Prompt Anti-Patterns That Destroy AI Code Quality

Most bad AI code is bad-prompt code. The model is doing exactly what you asked, just for a slightly different question than the one in your head. Six anti-patterns produce most of the damage.

Anti-pattern 1 — The kitchen sink

Asking for ten requirements in one prompt. The model can hold five well, eight badly, ten not at all. By the time it gets to your last bullet it has dropped your third one. Result: code that looks complete but missed half the requirements.

# Bad: kitchen sink
"Build a user signup endpoint with email validation, password hashing,
  rate limiting, CAPTCHA, JWT issuance, refresh tokens, audit logging,
  tracing, error handling, and unit tests."

# Better: chunked
"Build a user signup endpoint that accepts {email, password},
  validates email format, hashes password with argon2, and returns 201.
  Nothing else yet — no rate limit, no JWT, no logging."

# Then in turn 2: "Now add JWT issuance." Etc.

Anti-pattern 2 — Leading the witness

"Why is my function broken?" assumes the function is broken. The model will find a reason even if there isn't one. Now you have a fix for a non-bug, on top of the real bug.

# Bad: leading
"Why is my login function returning the wrong user?"

# Better: neutral
"My login function returns user A when I pass user B's credentials.
  Here is the function. Here is the test that exposes the bug.
  Without changing code, list the most likely root causes."

Anti-pattern 3 — The vague verb

Anti-pattern 4 — Pretend the previous turn didn't happen

Models trust your latest message more than the file. If you say "that's wrong, try again" without saying what's wrong, the model will rewrite the working parts and break them, while keeping the broken part. Always preserve what works in your correction.

# Bad: blanket rejection
"That's wrong. Try again."

# Better: scope the correction
"The validation logic in lines 12-18 is correct, keep it.
  The error handler at line 24 catches everything as 500 — that's wrong.
  Validation errors should return 400, auth errors should return 401.
  Only change line 24."

Anti-pattern 5 — Asking for code without examples

Almost every confusion in code generation can be killed by one input/output example. "Parse this CSV" is ambiguous. "Parse this CSV — here is a 3-line sample, here is the JSON I want out" is unambiguous. The example does the work that ten paragraphs of constraints would.

Anti-pattern 6 — Ignoring the file you're editing

Asking for code as if the project doesn't exist. The model invents an architecture, picks libraries you don't use, and follows conventions that fight your codebase. Cursor, Claude Code, and Copilot all index the project — but only when you point them at it.

# Bad: detached
"Write a React component that fetches users."

# Better: anchored
"In src/components/UserList.tsx, write a React component
  that follows the conventions in src/components/PostList.tsx.
  Use the existing hook in src/hooks/useFetch.ts. Do not add
  a new data-fetching library."

“Most prompt engineering is just remembering to write down what you already know.”

The big idea: bad prompts produce bad code with mathematical reliability. The fixes are not exotic — chunk requirements, stay neutral, use specific verbs, scope corrections, include examples, anchor in your codebase. Sixty seconds of prompt hygiene saves an hour of debugging.