What you'll be able to do after this guide: explain how a set of scheduled Claude tasks can find, research and prepare prospects for you every day, and describe the job each part of the system does.
Start with a picture of a working day
Imagine a small sales team at a software company. We'll call it Fernhill. Nobody at Fernhill opens a spreadsheet of leads in the morning. Instead, by the start of the working day:
- 15 new companies have been found, checked against the CRM so none are duplicates, researched, scored and written into the CRM.
- Each strong prospect has a named decision-maker, the evidence behind the score, and a drafted first message.
- Anyone who accepted a connection request has been matched to their CRM record, and a follow-up is drafted and waiting for a person to send.
- Anyone who has gone quiet for a week has an email task in the founder's task list, with the draft already inside it.
- Any genuine reply from a prospect has become a deal at the right stage.
No person did any of that by hand. A handful of scheduled Claude tasks did it early in the morning, and a person spends their morning on the part that needs a person: reading, sending and talking.
The rest of this guide explains the pieces that make that morning possible.
The four building blocks
1. Claude scheduled tasks. A scheduled task is a set of written instructions that Claude runs on a timetable you choose, such as every weekday at 07:45. Each run starts fresh and remembers nothing from the last run. That matters, because it means everything the agent needs to know must be written down somewhere it can read.
2. A CRM connected over MCP. MCP (Model Context Protocol) is the standard that lets Claude use other apps through "connectors". With a CRM connector, Claude can search records, create companies and people, write notes and create tasks, just as a person would. We use Attio. Any CRM will work if it offers an MCP connection and lets you create your own custom attributes (fields), because the whole system runs on those fields.
3. Written instructions. Each agent follows a document that spells out exactly how to do its job: where to search, how to score, what to write, what never to do. Good instructions are the difference between an agent you can trust unattended and one you have to watch.
4. A person in the loop. Agents prepare, check and record. People decide and send. Outcome decisions such as "this is now a customer" and anything said to a prospect stay with a person.
One job per agent
The most important design choice is to give each scheduled task one job. A single "do all my prospecting" agent is hard to test, hard to fix and hard to trust. Separate agents, each with a narrow job, can be switched on one at a time and improved independently.
