Customize Instructions and Prompts¶
Use Settings -> Instructions to set personal preferences for how Numi responds to you. Editing the server's base system prompt in code is only needed when changing default behavior for every user on the instance.
Set your personal instructions¶
- Open your profile menu at the bottom of the sidebar and select Settings.
- Select Instructions and enter your standing preferences, for example:
- Click Save Changes.
- Start a new chat and send a test question to verify the behavior.
Personal instructions are saved per account and take effect on the next model request. Keep this field focused on response style, formatting, and workflow preferences; factual background about your projects and stack is handled automatically by Memory.
You can also update personal instructions through the settings API:
curl -X POST http://localhost:4567/api/v1/me/settings \
-H "Authorization: Bearer YOUR_TOKEN" \
-H "Content-Type: application/json" \
-d '{"personal_preferences": "Keep answers concise. Use metric units."}'
Developer: change the base system prompt¶
Shared instance instructions live as constants in
src/numi_chat/agent/prompt_builder.py, where _base_system_sections() defines
their order and build_system_prompt() assembles the system message.
- Edit the section constant that governs the behavior you want to adjust (such
as
TONE_AND_STYLE). - Preserve the citation and tool-result formatting rules: the web UI relies on
markers such as
[web:1],[file:2], and grouped[web:1, web:2]tags to link inline claims to source cards. - Restart or redeploy the server and start a New Chat to test (existing chats keep the system message that was stored when they were created).
How the prompt is assembled¶
Numi structures outgoing messages so the prompt only grows at the end, keeping historical message bytes identical across turns so provider prompt caches stay valid:
| Component | Contents | Where it is built |
|---|---|---|
| Stored system message | Shared assistant instructions, formatting, citation, and tool rules | build_system_prompt() in prompt_builder.py |
| Opening session context (prepended to the first user message) | Account instructions and selected memory (<user_context>), active model identity (<active_model>), uninstalled optional tool groups (<optional_tools>), and available skills (<available_skills>) |
context_layers.py and Agent._first_user_context() |
| Per-turn user envelope (stored with each user message) | The user's message (<user_query>), uploaded-file metadata and bounded PDF excerpts (build_additional_data_section()), and the message's send timestamp (<current_time>... UTC</current_time>, stored as a time_note part) |
Agent._persist_user_turn() and context_layers.current_time_note() |
Because the opening <user_context> block contains no live wall-clock timestamp
and each user message replays its own recorded <current_time> note, replaying
prior turns sends the exact same prefix bytes on every follow-up turn.
Limits and troubleshooting¶
By default, personal instructions are capped at PROMPT_MAX_PREFERENCE_CHARS (2,500 characters); longer text is truncated when building <user_context>. See Prompt settings for all prompt and history character budgets.