hermes

Recovered with lab prompts hermes from the trace proxy (newest capture 20260711T110400Z). 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 prompts are stored raw under prompts/hermes/; this page reviews what each one does and links to the file. Sizes are in tokens, counted with the o200k_base encoding.

Agent prompt

• 1,566 tokens
• tools (19): terminal, patch, read_file, write_file, search_files, execute_code, memory, delegate_task, todo, vision_analyze, image_generate, text_to_speech, and more
• raw: prompts/hermes/1-agent.md

hermes presents itself as a general assistant, and its toolbox is the widest-ranging in spirit: alongside the usual shell, file, and search tools it carries vision, image generation, and text-to-speech. The prompt stays mid-weight by leaning on the tool schema for signatures and using its prose to set the assistant's manner and its rules for delegating work. Worth noting from the traces: on a bare Hi! hermes fires more than twenty HTTP requests, but only one is a real completion. The rest are capability probing, hitting Ollama- and OpenAI-style discovery endpoints like /api/tags, /props, and /version, most of which the proxy answers with a 404 before the single scored call goes through.

Title generator

• 66 tokens
• no tools
• raw: prompts/hermes/2-title-generator.md

The second prompt is the smallest capture in the whole set: a two-sentence instruction to name a conversation in three to seven words and return only the title. It is the tool-free side call hermes makes to label a thread, and it costs almost nothing next to the agent prompt.