What you'll be able to do after this guide: create a scheduled task in Claude, choose its timing and permissions, test it safely, and switch it on.
In the Claude app, choose Scheduled in the sidebar. Click New task. You'll see two options:
Create with Claude: Claude asks you a few questions and drafts the task for you.
Set up manually: you fill in the form yourself.
For a prospecting agent, choose Set up manually, because you already have a prompt from Write the Spec and the Task Prompt.
Name. Start with the stage number, for example "1 - Source & Research (Segment A)". Numbered names keep the list of tasks in pipeline order and make it obvious what runs before what.
Instructions. Paste your scheduled task prompt.
Work in a project. If your specs live in a Claude project, choose it here so the task can read them. If they live in a folder on your computer, the Claude desktop app needs to be open for the task to reach them.
Model. The default is fine to start with. Research-heavy agents benefit from a more capable model.
Frequency. Options include Manual, Hourly, Daily, Weekdays, Weekly and Monthly, with a time. Sourcing agents usually run once each morning: Daily if the spec gives them weekend work such as a catch-up, otherwise Weekdays (as in the example above). A backfill agent can run hourly.
Permissions. Choose how much the task can do without asking.
Manual: Claude asks before using new tools or opening new sites. Use this for the first supervised runs.
Auto: runs without asking. This is the usual setting once a task is proven, because nobody is there to approve an unattended run.
Skip all approvals: skips the task's own approval prompts, even for actions Claude would normally check with you. We don't recommend it for anything that touches prospects.
Your connector tool permissions still apply on top of this. A CRM tool you blocked stays blocked whatever the task setting.
Advanced settings: Notify me when this task finishes. Turn this on during the first few weeks so each run's summary reaches you.
Agents depend on each other's output, so order matters more than the exact minute.
| Time (example) | Agent | Why then |
|---|---|---|
| 07:12 daily | 1 - Source & Research (Segment B) | Fresh prospects ready before the working day |
| 07:20 weekdays | 4 - Acceptance check | Must run before the message and email stages |
| 07:47 daily | 1 - Source & Research (Segment A) | Staggered so two sourcing agents aren't writing to the CRM at the same moment |
| 08:35 weekdays | 6 - Email task | After the acceptance check, so recent acceptances don't get an unnecessary email |
| Hourly | 2 - Backfill | Small batches, all day |
Three timing tips:
Avoid exactly on the hour. Many scheduled jobs start at :00, so runs at odd minutes such as :47 tend to start on time.
Put working hours in the schedule, not the prompt. An instruction like "stop if it's outside office hours" still costs a run every time it fires. Choose a weekday schedule instead.
Check the live schedule, not your notes. Over time, what a document says a task does and what it is actually set to do drift apart. The scheduler is the truth.
Don't switch on a new agent straight away.
Dry run in chat. Paste the prompt into a normal chat, add "This is a test. Research only 3 companies, and show me what you would write to Attio instead of writing it." Check the reasoning.
Supervised run. Save the task with Frequency set to Manual and Permissions set to Manual, and run it. Watch which tools it uses. Check every record it wrote.
Compare with your own research. Pick one company it researched and research it yourself. Every difference is a rule to add to the spec.
Schedule it. Switch to your chosen frequency and Auto permissions.
Read every summary for the first two weeks. Expect to update the spec once or twice in the first week as real cases appear.
Each run starts fresh. It doesn't remember yesterday. Anything it should know must be in the spec, the CRM or a log.
Each run reads files as they are now. That's why specs are re-read every time, and why an agent must never rewrite a shared file from a copy it read earlier.
Browser and local files need your computer. If a task uses Claude in Chrome or reads files on your computer, the Claude desktop app must be open and the computer awake when it runs.
Working files need unique names. If agents save working files, have each use its own file names, so one task never picks up another's leftovers.
To replace a task, create the new one and pause the old one in the same sitting. Leaving both running means two agents writing the same fields with different rules, which is hard to spot later.
Next: Guardrails Every Sourcing Agent Needs.
