How To Write A How-To That AI Can Follow

A human reading your how-to fills in the gaps. They infer that “the settings” means the app’s settings, not the operating system’s. They notice a step assumes you already installed something and go do it. An AI agent following the same instructions does none of that. It executes what you wrote, in order, literally, and a gap that a person would step over is where the agent falls down.

That changes what a good how-to looks like. The instructions that read fine to a skimming human are often the ones an agent cannot complete, because they lean on inference the agent will not supply.

State Every Prerequisite Before Step One

The most common failure is a procedure that assumes a starting state it never names. The reader is presumed to have an account, a file open, a permission granted. A human notices the gap and fixes it. An agent starts at step one and hits a wall at step three.

List the prerequisites explicitly, above the steps, as their own short block: what must be installed, signed into, downloaded, or open before the procedure begins. This is exactly the discipline Google’s structured data documentation encodes with tools like supply and tool properties, and the reasoning holds whether or not you add the markup: a procedure is only reproducible if its starting conditions are stated.

Prerequisites first also helps the human, who would rather learn on line one that they need admin rights than discover it halfway through.

One Action Per Numbered Step

Cramming three actions into one step is how procedures become unfollowable. An agent that reads “Open Settings, scroll to Engine, and toggle Local, then confirm the download” has one step to succeed or fail at, when it actually needs four checkpoints. Split them. One verb, one action, one observable result per line.

Number the steps rather than using bullets. Numbering carries order, and order is the whole point of a procedure. Bullets say “these things are related”; numbers say “do this, then this.”

Before:

Head over to the settings and switch things to local, then once that’s sorted you’ll want to grab the model so it’s ready to go.

After:

1. Press Ctrl+Shift+Space to open the overlay.
2. Click the engine toggle and select Local.
3. When prompted, click Download to fetch the local model.
4. Wait for the progress bar to reach 100 percent before continuing.

The second version works because each step is one action with a checkable outcome, and nothing depends on the reader guessing what “that” or “things” refers to.

Kill Every Ambiguous Pronoun

“Click it.” “Open that.” “Once this finishes.” Each of those words points at something the writer had in mind and the reader has to reconstruct. A human usually reconstructs correctly. An agent resolves the pronoun to whatever the retrieval context makes handy, which may not be what you meant.

Replace pronouns with the actual noun, even when it feels repetitive. “Click the Download button,” not “click it.” “When the model download finishes,” not “once that’s done.” Repetition reads slightly clumsy to a human and reads as clarity to a machine, and clarity wins here.

The same rule applies to references across steps. “Repeat the previous step” forces the reader to count backward; “repeat step 3” does not. Anchor references to numbers, not to memory.

Make Each Step Verifiable

A step an agent cannot confirm it completed is a step it cannot recover from. Where you can, end a step with the observable result: what the screen shows, what value appears, what state the system is now in. “The status line reads Ready” gives both a human and an agent a way to know the step worked before moving on.

This is also what makes a how-to survive extraction into an answer. Answer engines and agents pull procedures the same way they pull definitions, in chunks, so a step that only makes sense in the context of the step before it loses meaning when lifted. The same self-contained discipline that helps AI search quote a page helps an agent run it. If you write support content, the crossover is covered in turning support docs into AI answers.

A Wrivio Context for cleaning up a procedure could say:

Rewrite this as a numbered how-to with one action per step. Move all prerequisites into a short list before step one. Replace every pronoun such as it, this, that, and them with the specific noun it refers to. Reference earlier steps by number, not by “previous.” Keep every command, setting name, figure, and value exactly as written. Do not add steps, tools, or requirements that are not in the original.

Press Ctrl+Shift+Space, paste the draft procedure, and check the diff. Watch for the rewrite inventing a plausible step you never wrote, because a procedure with a fabricated step is worse than a terse one.

Structure The Page So The Procedure Is Findable

Put the procedure under a heading phrased the way someone would ask for it, and keep the steps in one contiguous block rather than scattered across prose. A retrieval system matching “how do I enable local mode” wants a heading that says that and a numbered list directly beneath it, not the answer woven through four paragraphs.

The FAQ discipline is the same discipline in miniature, and it is worth reading alongside this one: how to write an FAQ section that gets extracted. Both come down to the same thing. Say the thing plainly, in order, where a machine expects to find it, and stop assuming the reader will fill in what you left out.

Common Questions

Why can an AI agent not follow instructions a person understands fine?

Because a person infers missing prerequisites, resolves vague pronouns, and skips over gaps, while an agent executes what you wrote literally, so any assumption you left implicit becomes a place the agent fails.

Should how-to steps be numbered or bulleted?

Numbered, because numbering encodes order and lets both readers and agents reference an earlier step precisely, while bullets only signal that items are related.

How many actions should one step contain?

One, with a single verb and an observable result, so that success or failure is checkable at each step rather than bundled into a line that half-worked.

Do I need HowTo structured data for this to work?

No, the plain-language discipline of explicit prerequisites, one action per step, and no ambiguous pronouns is what makes a procedure followable, and structured data is an optional enhancement on top of already-clear steps.

Download Wrivio for Windows to rewrite tangled instructions into numbered steps an agent or a colleague can follow without guessing.