Common prompting mistakes
Every one of these is common, and every one is fixable in a sentence.
Asking for too much at once
> Refactor the auth module, add tests, update the docs, and fix the lint errors.
The agent will do some of it, do part of the rest, and report success. You then have to work out which parts actually happened.
Split it. One task, one verification, then the next. This also keeps the context small, which keeps quality up.
Not saying what "done" means
> Make the tests better.
Better how? More coverage, faster, clearer names, fewer mocks? The agent picks one, probably not yours.
State the finish condition: "add cases for empty input and for the timeout path, and make sure npm test passes."
Assuming it remembers
A new session knows nothing about the last one. Even within a session, older turns may have been trimmed to fit the context window.
If something matters across sessions, it belongs in a file the agent reads, not in a message you sent yesterday.
Correcting without removing the cause
You tell it to stop adding comments. Three turns later it adds comments again.
The original instruction is still in the context, competing with your correction. Editing the instruction beats piling corrections on top of it.
Trusting confident output on unverifiable claims
An agent will tell you a library has a function it does not have, in the same tone it uses for things that are true. This is most dangerous with version specific APIs and with anything released after the model's training cutoff.
For library specifics, check the documentation, or use a documentation tool such as the Context7 MCP server, which puts current docs in the context instead of relying on memory.
Letting the session run too long
Quality degrades as context fills. If a session has gone twenty turns and the agent is getting worse, that is expected behaviour, not bad luck.
Start a fresh session and bring only what matters. This costs less and works better than pushing through.
Pasting an error with no context
> Fix this: TypeError: cannot read property id of undefined
The agent does not know which file, which recent change, or what you expected. It guesses, and it guesses wrong.
Include the file, the command you ran, and what you changed just before it broke.
Reviewing only the diff
The diff shows what changed, not what should have changed and did not. A change that passes review and silently skips the error path looks perfect in a diff.