Guide
How to write an AI agent work order: goal, checks, boundaries
Last updated August 12, 2026
A prompt asks an agent to do something. A work order tells it what done means, how to prove it, and what to leave alone. Same input, very different outcome. Here are the four parts, in the order that matters.
1. The goal names a finished state
The goal is one sentence describing the world after the job is done, not the work itself. "Implement the export feature" describes activity. "Export produces a CSV with one row per order that opens in Excel" describes a finished state. If you can attach "...and I will know because" to the goal, you are close.
2. Acceptance criteria are checkable by a stranger
Every criterion should be provable by someone who has never met you and is mildly hostile. If a stranger cannot grade it, neither can a verifier, and the agent can report SUCCESS while being technically correct. Prefer Auto checks (provable by command, URL, file, or test) over "feels right" checks. See the full guide on acceptance criteria.
3. Boundaries are half the contract
Non-goals are acceptance criteria in disguise. Without them the agent will refactor, upgrade, and "improve" its way past your deadline. Write what it must not touch: no dependency upgrades, no redesign, no production deploy, no emails to real customers. Short list, plain words.
4. Budget and preauthorization
Set how many attempts the agent gets before it must stop, and list any outward or irreversible actions it is allowed to take, word for word. Leave the risky list empty unless you mean it. A contract with no budget can loop forever; a contract with no preauthorized list cannot ship anything.
Before and after
Before (a prompt)
Make me a weekly sales recap and put it somewhere my team can see it.
After (a work order)
Goal: A short weekly recap of my sales calls lands in a shared doc my team can open. Checks: - Auto: shared doc link resolves for the team - Auto: one section per call with a next-step verdict - Auto: no customer names or numbers in the doc Boundaries: no CRM changes, no emails to customers, no pricing discussion Budget: 20 attempts
Start from a ready-made work order template, let Grok draft one from plain English, then run the $1 audit before anything expensive. And when it runs, verify the work instead of trusting the report.
Put your next agent task through the press
Talk to Grok, get a tight work order, and hit it with a $1 audit before anything expensive runs.