1. Three pieces, and that is the whole machine
Claude Code is Anthropic’s coding agent. You run it in a terminal, and it also lives in editors. Install it from the current Claude Code instructions. The install line changes, and this page will not freeze a copied command.
Once it is open, the work has three pieces. A model reads your request and decides the next step. Tools let it act: read a file, edit a file, run a command, or call something outside the chat. A loop feeds each result back, and the model decides again, until the job is done or the run gets stuck.
That is an agent. The chat is only how you watch the loop. A chatbot that answers once and stops is a different thing.
2. The harness is everything around the model
The model is not the product. The harness is the rest: which tools exist, when Claude must ask, the memory file, and the instructions it loads before you type. Two people can use the same model and get different work, because the harness is different.
A coding harness suits software work. An assistant that only chats suits a different job. Using one for the other’s work usually disappoints. Cloudadorn’s Grok Build series teaches the same shape for a different product. The habits rhyme. The controls do not.
3. When a run fails, read the step that failed
A bad result is usually one bad step. Open the trace and find the turn where it went wrong. A command that failed. A file it misread. An assumption it wrote down and then treated as fact.
Example. You asked it to fix a failing invoice test. It edited the export page, ran the test, and the same failure came back. The useful next sentence names the step: “The failure is in the invoice test, and you edited the export page. Open the invoice test.” Starting the brief over hides the step you already found.
4. The conversation is a budget
The model can hold only so much text at once. Some of that space is already used, by the tools, the memory file, and the instructions. A huge log or a whole folder dumped into the chat spends the rest. When the window fills, later answers get worse.
The habit is smaller inputs: the file that matters, the error that matters, the decision you already made. Prices, plan limits, and model rankings change. Re-test a tool on your own tasks every few months. A ranking is not a measurement of your work.
5. What to try on the next real task
Pick a small job you already understand. Tell Claude what finished looks like. Let it work. When it stops, read the tool calls, not only the last paragraph. If the result is wrong, point at the step. If the result is right, you still read the change before you keep it.
Keep the instructions in plain files in the repository, so you are not locked to one product’s private notebook. The next page is the job that stays yours: you manage, and the agent builds.