How I work

Evidence before claims. An agent may not say “done” or “fixed” without a check it really ran. It names the check. A confident sentence is not proof.

The doer is not the checker. For large or risky work, a separate auditor reviews the result and returns pass or fail. Workers never approve their own work.

Right-size the process. Small, reversible jobs get one run and one check. Full planning and audit are kept for work that is big, risky or hard to verify.

A human approves risky actions. Restarts, deletions, and anything irreversible, paid or public wait for my explicit OK. Agents can prepare the change. They do not press the button.

A repeat problem is not fixed without a safeguard. When the same failure comes back, I stop patching. I look for the root cause and add a check that would catch it next time.

Privacy by default. Credentials, chats and personal files are kept out of public work. Files are scanned for secrets before they enter the shared memory store.

Write each fact down once. One fact lives in one place. Other files link to it. Copies drift, and a stale copy misleads the next reader.