In a previous video, we used a proxy to see what Claude Code sends to the model. Let's read the system prompt we found. These are instructions supplied by the harness alongside the task you give it. We will look at a few sections and connect them to behavior you have already seen when working with an agent.
This captured prompt contains about a thousand words. Our own request was only a short greeting, but the harness supplied instructions about how the agent should work. We can see the whole document here for scale. We will read a few excerpts at a larger size instead of trying to read everything on this screen.
The headings give us a useful map. There are instructions for working in the harness, keeping memories, understanding the environment, and continuing a long conversation. There is also session-specific guidance and a location for temporary files. Let's start with the opening, then visit the sections that explain familiar parts of the agent's behavior.
The opening tells the model which application it is working in and gives it a role: helping with software engineering tasks. It also includes guidance about which requests to help with and which to refuse. Refusing destructive requests is also behavior learned during training, and here it appears again as a written rule. These instructions are part of what the model receives before it answers us. Reading them makes some of the agent's expected behavior explicit.
The Harness section gives practical instructions about doing the work. Here the model is asked to prefer dedicated file and search tools when they fit the task. If you have watched the agent choose a file-reading tool instead of running a shell command, this is one of the instructions that helps explain that choice.
Other lines are about the quality of the work and how it is reported. Match the style of the surrounding code. Say when tests fail or a step was skipped. These are ordinary written instructions, much like instructions you might give a collaborator. A capable model can apply them across many different tasks without a separate programmed rule for every situation.
The Memory section explains how to keep useful information for future sessions. It tells the model to write notes, update existing notes instead of making duplicates, and remove information that turns out to be wrong. The harness provides file tools, storage, and a way to load saved information later. The instructions guide the model in using those facilities. This is a concrete example of a feature of the harness that was built partly by telling the model how to behave.
The prompt also tells the model where it is working. Here it names the working directory, says that it is outside a Git repository, and identifies the platform and shell. It even names the model as Haiku four point five and gives its stated knowledge cutoff as February twenty twenty-five. The model receives these facts from the harness. They help it choose suitable commands and understand the setting for the task without asking us for that basic information each time.
The Context management section tells the model what to expect when the conversation grows long. Earlier material can be summarized so work can continue within the available context. The instruction here is to keep working rather than finish early because space is running low. We have already discussed compaction. What is interesting here is seeing the harness explain that process to the model in words.
These examples help explain why an update to a coding harness can change how the agent behaves even when the underlying model stays the same. Part of the update may be a change to these instructions. Strong models make this a relatively inexpensive way to add or refine behavior: describe what the model should do, and provide the tools it needs to do it. Anthropic also reported removing about eighty percent of the system prompt for Opus 5 and Fable 5 with no measurable loss on their coding evaluations.
Reading the prompt lets us see that part of the design. Your own CLAUDE.md is supplied separately, and in our saved example with that file it arrives inside the first user message. We found guidance about writing code, reporting results, keeping memories, understanding the environment, and continuing long tasks. When you inspect your own agent's prompt, look for instructions that explain behavior you recognize. The text may differ from this capture, but you now know what kinds of things to look for. Below this video there is a page from a later capture. It puts the prompt that Claude Code sends to Haiku next to the one it sends to Fable, with the differences marked word by word. Fable, the more advanced model, gets a much shorter system prompt, less than half the length.
Happy prompting. See you in the next video.