Skip to main content
Instructions tell the agent what success looks like. Good instructions are short, concrete, and written like you would brief a teammate. Good vs vague:
  • Good: “Investigate the error, propose a fix, and report back in Slack.”
  • Vague: “Look into this and do your best.”

​
What great instructions include

  • The outcome you want (example: identify the root cause and propose a fix)
  • Constraints (example: avoid risky changes, do not touch production data)
  • Where to report results (example: reply in Slack or comment in the ticket)

​
A simple template

1

Define the goal

Describe the result you want in one sentence.
2

Add constraints

Call out what the agent should avoid or prioritize.
3

Specify the output

Say where and how you want the result delivered.
If the trigger already has rich context (like a Sentry issue or a Linear ticket), keep the instructions short and let the context do the rest.