Turn a repeat task into a reusable AI routine with a review step
7 min read1 quick quiz1 templateSources cited
A reusable AI routine is a written card for one repeat task. It states the goal, the inputs, the steps, the output format, and the checks a person must do before sending. Write the card once. Test it on past work. Then use it every time the task comes back.
The short version
Pick a task you do at least three times a month.
Write three to five success criteria that you can check.
Test the routine on three past examples. Add a human review step before anything reaches a client.
Why a routine beats a fresh prompt each time
A fresh prompt gives a different result each day. You also spend time rewriting the same instructions. A routine fixes both problems. It also makes your work easier to hand over, to check, and to explain to a client.
Anthropic’s prompt engineering overview gives a useful order of work. Before you write a prompt, you need three things: clear success criteria, a way to test against them, and a first draft. The page then lists techniques such as clear instructions, examples, structure, and prompt chaining. Prompt chaining means splitting one job into small steps. We checked this page on October 12, 2026. It does not carry a date. It describes Claude, but the order of work applies to other tools by analogy. We did not test any tool for this article.
Step 1: choose a good task
Quick check
Which task is the best first candidate for a routine?
A weekly status email for one client, built from the same task list each time.
Yes. It repeats, it has a clear input, and you can check the result against the list.
A one-time reply to an angry client about a missed deadline.
Not quite. It happens once and needs your judgment about the relationship. Write it yourself.
A legal contract summary that you send without reading the contract.
Not quite. The risk is high and nobody checks the source. Do not hand this to a routine.
Click an answer to see why.
Step 2: write the routine card
Use one card for each task. Keep it to one page.
| Field | What to write | Example: weekly status email |
|---|---|---|
| Goal | One sentence on the result. | A 120-word email that tells the client what we finished and what blocks us. |
| Inputs | What you give the tool. | The task list export for the week. A sample email the client liked. |
| Steps | Numbered, short. | 1. List finished tasks. 2. List blockers. 3. Write the email in the sample tone. |
| Output format | Length, headings, tone. | Under 150 words. Three short paragraphs. No new facts. |
| Success criteria | Checks a person can do. | Every task in the email appears in the export. No invented numbers. Blockers name an owner. |
| Do not | Limits and data rules. | Do not add tasks. Do not paste client data that the client has not approved for AI tools. |
| Human review | Who checks, and what. | You compare each line with the export. You send the email. The tool never sends. |
Step 3: test it on three past examples
Run the routine on three past weeks that already have a correct answer. Compare each output with the success criteria. Mark each criterion pass or fail. If one criterion fails twice, fix the card. Do not fix the output by hand and move on.
This test takes more time than a quick prompt. It also shows you where the routine fails before a client sees the failure.
Checklist
Routine card checklist
☐ I do this task at least three times a month.
☐ The goal fits in one sentence.
☐ I wrote three to five success criteria that I can check.
☐ I listed what the tool must not do, including client data limits.
☐ I tested the card on three past examples.
☐ A named person (me) reviews the output before it leaves my desk.
☐ I know where I will record failures.
Copy this list into your notes and tick each item off.
Example: where a routine can fail
This example is made up for illustration. It is not a measured result.
A virtual assistant uses the status email routine for six weeks. In week seven, the client changes its task tool. The export now has a new column name. The routine reads the wrong column and lists tasks that are not finished. The review step catches the error because the assistant compares each line with the export. Without that step, the client would receive a wrong report.
Routines drift when your inputs, your tools, or your client’s needs change. A monthly check prevents most surprises.
Limits and tradeoffs
- A routine saves time only on tasks that repeat. Building the card takes real effort.
- A routine does not remove your responsibility. The client hired you, not the tool.
- Tools and features change. Re-test the card after a major tool update.
- Check your client’s rules on AI tools and data before you build the card. Some clients forbid pasting their data into some tools.
Research on knowledge workers adds one more reason for a review step. In a survey of 319 workers, higher trust in AI went with less self-reported critical thinking. The survey is self-reported, so it shows a link and not a cause. See the Microsoft Research listing for the CHI 2025 paper. Our guide on five checks for AI-assisted client work gives a review list you can paste into any card.
Advanced extension: chain the steps and keep a failure log
Split the routine into two passes. In pass one, the tool drafts the output. In pass two, you ask the tool to compare the draft with your success criteria and list any criterion it fails. Then you check the list yourself. Add a failure log with three columns: date, what failed, and the card change you made. Review the log once a month. After three months, you have evidence of your quality process that you can show to clients.
Do this week
Choose one repeat task. Write the routine card in 20 minutes. Test it on one past example before you use it on live work.
Sources
- Anthropic. “Prompt engineering overview.” Claude Platform Docs. No publication date shown. Checked October 12, 2026. platform.claude.com/docs
- Lee, H.-P. et al. “The Impact of Generative AI on Critical Thinking.” CHI 2025, via Microsoft Research. Checked October 12, 2026. microsoft.com/en-us/research
