Only the operator's own words instruct

The idea

Only a verified sender's own lines are instructions; forwarded, quoted, fetched, partner and guest-command text is data, hidden characters are stripped, and actions after instruction-shaped content wait for a person.

The problem
Outside text could steer workers into harmful actions.
Why it matters
Outside content cannot approve, instruct or leak secrets through replies.
Where it stands
In production as of

A step-by-step replay of this idea is not published yet. The library adds one when a recorded run shows it.

  • Business-only WhatsApp guards

    The crew runs only on the customer's own number, answers clear consumer-assistant use with one scoped reply, discloses it is AI, and keeps WhatsApp content out of other orgs and shared outputs.

  • Work between organizations

    Two orgs connect only after both owners accept; one gateway employee per side sends and receives typed partner requests, partner text is data, cost policy applies and either side can revoke.

  • Worker's own browser

    The worker drives a real browser on its own computer through governed tools and can check its own app on that computer; the platform treats every page as outside content, and admins see the worker's screenshots.

  • One governed tool gateway

    Every action goes through one gateway. It checks authorization, policy, risk, budget and approval, writes the action ledger first, and reads back each external write. A model answer can only add a requirement.

  • Gate on all outgoing text

    Commits, PR text, comments, messages and reports go through rule-based checks (length, structure, banned phrases, attribution, secrets) and a style check against examples a person approved. Failing text gets a limited number of rewrites.

Turn the numbers into shipped work.

Crewshal runs these choices for you: a persistent AI worker with memory and rules, on your Claude and Codex subscriptions.