documentation
04 admin

Global agent rules

Global rules are the deployment-wide constraints every agent sees in its system prompt. Use them for the principles that shouldn't vary across specialists — privacy, brand voice, escalation rules — and layer per-agent rules on top for the specifics.

What this page is for

Every agent’s system prompt is built from four layers: the agent’s personality (written by the agent’s creator), any skills the agent is using, any agent-specific rules, and the global rules from this page. Global rules apply to every agent. They’re the place for deployment-wide principles that shouldn’t be restated in every personality.

Typical things to put in global rules:

  • “NEVER share customer personally identifiable information outside approved tools.”
  • “ALWAYS cite sources when summarizing external web content.”
  • “AVOID speculating about legal, medical, or financial advice — refer users to a professional.”
  • “PREFER using the data_store tool over pasting long results into chat.”

Global rules sit on top of individual agent personalities. They’re the floor of behavior, not the ceiling — an agent’s own personality can go further, but it can’t negate a global rule.

Rule categories

Each rule has a category that controls how it’s phrased in the system prompt:

CategoryMeaningExample
ALWAYSA rule the agent must follow”ALWAYS verify the tool’s result before reporting success.”
NEVERAn explicit prohibition”NEVER send emails to unverified addresses.”
PREFERA soft steer toward a behavior”PREFER markdown over plain text for multi-paragraph responses.”
AVOIDA soft steer away from a behavior”AVOID speculating about current events without a recent search.”

ALWAYS and NEVER read as hard rules to the agent. PREFER and AVOID read as preferences — the agent will follow them unless overridden by something stronger. Use the soft forms for style or taste; use the hard forms for safety and compliance.

Common tasks

Add a global rule

  1. Open /admin/rules
  2. Click + Add Global Rule
  3. Pick the category
  4. Write the description (one or two sentences — short rules work better than essays)
  5. Optionally add an example
  6. Save

The rule appears immediately in every agent’s next turn. You don’t need to restart anything.

Preview how rules look in the prompt

Click Preview Prompt at the top of the page. A block of text opens showing how the rules compile into the system prompt. Use this to sanity-check that your rules read clearly in context — rules that look fine in the edit form sometimes read awkwardly when stacked.

Edit or delete a rule

Each rule row has Edit and Delete buttons. Edits go live on the next agent turn. Deleted rules are gone — there’s no undo, so take a screenshot if you might want it back.

Reorder rules

Rules are listed in the order they’ll appear in the prompt. Drag the grip handle on the left to reorder. Order matters for how the agent reads them; keep important rules (safety, compliance) at the top.

Agent-specific rules

Each agent can layer its own rules on top of the global ones from its detail page. The effective rule set for an agent is the global rules plus the agent’s own. There’s no subtraction — an agent can’t remove a global rule locally.

Use agent-specific rules for behavior that only applies to that specialist. A research agent might have a rule “ALWAYS verify facts with at least two sources”. A code assistant might have “PREFER idiomatic code over clever one-liners”. These don’t need to be global because they’re specific.

Common pitfalls

Writing too many rules. A system prompt with 30 global rules becomes a wall of text the agent skims. Keep the global list short and focused on the things that matter most. Agent-level rules exist for the specifics.

Writing rules as essays. “ALWAYS verify the tool’s result before reporting success” works. “The agent should take care to verify the result of any tool call it makes before claiming the work is done, as per the guidelines in our quality document” doesn’t. Short and direct beats long and thorough.

Using PREFER/AVOID for safety. A soft steer is not a guardrail. If it’s a safety rule, use ALWAYS or NEVER so the agent treats it as binding.

Forgetting about the approval gate. Rules nudge behavior. Approval rules block behavior. For actions that must not happen without human review (destructive shell commands, external emails), add an approval rule in addition to a global rule.

Where to go next