Wednesday, May 27, 2026

Improving the Terrible English Used by Claude Code

  by Steve Endow

[This blog post was 100% human written. None of this post was drafted or written by an AI.]

========

UPDATE 8/6/2026:  This post by Kun Chen seems to explain the bizarre language that is being used by Claude, and perhaps other models.  Chen attributes the behavior to "Reinforcement Learning with Verifiable Rewards" (RLVR).

https://x.com/kunchenguid/status/2084004473408753833

========

I am hardly the grammar police, but I value linguistic precision.

The bizarre English language quirks baked into Claude Code cause it to use terrible words and phrases that make no sense (to a native US English speaker) and significantly weaken the precision of it's language when communicating with a developer.

What do you mean?

Here are some of the terrible phrases that Claude Code spat out while I was trying to develop software:

Load-bearing

Ground truth

Rung scaffold

First-class citizen

Territory

Cast a wide net

This is the smoking gun


Here are some eye rolling phrases I've seen in public PRs that were obviously fully AI generated and never read by a human:

review knowledge-article front-matter

post a non-blocking, feedback-instrumented advisory comment

Harvest + aggregation of acceptance/reactions

Front-matter validator

seed-free graceful degradation covered

Corrects both the naive belief and the folk rule 

a genuine before-image

silently no-ops

Custom Excel file format using SpreadsheetML in an AL codeunit

 By Steve Endow [This blog post was 100% human written. None of this post was drafted, written, or edited by an AI.] I recently worked on a ...