Five prompt lines that cost you twice
Read time: 3 minutes
You added the line to make it better.
"Double-check your work." "Be maximally thorough." "CRITICAL: YOU MUST ALWAYS."
Anthropic's engineering team just measured what those lines do. They make the answer worse and they make it cost more, at the same time.
On one customer support benchmark, running a prompt audit and deleting lines like those cut cost 14.6% and raised accuracy 5.3%.
That was the audit alone.
Nothing added.
The five patterns
1. Verification rituals
"Double-check your work." "Verify twice before responding."
Delete it. You are paying for a second pass you asked for.
2. Thoroughness boosters
"Be maximally thorough." "CRITICAL: YOU MUST ALWAYS."
Name the output you want instead. Length is not rigor.
3. Mandatory procedures
"Think step by step in a scratchpad." A required process you wrote.
Cut the scaffolding. It stacks on reasoning the model already does.
4. Stale examples
The examples you pasted in a year ago, for a weaker model.
Re-run them. Old examples teach new models old habits.
5. Contradictory rules
Two lines in your file that cannot both be obeyed.
Pick one. Newer models follow both literally, and stall.
What the full pass did
Receipt: with the cost-optimize pass on top, LegalBench came in about 58% cheaper, the retail tau2 benchmark 73%, and SWE-bench 55%. Same work, same models.
The pattern under all five is the same.
Every line was written for a model that needed the help. The model stopped needing it and the line stayed.
I had three of the five in my own CLAUDE.md.
Open your prompt file this week and delete, do not add.
Start with the longest instruction in it and ask what it was protecting you from.