Build your first repeatable AI workflow

A workflow, not an agent

If you ask an AI for the same kind of thing every week, you're probably rewriting the same request each time and getting slightly different results. A repeatable workflow fixes that: the same steps, the same instructions and the same checks every time, saved so you can reuse them.

That's different from an autonomous agent, which decides its own steps and can take actions such as sending emails or editing files. Agents are useful for some jobs, but they're harder to check. For a weekly task, a workflow where a person runs it and approves the result is usually the better place to start.

Pick the right task

A good first workflow passes five tests.

1. 1It repeatsWeekly or more often. Anything rarer isn't worth the setup.
2. 2The input looks similar each timeMeeting notes, enquiry emails, a weekly report. If every input is different, the instructions can't be.
3. 3You know what good looks likeYou could judge a result in a minute or two. If you can't describe a good result, an AI can't aim at one.
4. 4Mistakes are cheap to catchYou'll check every output before it's used, and you can tell when something is wrong.
5. 5The data is OK to shareYou're allowed to paste it into the tool you'll use. If your organisation has rules about AI tools, they apply here.

Our worked example is a task most people have: turning messy meeting notes into a short summary and a table of actions with an owner and a due date, ready to paste into an email.

Break it into steps

Write down what you do today by hand, then mark which steps the AI drafts and which a person checks. Keep the AI's part small enough that you can see whether it did it right.

1. 1Prepare the input (you)Paste the notes as text. If they're in a PDF or Word file, convert them first; a converter such as MarkItDown turns most office files into plain text.
2. 2Draft the summary and actions (AI)One run, with the saved instructions.
3. 3Check every row (you)Each action against the notes, with the short checklist below.
4. 4Send or file it (you)Only after the check. The AI never sends anything.

Choose a tool

Any AI chat tool can run this example, including free plans. If your tool lets you save standing instructions, use that so you don't have to paste them each time. If it doesn't, keep them in a text file and paste them at the start of a new chat. A new chat each time also stops last week's notes leaking into this week's answer.

Note which tool and model you used. Providers retire older models on published dates, and a workflow that worked on one model needs re-testing on its replacement. The current options are below, with dated sources.

Write instructions that hold up

A workflow's instructions are a brief you write once: the goal, the context, the rules and the exact output format. The rules that matter most are the ones that stop the AI filling gaps with guesses, and the ones that make its output quick to check.

Three rules are doing the work. "Never guess" with a fixed placeholder turns a hidden gap into a visible one. The "From the notes" column lets you check each row in seconds. And the "to confirm" list gives the AI somewhere to put its doubts instead of resolving them quietly.

Context, documents and structured output

• Give it only what this run needs: this week's notes, not the whole folder. Anthropic's documentation notes that accuracy and recall can fall as a model's context fills up.
• Paste text rather than screenshots where you can. Converted text is easier for you and the model to check.
• Ask for the same format every time. A fixed table is faster to check, easy to paste, and makes a missing column obvious.
• If one input is very long, split it into parts and run the workflow on each.

Check the output

Every run gets a quick human check before anything leaves your hands. For this example it takes about two minutes.

• Every action matches its "From the notes" quote, and the quote is really in the notes.
• No owner, name or date appears that isn't in the notes.
• Nothing important is missing: skim the notes for anything that sounded like a task.
• Anything changed later in the meeting shows its final version.
• The "to confirm" items are either resolved or sent on as questions.

If a workflow produces factual claims, such as figures in a report or statements about rules, check those the way our guide to checking an AI answer describes.

Test it on three real examples

Before you rely on a workflow, run it on three past inputs where you already know the right answer: one easy, one messy and one tricky. We did this for the example with three meeting notes we wrote, comparing the one-line request with the standing instructions.

The one-line request wasn't bad. Its mistakes were plausible guesses, which are the hardest kind to notice. The instructions didn't make the model smarter; they made its gaps visible. They weren't perfect either: one summary left out two items that appeared only in the action table, which is why the human check reads the table, not just the summary.

When it fails, fix the instructions

When a run goes wrong, correcting the output fixes this week. Changing the instructions fixes every week after. Note each failure in a test log, change one rule, and re-run your three examples.

Save it so it works next week

A workflow only saves time if you can find it and trust it next week. Keep it on a workflow card: one text file with the task, the steps, the instructions, the output format, the checklist, your test log and a version line. When you change a rule, add a line saying what and why. When your tool or model changes, re-run your three test examples before you rely on it again.

Once a workflow is proved by hand, you can go further: package it as a reusable skill in tools that support them, or connect it to an automation tool so the input arrives on its own. Keep the human check either way.

Try it

What to remember

• A workflow makes results more consistent and easier to check. It doesn't make them right; the check does.
• Our test used three examples on one model. Yours will behave differently, which is why you test on your own examples.
• Don't paste confidential or personal information into a tool your organisation hasn't approved for it.