[MODEL] Recent versions of Claude are extremely verbose, and requests to fix this in CLAUDE.ms and hooks are ignored

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Claude Code has become way more verbose than it used to be. In conversations with user, and well as in documents it writes. Also very opinionated, and not professional -a lot of editorial type comments.

Also conversations are highly detailed and high on statistics (granted this is a stats project), but I have repeatedly told Claude that I am a novice, and to keep things at a basic stats level.

Basically not following CLAUDE.md

What Claude Actually Did

Very verbose, more detail than needed, convoluted sentences.

Basically not following CLAUDE.md. Attached CLAUDE.md, and an original document it wrote.

CLAUDE.md.txt

PLAN.md

Expected Behavior

Should be tight, to the point, professional. Have attached a shorter/better

PLAN_short.md

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

Just ask claude a question , or ask it to write a bug report. I am attaching hooks that I use.

brevity-check.py
brevity-judge-prompt.md
brevity-judge.py
brevity-reminder.sh

Claude Model

Opus

Relevant Conversation

Here is an example response which is way more detail than I want - 

1. BH-FDR on C1/C6 is backwards.
BH controls false positives — it inflates p-values so rejecting a null gets harder. C1 ("delta sits at the floor") and C6 ("null survives the coarser partition") are affirmed by not rejecting. Correcting them makes their preferred answer more likely, so it's a bias in their favour, not a safeguard. The right instrument is what lines 198–200 already do for C6: a non-inferiority / equivalence bound (upper CI below a pre-registered margin) plus a power statement. C1 has no such bound — it's stated as "CI straddles the floor", which any noisy estimate satisfies. Restrict the BH set to C2, C3, C5 (C4 depends on how it's stated), and give C1 the same TOST-style bound C6 has.

2. C5's threshold is stated twice.
The 0.05 effect size is already baked into m1_threshold = 0.430 (0.380 floor + 0.05). So R7.2's wording — "lifts the floor by at least the pre-registered 0.05 (M2: seq > 0.430)" — asks for 0.05 on top of a gate that already includes 0.05, i.e. 0.100 over the bare floor. R7.7 asks the other question: does M2's +0.008 margin over 0.430 have a CI excluding zero. That's the single-count version and it matches how the gate is recorded in summary.json. This is exactly the interchange lines 39–41 warn against.

Fix is one line: restate C5 as "ESM-3 scale clears the pre-registered M2 gate (family-split F1 > 0.430); affirmed if the paired bootstrap CI on the +0.008 margin excludes zero" — and note in R7.1 that the 0.05 lives in the threshold, not in the test.

Impact

High - Significant unwanted changes

Claude Code Version

Opus 5

Platform

Anthropic API

Additional Context

A lot of time is spent on working with Claude to be more to the point. I am spendinmg more time on this than on coding.