Docs
Docs
Write policies

Writing effective policies

A good policy says which emails, what your agent may do with them, and what stays off-limits. One rule per line, each starting with Allow, Ask me before or Deny. If you'd rather not start from a blank page, pick a recipe and adjust it.

Say which emails

Every allow needs a scope: a label, a sender, a domain or a date. Without one, the review stops.

  • Good: "Allow reading emails from my daughter's school at stmarys.edu."
  • Bad: "Let my agent read my emails."

Name the action

Say read, archive, draft, send, reply or forward. Clam can't guess what "handle" means.

  • Good: "Allow reading work email and creating drafts. Ask me before it sends anything."
  • Bad: "Handle my work emails."

Block sensitive labels in writing

Anything you don't allow is already blocked, but a written block holds when you widen the policy later.

  • Good: "Never access emails about my finances, health, or passwords."
  • Bad: "Stay away from my private stuff."

Ask before anything leaves your mailbox

Reading is easy to undo. Sending isn't. Let the routine case through and ask about the rest.

  • Good: "Allow sending receipts to my accountant at billing@acme.com. Ask me before it sends to anyone else."
  • Bad: "Allow sending replies to work email."

How an approval reaches you: Approvals.

Give Clam something it can check

Clam matches on addresses, domains, labels and dates, not on names or the text of an email. Write the address, or tell Clam who a person is once under Tell Clam what only you know on your mailbox's Settings page.

  • Good: "Allow forwarding receipts to billing@acme.com."
  • Bad: "Allow forwarding receipts to my accountant."

Likewise "newsletters" is a label and "emails that mention discounts" is text, and a date is "before 1 January 2024", not "before 1/1/24".

Before you save

Start with the smallest policy that gets today's job done; widening a rule later is easy, narrowing one isn't.

Click Review and check that each action shows ALLOW, DENY or ASK as you meant, with no warning left. Then Save & activate and try it on your agent: ask for something allowed, then something forbidden, and watch your rules hold.