"Claude became unreadable" is not a model problem.
It is a briefing problem.
Four everyday prompts went through seven output styles: five built in and two I wrote. Same model, a fresh session per cell, hooks off. 28 answers and 28 real terminal screenshots. On the Friday question five of the seven opened with the identical sentence.
Quick check: why it is not working for you
Three yes or no questions. They explain most cases of "I tried it, no difference". Answers stay in your browser.
- Did you /clear or start a new session after switching style?
- Did you switch through /config rather than the old command?
- Is the work you are reading from the main chat rather than sub-agents?
The same decision came back again and again
What was askedI have three free hours on Friday afternoon. Options: fix the flaky tests everyone complains about, write the onboarding doc for a new hire starting Monday, or prototype the dashboard idea I pitched last week. Help me decide.
The first sentence tells you most of it. Click a card to read that answer in full.
The same decision came back again and again (5/7)
Write the onboarding doc.
That is not a page bug. It is the finding: the style does not change the decision, it changes how much text it takes to reach it.
The full answerConcise
What to notice: Concise
- Same decision, same table, a third to a half shorter. Narration and the closing offer are gone; the substance stays.
- It never promised short. It removes preamble; a 500-word question still gets a 280-word answer.
original answer, English
Write the onboarding doc. It's the only option with a hard deadline (Monday), and the cost of not having it lands on someone else: the new hire loses their first day or two, and you spend Monday answering setup questions live instead of doing anything else.
Ranking:
| Option | Deadline | Cost of skipping | Fits 3h? |
|---|---|---|---|
| Onboarding doc | Monday, fixed | New hire stalls; you get interrupted all week | Yes, 60-120 min for a useful v1 |
| Flaky tests | None | Ongoing annoyance, unchanged | Poorly: root-causing flakiness is open-ended and you lose state over the weekend |
| Dashboard prototype | None | Idea cools slightly | Yes, but nothing breaks if it waits |
Suggested split: spend up to 90 minutes on the doc, scoped to what a new hire needs in week one (env setup, how to run/test, who owns what, first ticket). Don't try to make it comprehensive; have them extend it as they onboard, which also gives them a real first contribution. Use the remaining time on the dashboard prototype, since a fresh idea benefits from momentum and it's better Friday-afternoon energy than debugging. Leave the flaky tests for a Monday or Tuesday morning block where you can keep context across the day.
What it looked like on screen
Answers and screenshots are exactly as they came back. That is the evidence.
The style does not change the answer. It changes the packaging.
Words per answer, on the exact same prompt. On the Friday question all seven said the same thing: write the onboarding doc.
| style | Friday | feature flags | LinkedIn rewrite | URL shortener |
|---|---|---|---|---|
| Default | 382 | 501 | 151 | 589 |
| Concise | 217 | 283 | 116 | 399 |
| Explanatory | 584 | 859 | 172 | 685 |
| Learning | 421 | 556 | 176 | 615 |
| Proactive | 315 | 626 | 196 | 498 |
| Yonatancustom | 321 | 367 | 149 | 508 |
| Tersecustom | 111 | 123 | 83 | 128 |
Concise cuts a third to a half off Default, then keeps going. Its instructions, read out of the binary, cap length nowhere. If you want short, write a number.
Two files. That is the whole thing.
A style is one markdown file. Not five skills and seven hooks. Both are here as plain downloads, no form and no sign-up.
Terse
terse.md
Four lines asking for 120 words maximum. In practice: 83, 111, 123 and 128. Two of four went over, and it is still a fifth of Explanatory.
Yonatan
yonatan.md
A full house voice: answer first, no flattery, evidence before conclusions. Unprompted, it produced a table with a three-colour risk column.
Where it goes, and how to switch
cp terse.md yonatan.md ~/.claude/output-styles/ # or, for one project only: cp terse.md yonatan.md .claude/output-styles/
After copying: /config, then Output style, then /clear. The /output-style command was removed in 2.1.91, so if you tried it that way it genuinely did nothing.
All 28 answers, right here
The full run is above: pick a prompt and a style, read that answer in full with the terminal screenshot for that cell. Check any number yourself.
4 prompts x 7 styles = 28 cells · Fable 5 · Claude Code 2.1.251
Want this voice to be your company's?
One style carrying your house rules, your register and your bans, so everyone on the team gets the same output. 15 minutes is enough to know if it fits.