Skip to content
FlintFlintby Torch
← LevelsBuilder · B1

Four words before you start

Build-along · ~8 min at your own keyboard

Most confusion in the first month comes from four words used loosely. The model is the trained network that turns text into more text. It keeps nothing between calls: on every turn it sees only what is sent to it. The harness is the program around the model. Claude Code, the Claude app and ChatGPT are harnesses: they decide what the model sees, which tools it can use, and what gets saved between sessions. The context window is everything the model can see on one turn, counted in tokens: your messages, its replies, the files it read, tool results, standing instructions. It has a ceiling, and answers get worse well before the ceiling as it fills. An agent is a model in a harness using tools in a loop: it acts, looks at the result, and picks its own next step instead of only replying. These words tell you where to look when something goes wrong. It forgot what you said yesterday: the model keeps nothing, and the harness didn't save it. Answers got worse late in a long session: the context is too full. It did something you didn't ask for: the agent chose its own next step. The rest of this track builds on these four.

The long version: Getting started ↗

Do this

0/4

Checkpoints

  • ▸You should now see the chat's own one-line answer for each of the four words.
  • ▸You should now see a new chat that can't tell you what you asked in the old one.
  • ▸You should now see four definitions written in your own words.
  • ▸You should now see one weekly job with its context listed and an agent-or-not call.
The three ways this goes wrong
  • You wrote that the model remembers your past chats. — The model keeps nothing between turns. A memory feature is the harness saving notes and sending them back in. Rewrite the definition to name which part does the remembering.
  • Your definitions match the chat's answer almost word for word. — Close the chat and write each one from memory, the way you'd say it to a colleague. Then check it against the lesson.
  • Every job you think of comes out as needing an agent. — Ask whether you could write the steps down in advance. If you could, it is a workflow: you fix the steps and the model fills them in. Save agents for jobs where the next step depends on what the last one found.

Where these ideas come from: Model, Harness, Context window, Agent, Stateless, Token (AI Hero dictionary). Paraphrased here.

Share what happened

Paste your four one-line definitions, in your own words, and the two lines about your weekly job: what goes in the context, and agent or not, with why.

Strip API keys, passwords, and tokens before you paste. Swap them for [KEY].

🔒 Don't paste confidential, personal-health, or payment info.