Your first automation, on one page

Why it is worth fixing

Nobody can call an automation a success if nobody said what success meant. A one-page plan forces the questions that matter before money is spent: is this task worth automating, what exactly happens, what could go wrong, and how will we know it helped? It also makes any conversation with a supplier clearer, because you are describing a job rather than asking to “use AI”.

Step by step

1. Run the three-sign test

   Ask three questions about the task. Is it repeated every day or every week? Do the rules fit on one page? Is a mistake easy to spot and put right? Yes to all three makes a strong first candidate. No to the second means writing the rules down first: software can’t follow rules nobody has written.

2. Not sure which task? Let your own records suggest one

   If you use an AI agent that supports skills, such as Claude Code, Codex or Cursor, the free, open-source Automation Mining skill can read one work source, such as your past agent sessions or, with a Slack connection, a channel you belong to. It returns up to five proposals, each with the records it is based on and a count of how often the task came up. It suggests; it doesn’t build. It is a young labs experiment from a workflow-automation company, tested on made-up companies rather than real teams, and its installer needs Node.js and Git. Tell the channel first, and keep client and candidate details out, because what it reads goes to your AI provider.

3. Write the task as four blocks

   When: what starts it, such as an application arriving. Do: the steps, such as sending a thank-you and adding the applicant to your tracker. Check: the rule it must pass, such as “is a CV attached?”. Hand off: when a person takes over, such as messaging the consultant when the CV is missing. Workflow tools such as Make are built on the same idea: a trigger starts the workflow, and the steps follow.

4. Write the four-line measure

   The job: one task, in one sentence, not “use AI”. The number: what you will count, such as hours a week, errors or days to place. The baseline: where that number is today. The target: where it should be, and when you will check. Without these lines, nobody can say whether it worked.

5. Count the baseline, don’t guess it

   Before anything is built, count the current number for a week or two: how often the task happens, how long each one takes, how many go wrong. A tally on paper is fine. A guessed baseline makes any result look good or bad by accident.

6. Decide what stays with a person

   Keep judgement, exceptions and relationships with people: anything that decides about a candidate, anything a client would expect a person to handle, and anything that doesn’t fit the rules. Write down who receives each hand-off and how quickly they should pick it up.

Checklist

• The task passes all three signs, or its rules are now written down
• The four blocks are written: When, Do, Check, Hand off
• The four lines are written: the job, the number, the baseline, the target
• The baseline was counted, not guessed
• Every hand-off names a person
• Anyone whose messages you mine for ideas knows, and client details stay out

When it is worth getting help

Once this page is filled in, some first automations can be set up yourself with an off-the-shelf workflow tool. It is worth getting help when the task crosses several systems, when one of them offers no direct connection, or when a mistake would reach a client or a candidate.

Rec Int. builds automation for UK recruitment agencies around the systems they already run, using browser automation where a system offers no direct connection.

Questions people ask

Sources

• Automation mining: find repeated manual work (README) GitHub
• Types of modules Make Help Center

If you would like to talk through the page once you have filled it in, book a discovery call.