Prompt Patterns I Stopped Using

The way I prompt AI agents in September looks almost nothing like the way I did in April. Most of the change is subtraction. I stopped using patterns that felt clever but added friction, and the work got faster. Here's the drop list.

Dropped: Role-Play Openers

'You are an expert Python developer.' I said this at the top of every prompt for a month. It didn't help. The model is good at Python whether I remind it or not. The only thing the role-play accomplished was burning tokens. Gone.

Dropped: Long Chain-of-Thought Invocations

'Think step by step, then output the answer.' For models that reason by default, this is redundant and sometimes makes the output worse by forcing visible reasoning where none is needed. I let the model decide whether to show work.

Dropped: Format Straightjackets

'Respond ONLY with a JSON object matching this schema...' I used to specify format aggressively. The model mostly complies but the failure mode is ugly. Now I let it answer, and if I need JSON I run a small parser that extracts a JSON block.

Dropped: Few-Shot Examples for Simple Tasks

Examples are powerful for weird tasks. For common tasks they just bloat the prompt. If the task is 'summarize this article in three sentences,' no examples needed.

What I Use Instead

Short, literal, verb-first prompts plus a concrete goal. 'Add a rate limiter to /posts. Test with pytest. Commit if green.' Thirteen words. Better outcomes than any of the patterns I dropped.

See the Anthropic prompt engineering overview for the current state of the art.