When I first used an agent, I would type, “help me fix this.” It did something, but often not what I had in mind. The problem was obvious in hindsight. I had not said what to change, what to preserve, or how I would judge the result.
Working with an agent can feel like onboarding someone new to a team. Ability matters, but so does the handover.
Describe the job before assigning it
I now tell the agent where the source material is, which files are off limits, and what a finished result should look like. For risky actions, it must ask first. Writing this down also exposes the gaps in my own thinking.
Vague instructions produce vague work. An agent may guess correctly, but I do not want a workflow that depends on luck.
Memory should be visible
Repeating the same background every day is tiring. I prefer memory stored in files: project status, working preferences, and mistakes we have already found. I can inspect it and correct it.
“It understands me better” often means something simple. Yesterday was written down.
Permissions should arrive slowly
I usually let an agent read and suggest before I let it execute. Reversible work can be looser. Accounts, publishing, and deletion need more care.
This is not unusual. A new intern does not receive every company key on the first day either.
The responsibility stays with the person
An agent can research, draft, edit, and run a process. A person still decides why the task exists and what result is acceptable. My biggest change has not been speed. I now write better rules, keep records, and define checks before the work starts.
The agent did not remove the need to think. It made vague thinking harder to hide.