The system prompt

One file, src/engine/prompt.ts, 53 lines. It is short on purpose: every line in it earned its place by fixing something that was visibly wrong.

It enters two different ways

Engine	How	What that means
claude	--append-system-prompt	It is appended to Claude Code's own prompt, not swapped for it. Instructions about editing files and git are still in there, and we cannot remove them
Kimi, OpenAI, …	first role: "system" message	It is the whole prompt. Nothing else

That difference is half the reason the HTTP engines exist. Driving the CLI, we are a guest adding a paragraph to the end of somebody else's instructions. With our own loop the prompt is ours, and not one word of it is about git.

What it says

That there is a panel to its right. Without this the agent answers with the full inventory in prose — device by device, IP by IP, restating what the panel already draws — and then adds an ASCII diagram of the topology on top. Duplicated and worse.

The shape of a reply. The chat is ~70 columns: two or three sentences for normal work, no ## headings on short answers, tables only for verification results (pings, health checks) where they actually help, code blocks only for CLI you are going to paste.

One specific, expensive instruction:

Use pt_export_topology, not pt_query_topology. Both list devices, but only export carries the links, and without links the panel cannot draw the hierarchy: you get a flat list grouped by subnet instead of the router → switch → hosts tree.

That one was learned watching the panel go flat for no visible reason. It is not a just-in-case line; it is the difference between a tree and a list.

How to work. Verify against the device, not against the plan — a tool returning OK does not prove Packet Tracer did what you asked. If something fails, say so plainly with the cause. Look before you build: pt_bridge_status, pt_export_topology.

The language, explicitly. Otherwise the agent replies in the language of its own instructions regardless of what the interface is set to, and an English interface answering in Spanish is in no language at all. /language changes both, and the session relaunches so the new prompt takes effect.

What it deliberately does not say

It does not ask for structured output. There is no MODULE=… envelope, no agreed block for the sidebar, no "answer in this shape and then your prose". The panel is derived from tool results — see architecture for why that is a correctness decision and not a stylistic one.

It does not describe the 61 tools either. Those arrive from the MCP server with their JSON Schema attached; restating them in the prompt would be a second copy to keep in sync, and it would go stale the first time the MCP adds one.

Changing it

Keep it short. A system prompt is charged on every request of every turn, and a long one buys less than it costs: the models that follow instructions do not need the repetition, and the models that drift will drift anyway.

The test to apply before adding a line: did the agent actually do the wrong thing without it? Every line currently in there passes that test. Add one that does not and you are paying tokens for a superstition.