Run Your First AI Partner Workflow with One Recurring Task

The goal of your first workflow is not “never look at it again.” It is to swap in a fresh input and still receive something you can check in minutes. A small workflow card is enough to make that repeatable.

1. Choose a recurring task small enough to learn from

Look at the last two weeks and pick one task that meets all four conditions:

• It appears at least twice a week and has a reasonably stable input source.
• The AI partner can produce a draft, checklist, classification, or options first.
• You can judge the result in a few minutes.
• A mistake can be sent back for revision instead of paying, deleting, broadcasting, or shipping.

2. Write a workflow card

Do not begin with a long prompt. Define five positions first:

1. Trigger: when the work starts, such as Friday afternoon or when a new feedback batch arrives.
2. Input: where the material comes from: a folder, conversation, or meeting note.
3. AI preparation: a reviewable output, such as three feedback groups plus open questions.
4. Your confirmation: decisions that stay with you: priority, promises, price, and whether to send.
5. Delivery: where approved work goes next: a todo list, document draft, or next-input tray.

3. Make the smallest setup in Agent Workspace

1. Open Agent Workspace and claim or choose the AI partner closest to the task.
2. Put the card's input, AI preparation, and human confirmation into its role or the task. Tell it to list missing material instead of guessing.
3. If the input comes from Feishu or another channel, connect only that channel to this partner.
4. Prepare one real sample whose correct outcome you already understand.

4. Run the first round manually and check four things

• Complete: did it miss anything important in the raw material?
• Traceable: can each conclusion point back to a source instead of filling gaps?
• Decision-ready: is the result shaped so you can compare, edit, or approve it?
• Within bounds: did actions that need your judgment stop at the review point?

Turn your corrections into a fixed checklist. “Remember next time” inside one conversation is not a repeatable rule.

5. Use fresh input for two more rounds

Use the checklist to correct format and omissions in round two. In round three, switch to fresh real material so you know the workflow did not merely memorize the first sample. After all three rounds can be checked quickly, add only one thing: a fixed trigger, automatic input gathering, or delivery after approval.

Expand one box at a time. If the new action fails, you should be able to return to the previous working loop.

Troubleshooting

• The format changes every time: provide an explicit output shape and say that missing values stay blank.
• It invents missing facts: require a source for each conclusion and move unsupported items into “needs review.”
• Review takes longer than doing it yourself: shrink the task to one classification, one draft, or one checklist.
• Results look good but still feel risky: keep human confirmation, then add run history and a clear rollback path.

Data boundary

Work data stays on your device by default; only channels you connect and model calls go online as the task requires. Connect only what this workflow needs and provide only the material required for the current task.