
Where people usually go wrong
  • Solving coordination problems by “adding one more Agent”
  • Handing the acceptance right to the executor

Single next move: for the line you manage, stand up a three-month task ledger: who executes, who verifies, what counts as closed.

L6 Set the board 87–100

You are moving the rules themselves: where data comes from, who sets the standard, who carries the cost; you can see where the next bottleneck will bite.

What you should know
  • The mutual constraints among compute, data, evals and organisation
  • How open/closed source, regulation and standards change what is possible
Where people usually go wrong
  • Treating “I can do it” as “this should be done”
  • Ignoring responsibility and externalities

Single next move: pick a system you are running, write which layer is most likely to block it in six months, and move that layer first.

How to use the three layers

Layer 1 (the questions) only samples, it does not judge; layer 2 (this table) translates the score into “the layer of mechanism you should add”; layer 3 (the panorama) puts that mechanism back into the world — showing which stretch of river, bridge, street, city, wild it occupies. After all three, your judgment about your own understanding will be far more accurate than a score.

A counter-intuitive way to use it: read the panorama first, then come back and test. Many people only discover after testing that what they thought they “knew” was just a name they had heard — because the panorama puts the noun back among its upstream and downstream, and you instantly see which side you are missing.

Take the test →  Open the panorama directly

Lupu site group · three stations
① The Agent Era | ② The AI Scale | ③ 10,000 Sites
Next stop → ③ 10,000 Sites
季夏为萤路普AI | 滇ICP备2023002020号-2