Three Output Styles That Cut Claude Code's Wall of Text

Claude Code answers are complete, correct — and exhausting. alexgreensh/attention-span, a 2k-star collection of output styles, changes how Claude talks to you, not how it codes: answers come first, stay short, drop the jargon, and are formatted to skim. Each style is a single markdown file you drop in and switch on.

Why This Skill Matters

Output style is an underused Claude Code setting: it reshapes the responses without touching the model's engineering behavior. This repo ships three of them. Attention-kind (the flagship) is an ADHD-friendly style that marks every point with a →, bolds the words that matter, and expands only on what deserves attention. Spartan strips the warmth out and compresses hard. Rundown formats status updates as TL;DR briefings with a state checklist and numbered next moves.

The flagship's claims come with a reproducible benchmark (writeup dated 2026-08-11, no LLM judge — hidden test suites only): across 12 coding tasks, pass rates were equal with the style on and off — both 97%. Output ran about 43% shorter on average (median 41%), the answer appeared in the first line 75% of the time versus 3%, and deliverables came out clean 88% of the time versus 12%. The README is explicit that this is a delivery change: it does not claim better answers.

Installation

The one-step path is the plugin, which includes all three styles, the /style switcher, and the user-invoked skills:

/plugin install attention-span --marketplace alexgreensh/attention-span

On Claude Code older than 2.1.275, run /plugin marketplace add alexgreensh/attention-span first, then /plugin install attention-span.

Prefer the manual route? Drop one style file into your output-styles folder and set it as the default:

mkdir -p ~/.claude/output-styles
curl -o ~/.claude/output-styles/attention-kind.md \
  https://raw.githubusercontent.com/alexgreensh/attention-span/main/output-styles/attention-kind.md

{ "outputStyle": "Attention-kind" }

Restart or /clear, and it is on every session. Cost: about 650 tokens of input, added once per session and cached after the first request.

Real Workflow: Try a Style for One Conversation

Committing a default is optional. With the plugin installed, type:

/attention-kind

That conversation now answers first and stays scannable. Ask the README's own comparison question — which database for a new social app — and the default 430-word essay collapses to 94 words spread over five arrow-marked points, PostgreSQL recommendation on line one. When the conversation ends, the style goes with it. Run /style spartan to switch mid-flight, or /style default to put the built-in style back.

Real Workflow: Compress Someone Else's Wall of Text

The styles shape how Claude reports its own work. The fourth skill, /tldr, points that machinery at material you did not write: a document, a thread, a transcript, pasted text. It compresses the input into the same scannable briefing format, which makes it the piece to reach for when digesting long reports rather than writing your own.

The skills are user-invoked only (disable-model-invocation), so they cost zero passive context until you call one — and they are generated from the same style sources by scripts/gen-skills.py, so the wording cannot drift from the flagship styles.

Tips

• The repo is multilingual: official READMEs ship in Spanish, Chinese, Japanese, Portuguese, German, and French.
• Rundown is built for standups and status updates, and it optimizes for readability rather than brevity — reach for it when clarity matters more than length.
• The style body is plain markdown with a <!-- body-start --> marker; one sed line strips the frontmatter, so the styles also work in Codex (via AGENTS.md), Devin, and Antigravity CLI, with per-agent install commands documented in the README.
• Styles apply to the main conversation only — subagents run their own prompt.
• Already installed? grep attention-span ~/.claude/output-styles/*.md shows which version you are on; re-running the install command overwrites with the latest.

When Not to Use This

The README draws its own boundary: these styles do not produce better answers, and that is not what they are for — if your problem is the quality of the engineering rather than the shape of the prose, a style file will not move it. Their scope is the main conversation, so work delegated to subagents keeps its default voice. And if cutting the token bill is the actual goal, the README redirects you to its sister tools, Token Optimizer and Outsourcerer, on the grounds that the bigger cost is the work your agent does, not how it talks.

See the leaderboard for more skills.