Prompt Packs

MiniBot composes the system prompt from a base file plus runtime fragments.

Base System Prompt

  • File-based (default): loaded from ./prompts/main_agent_system.md (configurable via llm.system_prompt_file).

  • Inline fallback: set llm.system_prompt_file = null (or empty string) to use llm.system_prompt instead.

  • Fail-fast: if system_prompt_file is configured but the file is missing or empty, the daemon will fail at startup.

Runtime Fragments

  • Channel-specific: place channel fragments at prompts/channels/<channel>.md (e.g. prompts/channels/telegram.md).

  • Policy fragments: add files under prompts/policies/*.md for cross-channel rules (loaded in sorted order). Every file in the directory is loaded unconditionally — the loader globs the directory and never inspects which tools are attached. Prompt text such as “applies when the graph_* tools are attached” is guidance addressed to the model, not a condition the loader evaluates. For a fragment that must appear only when a given tool is attached, add a gated _<name>_fragment() method to minibot/app/handlers/services/prompt_service.py instead (see _skill_catalog_fragment for the pattern).

  • Composition order: base prompt → policy fragments → specialist roster → skill guidance (with a catalog snapshot when preload_catalog is on) → runtime capability status → channel fragment → environment context → tool safety addenda.

  • Prompts directory: configure root folder with llm.prompts_dir (default: ./prompts).

Shipped prompt files

The default ./prompts tree ships seven files, each playing one role in the composed prompt (or the compaction pass):

File

Role

main_agent_system.md

Base system prompt — persona and ground rules

policies/delegation.md

Cross-channel delegation policy

policies/graph.md

Cross-channel relation-graph vs durable-memory routing policy

policies/tool_usage.md

Cross-channel tool routing policy

channels/telegram.md

Telegram channel fragment

channels/console.md

Console channel fragment

compact.md

Memory-compaction pass instructions
prompts/main_agent_system.md

The identity and ground rules: “You are Minibot, a self-hosted personal AI assistant.” It covers privacy/secrets hygiene, directness, act-now-over-narrate execution, and the durable-memory workflow (search before create, update matching entries, persist only confirmed facts). Loaded as the base fragment via llm.system_prompt_file (default ./prompts/main_agent_system.md); startup fails fast if the file is missing or empty.

prompts/policies/*.md

Cross-channel rules appended right after the base prompt, in alphabetical order, regardless of the active channel and regardless of which tools are attached.

  • delegation.md — keep trivial requests local; prefer delegate (result needed) or spawn_task (final answer) for multi-step specialist work; use fetch_agent_info when the roster description is not enough; never claim a delegated result, since it arrives as a later message; continue locally when delegation is unavailable or fails.

  • graph.md — decide between the relation graph_* tools and the durable memory_* tools: a fact naming two things is an edge and belongs in the graph. Note this fragment ships even when the graph_* tools are disabled; it relies on the model no-opping when the tools are absent.

  • tool_usage.md — route the model to the right tool: memory_* vs history, apply_patch for existing-file edits vs write_file/move_file/delete_file for file management, reuse canonical path fields from tool output, absolute paths in yolo mode.

prompts/channels/<channel>.md

Channel-specific formatting and capabilities, appended only when that channel is active.

  • telegram.md — enforce Telegram Bot API formatting: plain text, Telegram-compatible inline HTML, or Markdown (no web HTML, no JSON wrappers like {"answer": ...}); when delegation results contain attachments, send each with send_file and confirm briefly.

  • console.md — text-only interface: always use kind="text", report file paths in the reply instead of sending files, and never call send_file.

prompts/compact.md

Instructions for the memory-compaction pass: produce an information-dense summary with labeled sections (Goals / Progress / Decisions / Constraints / Context / Next), retain relevant context over aggressive shortening, and drop chit-chat and tangents. It is concatenated after the base system prompt whenever a turn triggers compaction; if the file is missing, a built-in one-liner summary prompt is used instead.

Editing the System Prompt

  1. Edit prompts/main_agent_system.md.

  2. Review for content, security, tone, and absence of secrets.

  3. Commit with a descriptive message.

  4. Deploy via Docker or systemd — both setups include the prompts/ directory automatically.