System Prompts

What the model actually receives

Your system prompt comes first. SynapseOrch then appends sections of its own, so you don't need to write them:

• how to reason and call tools, and — for CLI and Bedrock models, which have no native tool calling — the tool list itself
• how to handle large tool outputs saved to the vault, and how to keep a task list
• the schema of any linked databases and the linked repositories (Code agents), the members' roster (Team agents), selected skills, and the memory index when memory is on
• the connected Google account, if there is one

The current date and time (UTC) and the remaining turn budget are added to the end of each user turn rather than to the system prompt, so the system prompt stays identical from turn to turn and prompt caching can reuse it. There is no need to tell an agent to look up the date.

Vault file injection

Reference vault files directly in the system prompt using @[path/to/file] — type @ in the editor to pick one:

You are a customer support agent for Acme Corp.

Company knowledge base:
@[knowledge/faq.md]
@[knowledge/pricing.md]
@[knowledge/policies.md]

Each time the agent runs — every chat message, every workflow step — SynapseOrch reads the files and puts their content into the system prompt, so edits take effect on the next message. A file is inlined up to the Character threshold in Settings → General (100,000 characters by default), and one prompt inlines at most twice that in total. A file past either limit, or a PDF or image, is replaced by a line naming it and the tool that can read it. A path that doesn't exist is left as typed. See Vault → What happens to a file that is too big.

Variables

{{ global.key }} in a system prompt is filled in from Variables each time the agent runs — type {{ in the editor to pick one. A system prompt fills in only global references. Shared workflow state is not available here; use a step's prompt template for that.

State references in workflows

Workflow steps have their own Prompt template, separate from the agent's system prompt. Reference shared state there with {{ state.key }}:

You are a writer. Summarise the following research:

{{ state.research_results }}

Write a 500-word article in markdown format.

The older single-brace {state.key} still works for flat keys. See Referencing state in text.

Tips for effective system prompts

Be specific about output format

Always respond with JSON in this exact format:
{
  "summary": "...",
  "confidence": 0.0-1.0,
  "sources": ["url1", "url2"]
}

Define tool usage policy

- Use scrape_url for any question about a specific web page
- Use the vault only to save final outputs — never intermediate steps
- Use run_sql_query only for SELECT queries unless the user explicitly asks to modify data

Name tools as the agent sees them; the names are listed under the agent's Tools. A prompt is guidance, not enforcement: to block writes to a database, leave Database write access off in Settings → General (the default), and to require a person's approval for destructive calls, use Tool Policy.

Set persona and tone

You are a concise technical writer. Avoid filler words.
Prefer bullet points over paragraphs. Use markdown headers.
Always include a "Next steps" section at the end.

Constrain scope

You only answer questions about our software product.
For anything outside that scope, politely redirect the user.

Role and context

You are a senior analyst at a hedge fund. You have access to our market data database.
Always quantify uncertainty — never present uncertain information as fact.

System prompt length

System prompts can be as long as needed, but everything in them — including inlined @[file] content — is sent as input tokens on every turn. Referencing a file keeps it maintained in one place; it does not make it cheaper. For a large set of documents, don't inline them: make the folder a knowledge base, which the agent searches for the passages it needs and cites, or let it read files with the vault tools.

Testing system prompts

Self-hosted, use the AI Prompt Writer in the agent editor — describe the agent and click GENERATE — for a starting point, then refine it by chatting with the agent. Chat shows each tool call and reasoning step so you can see how the agent interprets the prompt.