Weave gives you full control over every prompt it sends to the model. Nothing is hardcoded and hidden — every piece of text that shapes how the model behaves is editable. You can inspect the defaults, replace them entirely, or leave them blank to restore the built-in behaviour.
Where prompts live
Prompt configuration is spread across three places, reflecting how specific it needs to be:
- Settings → AI — global defaults and automated-behavior prompts that apply across all conversations.
- Settings → Models & Agents — a system prompt attached to a specific agent, used whenever that agent runs.
- Per-conversation override — a one-off system prompt set inside an individual chat, overriding the agent default for that session only.
Global system prompt
The global system prompt is prepended to every conversation and task, regardless of which agent runs it. Use it to establish a baseline persona, enforce rules that apply everywhere, or supply context that every model invocation should have — your name, preferred language, how you want the model to format replies.
Set it under Settings → AI → Global system prompt. Leave it blank to send no global prefix.
Agent system prompt
Each agent can carry its own system prompt. When an agent runs, its system prompt is combined with the global prompt and any active skills to form the full context prefix. This is the right place for instructions that only make sense for a particular model — a coding-focused agent might have strict tool-use rules that you wouldn't want on a general-purpose agent.
Edit it under Settings → Models & Agents → [agent] → System prompt.
Per-conversation override
Inside any chat you can set a system prompt that replaces the agent's system prompt for that conversation alone. The global system prompt and skills still apply; only the agent-level portion is swapped out. Useful for quick experiments or one-off sessions without touching your agent configuration.
Find it in the chat's settings panel (System prompt override).
Automated behavior prompts
Beyond what the model says in conversation, Weave uses prompts to drive its own internal behaviours. Every one of these is editable under Settings → AI:
- Chat title generation — the prompt that asks the model to produce a short title for a new conversation. Replace it to change the format, language, or length of titles.
- Chat-to-Weave summary — used when Weave converts a chat into a structured Weave. Controls how the model distils a conversation into a plan of tasks.
- Task decomposition — the prompt that breaks a high-level goal into the individual tasks a Weave executes. Adjust it to change how tasks are scoped, named, or ordered.
Each automated prompt falls back to a sensible built-in default when left blank, so you only need to touch the ones you want to change.
Skills as prompt extensions
Skills complement the system prompt layers above: each skill is a Markdown document that gets injected into the system prompt whenever it's active. Skills let you build a library of reusable instructions — tool guides, style rules, project-specific context — and mix them into any conversation without editing the underlying prompts.
The full system prompt sent to the model on any given turn is assembled in this order: global system prompt → agent system prompt (or per-conversation override) → active skills. Each layer is optional; omit any or all of them.
