alexgreensh/attention-span

Attention Span – output‑style plug‑ins for Claude Code

What it is

• A tiny library of output‑style markdown files that change how Claude Code (Anthropic’s code‑assistant) formats its replies. The underlying model and its reasoning stay the same; only the presentation is altered.
• Three ready‑to‑use styles are shipped:
  1. Attention‑kind – ADHD‑friendly, answer‑first, bold‑pointed, easy to skim.
  2. Spartan – ultra‑terse, no warmth, pure signal.
  3. Rundown – TL;DR briefing with checklist emojis, suited for stand‑ups or status updates.

Why it matters

• The same model output can be up to ~43 % shorter on average (median 41 %) while preserving correctness (97 % pass rate on hidden tests). Shorter replies mean less token consumption and faster reading for humans.
• The styles are pure markdown, so they can be dropped into any agent that respects a rules file, not just Claude Code.

How to get it

1. Plugin install (simplest)

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

   This adds the three styles, a /style command, and the skill shortcuts (/attention‑kind, /spartan, /rundown, /tldr).
2. Manual install – download the desired *.md file into ~/.claude/output‑styles (or a project‑local .claude/output‑styles).
3. Activate – set "outputStyle": "Attention‑kind" (or another name) in ~/.claude/settings.json, or run the /style command to switch on the fly.

Using the styles

• In Claude Code: after installation, type /attention‑kind (or the other commands) to make the current conversation follow that style.
• In other agents (Devin, Codex, Antigravity CLI, etc.)
  • Strip the front‑matter with a simple sed command (the README provides the exact one‑liner).
  • Append the clean markdown to the agent’s rule file (AGENTS.md, GEMINI.md, etc.) using the fenced‑marker blocks the README shows, then restart the agent.
• The body of each style is about 650 tokens; it is cached after the first request, so the extra input cost is quickly outweighed by the token savings on subsequent replies.

Complementary tools

• The repo mentions two sister projects that tackle actual token waste:
  • Token Optimizer – analyses structural, runtime, and behavioral token bloat and provides a live dashboard.
  • Outsourcerer – runs a squad of models, picks the best per‑task by benchmark, and validates their work.
  • Both can be combined with Attention Span for a full‑stack cost‑reduction workflow.

Who should try it

• Developers who spend a lot of time reading Claude‑generated explanations or code reviews and want them concise.
• Users with attention‑related challenges (ADHD, fatigue) who need the answer up front.
• Teams that use Claude for status updates, stand‑ups, or quick decision‑making and prefer a checklist‑style brief.

Limitations

• The styles only affect the main conversation; sub‑agents or tool calls keep their own default formatting.
• They do not improve the quality of the underlying answer—only its layout and length.
• Compatibility relies on the target agent accepting plain markdown rules; agents that strictly require front‑matter will need the strip step.

Bottom line Attention Span is a lightweight, open‑source plug‑in that lets you re‑tone Claude Code’s output without touching the model itself. By swapping in a few markdown files you get faster, more readable replies and a modest token‑cost reduction, making it a handy addition to any Claude‑centric development workflow.