Rules an agent can keep
On 6 November an AI agent of mine was told to post a link only "once both pages are live". It checked several times and held back: at first both pages were missing, then for a while the English address was showing the German text. That was a slip in publishing. The agent didn't fix it, but it didn't link to the wrong page either, and the log said why.
A condition you can check is what did the trick. "Make sure everything's right" would have been open to interpretation. My last piece was about budgets and outcomes. This one is about how to write such things down so that an agent can keep to them and a person can check.
An old checklist from staff appraisals does the job: SMART, shorthand that goes back to an article by George T. Doran (Management Review, November 1981). Specific: not "check the invoices" but which documents, which period, which points. Measurable: not "go easy on usage" but at most two sub-agents, and from 80% of the weekly limit only approvals and errors; that is how my own rules read. Achievable: does the agent have the data, tools and permissions? If not, it reports back rather than guessing. Relevant: what is the rule for? If it has no purpose, out it goes. Time-bound: by when, after how many steps a progress report, and an end date on every exception.
Anthropic takes a similar line. On the instruction files for Claude Code, it says the more specific and concise the instructions, the more consistently Claude follows them. They should be concrete enough to verify, and a file should ideally stay under 200 lines (as of 6 November 2026). It also names the limit: such instructions are context, not enforced configuration. A rule in a text file is not a lock. Whatever must never happen also belongs, in my view, in permissions, limits and approvals in the system itself.
A made-up example: "Watch out for mistakes in the invoices" against "Check October's incoming invoices for the mandatory details, list discrepancies by document number, change nothing in the accounts, put unreadable documents on the list rather than guessing, and stop at the end of October". A person then decides whether the rest is worth doing. The second version can be checked: breaches show up, and the rule can be tightened where it needs it.
Which rule for your agent or your team could not be checked today?