openclaw
A review of the system prompt openclaw sent to the model, recovered from the trace proxy: the largest agent prompt in the field, stored raw and linked.
Recovered with lab prompts openclaw from the trace proxy (newest capture 20260711T110411Z).
Every tool routes through the proxy, which records each completion after normalizing it to the chat-completions shape, so this is the exact text that reached the model, not a copy from the tool's source.
The prompt is stored raw under prompts/openclaw/; this page reviews it and links to the file.
Sizes are in tokens, counted with the o200k_base encoding.
Agent prompt
- 7,413 tokens
- tools (24):
apply_patch,exec,edit,read,write,web_fetch,web_search,memory_get,memory_search,sessions_spawn,subagents,cron, and more - raw: prompts/openclaw/1-agent.md
openclaw carries the largest prompt of any wired agent, and reading it explains why.
It is a personal-assistant brief that tries to cover the whole surface in prose: a broad toolbox of twenty-four tools spanning shell, patching, memory, web, cron, and multi-session orchestration, each introduced and hedged with when-to-use guidance.
The design bets that a thorough single briefing makes the model steadier, and the cost of that bet is a fixed prompt many times the size of tomo's, paid on every call.
On the Hi! baseline that fixed weight is what puts openclaw near the top of the token table before any work happens.