Same prompt, seven voices: Claude Code output styles, measured

"Claude became unreadable" is a briefing problem, not a model problem.

I kept seeing that complaint, so I stopped arguing and measured it.

The setup

Four everyday prompts, none of them code:

• what to do with three free hours on a Friday afternoon
• explain feature flags to a non-technical cofounder
• rewrite a paragraph for a post
• design a URL shortener

Each ran through seven output styles: the five built into Claude Code plus two I wrote. Same model in every cell (Fable 5, high effort, v2.1.251), a fresh session per cell, hooks off. 28 answers, 28 real terminal screenshots, all published.

1. The decision does not change. The packaging does.

On the Friday question, five of the seven opened with the identical sentence: "Write the onboarding doc." What moved was length. 111 words on Terse, 584 on Explanatory. Same recommendation, same three options, five times the reading.

Here is Explanatory and Terse answering the same architecture prompt:

2. Concise is not "short". It is "no narration".

I pulled its instructions out of the 2.1.251 binary:

keep your responses short and direct while doing the work just as thoroughly

plus six more rules. None of them caps length.

Measured: 217 words against Default's 382 on the Friday question, and 283 against 501 on feature flags. So if you are on Concise and it still feels long, you are half right. It removes preamble and closing offers. It never promised to stop.

3. If you want short, write a number.

My Terse style is four lines with "maximum 120 words" in it. It landed on 83, 111, 123 and 128 across the four prompts. Two of the four went over the cap, and it is still a fifth of Explanatory.

A number is a weaker instruction than people expect, and a much stronger one than no number at all.

4. One markdown file carries a whole voice.

My own house style produced a decision table with a risk column nobody asked for, then closed by naming the assumptions it had not verified.

That is not a skill, a hook, or an MCP server. It is one file in ~/.claude/output-styles.

Three traps behind "I tried it, no difference"

1. The style is read once at session start, so nothing moves without /clear.
2. Sub-agents do not inherit it.
3. /output-style was removed in 2.1.91. It is /config now.

All 28 answers

Every cell, side by side and switchable, with the real terminal screenshot for each:

yonyon.ai/styles