Describe outcomes, not steps
Describe the result you want and the context around it. The agent plans the steps itself.Give it context
The agent knows what you tell it plus what it remembers. Include the things only you know:- Who the work is for and how it will be used
- Constraints: deadline, tone, format, must-include, must-avoid
- Where to find inputs (“the spec is in the file I just dropped in”)
Answer its questions
The agent asks before it guesses on anything consequential: a destructive action, a spend, an ambiguous requirement. Answer the question. The result is better than if it had assumed.Work in passes
Treat the first output as a draft, then refine it:Hand off recurring work
Recurring and tedious work is where agents help most:- “Every weekday at 9am, summarize what changed in our shared inbox.” → a routine
- “Whenever I forward you a contract, pull out the key dates and obligations.”
- “Keep our launch checklist current as we close items in Slack.”
Todos
On longer work, the agent keeps a live plan in the conversation. You see it as a checklist: what it is doing, what is next, what is done. You do not fill it in. You can glance at it to see where things stand. It is not a project tracker. It exists for the current run and disappears when the work is finished. Recurring work belongs in a routine.Teach it once
When you correct the agent or state a durable preference, it writes that to memory. You don’t repeat yourself next time, and your teammates get the same behavior.Attach files and images
Drop files into the conversation: documents, spreadsheets, screenshots, audio. The agent reads, analyzes, transcribes, and transforms them, then returns files you can download.Shape its default behavior
Set a persona and standing instructions once instead of repeating preferences in every chat.