Skip to content

Re-run a skill on new inputs

Tool: Cowork
Source status: Author from scratch (short)

A reusable skill gets more valuable when you run it again: on a new file, a new meeting, a new email, a new week, or a new trigger.

Steps

  1. Start a new Cowork task.
  2. Use the same trigger phrase you put in the skill description, such as Run the weekly briefing skill or Use my launch-review skill on this file.
  3. Point Cowork at the new input: attach a file, paste a link, name a meeting, name a sender, or give a time window.
  4. Ask Cowork to keep the same output contract: Use the same sections and field names as last time.
  5. Compare the new output with the previous run. Look for what changed, what stayed stable, and what needs a rule update.
  6. If the task should run on a schedule, describe it in chat: Send me a daily briefing every morning at 9 AM.
  7. Manage scheduled prompts on CoworkScheduled. The page has Runs and Manage schedules tabs. You can have up to 5 scheduled prompts.
  8. If the task should react to work, use an event-driven task. Cowork can trigger on incoming email or a Teams message/@mention.
  9. When Cowork proposes a Set up trigger? card, review When, Run in, What it does, and the permissions needed.
  10. Keep the default draft-and-approve flow for actions that send messages. Cowork prepares the action and asks before sending.

You'll know it worked when... the same skill runs on fresh input without rewriting the instructions, and any scheduled or event-driven run still asks before sending externally visible messages.

Autonomy has guardrails

For event-driven tasks, Cowork's default is draft-and-approve. It can prepare an email or Teams response, but nothing sends without approval unless you explicitly configure otherwise.

Apply it to your scenario

Choose one moment where your scenario should repeat. It might be every morning, before every meeting, when a customer email arrives, or when someone @mentions you. Re-run the skill manually first. Then arm one schedule or trigger only after the output is reliable.