Tell the browser agent what finished work looks like.

Describe the outcome, not a brittle click-by-click script.

“Click the third menu, copy the number, then open the next tab” fails as soon as the page moves. “Find the five accounts with an overdue renewal and prepare a table with owner, value, and notice date” tells the agent what matters even if the route changes.

Use concrete objects and counts in the prompt: ten prospects, the Q3 invoice, the three unresolved tickets, or the notes from yesterday's call. The browser agent can choose the sensible path while you retain a result that is easy to inspect.

Give the agent the context it cannot safely guess.

An agent cannot know whether “the important customers” means annual-contract accounts, active trials, or the people who complained last week. It also cannot infer your house style or the commercial rule behind a discount without being told.

Point to the sources that settle those questions and state the rule in plain language. If the information is in an open tab, a connected app, or a document, say which one should guide the work and what to do when the evidence conflicts.

Specify the output that lets you judge the work.

“Help me prepare for this meeting” could produce a paragraph, a list of questions, or a full account plan. When the output format is left open, even good research can arrive in a form that is hard to use.

Ask for a table, a one-page brief, a draft email, a checklist, or a decision memo, and name the fields that must be present. This also gives the agent a place to put uncertainty rather than smoothing it away.

Put the approval boundary in the prompt before the work starts.

Reading and organizing information is different from sending a customer message, publishing a post, or changing a billing record. If the last action affects another person, say whether you want a draft, a prepared change, or a request for confirmation.

Once a prompt produces reliable work, save it as a named skill with its inputs and review rule. A routine can then run the preparation on a schedule, while consequential actions remain under the approval boundary you set.

Frequently asked questions