Directions are the natural-language instructions that tell the AI agent what a step should do and how to do it. Every workflow step has them, and they are the single biggest lever you have over how reliably that step behaves. This guide is about writing them well.
Directions are one part of a step's AI Instructions, alongside input and output properties, tool configuration, and human-in-the-loop settings. Read that article for what each part is for. This one is a how-to for the Directions field itself.
Start simple, then make it specific
You do not have to write perfect Directions in one pass. A common and effective pattern is to start with a plain description of the task, test the step on real work items, and tighten the Directions wherever the agent went off course.
The biggest single improvement is usually specificity. The more precisely you describe the task — and the method you want used — the more reliably the agent follows it. Compare a few ways of directing a step that needs to find a coffee shop's opening hours:
"Find the opening hours for the shop."
"Find the opening hours for the shop. Check the operating-procedure document to see how we look these up."
"Do a web search to find the opening hours for the shop."
"Use a Google Places search to find the opening hours for the shop."
Each is valid, but they are not equally reliable. The vaguest version leaves the agent to choose a method; the most specific one names the method and removes that choice. When a step matters, say what to do and how you want it done.
Write ordered bullets, not a paragraph
Break the work into a numbered or bulleted list of steps rather than one dense paragraph. A bulleted list reads as a sequence the agent should follow in order, which makes multi-step logic far more reliable than the same instructions buried in prose. Put one instruction per bullet, in the order you want them carried out.
Say what to do when something is missing
Reliable Directions cover the awkward cases, not just the happy path. State what the agent should do when a value is absent, a document is incomplete, or the inputs disagree — for example, "if the invoice has no purchase-order number, record it as unmatched rather than guessing one." Being explicit here prevents the agent from filling a gap with a confident but wrong answer.
Point at the material the step needs
Directions do not have to be only text. Where it helps, reference the specific file or example the step should use — for instance, "map each expense to a category using the accounting-codes spreadsheet linked here." Linking the exact reference material is more reliable than describing it and hoping the agent finds it.
You can also add positive and negative Examples to a step's AI Instructions. Examples are the right tool when a pattern is easier to show than to describe — a sample input file, an annotated document, or the exact output format you want. Use them alongside Directions rather than trying to describe a complex format in prose.
Let the step's structure carry the load
Good Directions are shorter than you expect, because much of what makes a step reliable lives in the rest of the AI Instructions, not in the prose:
Input properties focus the agent on the specific data it should use, and their names, descriptions, and types tell the agent what that data means. Choose them deliberately instead of restating the data inside the Directions.
Output properties tell the agent exactly what it is expected to produce. A well-named, well-typed output does more to steer a result than a sentence asking for it.
Tools define what the step can actually do. Disabling tools a step should not use, and constraining the ones it should, removes choices that would otherwise let the agent wander. You can even make a tool available only when the work item matches a condition — see Control which tools a step can use.
Keep the Directions about the how, and let input properties, output properties, and tool configuration constrain the what. For the reasoning behind this, see AI reliability mechanisms.
Keep the message logic separate
If your step only does internal work, do not use its Directions to describe messages to a person or an end user. Whether a chat step replies at all is controlled by its output properties and the conversation's settings, not by asking for a reply in the prose. See AI Instructions for how a step sends a message on a chat workflow.
The same skill applies to custom AI tools
A custom AI tool is implemented with a natural-language instruction written the same way you write step Directions: be specific, spell out edge cases, and describe what to produce. The tool's name, description, inputs, and outputs play the same structural role there that a step's properties and tools play for a step — so the same "let the structure carry the load" advice applies.
A quick checklist
Before you move on from a step, check that its Directions:
name the method you want used, not just the goal, wherever the method matters;
read as an ordered list of steps rather than one paragraph;
say what to do when a value is missing or the inputs conflict;
link the exact file or example the step should rely on; and
leave the data and results to input and output properties rather than repeating them in prose.