Put an AI agent on the follow ups you never get around to.

It never sends anything. Ever.

The first fear anyone has about letting AI near their leads is that it will email a client something embarrassing. So that is written into the agent itself, not left to hope.

NEVER:

- Send anything, on any channel, for any reason.
- Contact a client directly.
- State, imply, or negotiate a price.
- Edit the pipeline file or mark anything closed or dead.

WHEN YOU ARE UNSURE:

If there is not enough real material to write something specific,
do not manufacture warmth. Write "insufficient context" and tell me
what is missing.

That is a fragment. The full instructions run about forty lines and are written out completely in the PDF, ready to paste.

Not tips. A whole working agent.

• The full specification. All six fields filled in: what it does, what it reads, what it produces, what good looks like, what it must never do, and when it runs.
• The instructions you paste. Written out completely, ready to drop into a Claude Cowork project. Two copy and pastes and it exists.
• A free plan version. The same agent run by hand in ordinary Claude chat, for anyone not on a paid subscription. Fifteen minutes a week, same output.
• A first week plan. Three supervised runs, what will go wrong on the first one, and the rule for when it has earned the right to run on a schedule.
• The two rules that matter. Including the one that keeps an agent from doing something on your behalf that you cannot undo.