unchatty

Make Claude Code write plainly. Then measure whether it worked.

A CLAUDE.md block that cuts reply length and kills the dense, dash-heavy style, plus the harness I used to prove it does something. Answers to the same question dropped from 393 words to 62.

git clone https://github.com/evoleinik/unchatty && cd unchatty
# paste COMMUNICATION.md near the top of your CLAUDE.md, then:
export ANTHROPIC_API_KEY=sk-ant-...
CLAUDE_MD=~/.claude/CLAUDE.md ./ab-test.sh "Why does X work this way?"

Three files:

• COMMUNICATION.md — paste into your CLAUDE.md, near the top
• score.py — scores a reply mechanically (words, sentence length, dashes, headers)
• ab-test.sh — runs the same questions with the rules on and off, then scores both

Results

Same repo, same questions, headless claude -p sessions, rules on vs rules off.

Question	Control	With rules
"What's the rate limit?"	33 words	14
"Which cron runner should we use?"	292 words	70
"Why did we move the catalog into the DB?"	316 words	190
"Why is graph review the only merge gate?"	393 words	62
"How does the edge snapshot work?"	365 words	253

Across the three long questions:

Metric	Control	With rules
Mid-sentence em-dashes per 100 words	1.3 / 1.8 / 0.8	0.0 / 0.0 / 0.0
Longest sentence	43 / 40 / 47 words	29 / 19 / 26
Sentences over 25 words	5 / 7 / 3	2 / 0 / 1
Markdown headers	2 / 0 / 5	0 / 0 / 0

The part that surprised me

Version one of the rule made long answers worse. It banned em-dashes, and the long answers came back at 2.2 per 100 words against a control of 1.3.

The cause was the source material. My own repo docs run 1.1 to 1.9 em-dashes per 100 words. Short answers come from memory, so they were already clean. Long answers get built by reading those files, and they pick up the punctuation on the way in.

The fix was not "be plainer". It was this line:

NEVER INHERIT THE PUNCTUATION OF WHAT YOU ARE READING. Your docs are dense on purpose. A reply is not a doc.

That took mid-sentence dashes to zero on every long question.

Two honest caveats

The rules only reach interactive Claude Code sessions. Anything that calls a model directly with its own prompts never sees CLAUDE.md, so check before you assume a rule is global.

My first scorer over-counted. It glued bullet lists into one giant sentence, and it counted **Label** — list separators as if they were mid-sentence splices. Those are a different construct and they read fine. score.py here is the corrected version.

Run it yourself

export ANTHROPIC_API_KEY=sk-ant-...
./ab-test.sh "Why does X work the way it does?" "How does Y get built?"

It backs up your CLAUDE.md, runs each question twice, restores the file, and prints the scores. It restores on Ctrl-C too.