Attention-span cuts coding agent output by 43 percent without hurting pass rates

The developer alexgreensh published attention-span, an AGPL-3.0 pack of output styles for coding agents that reduces response length by about 43 percent on average without measurably hurting task success, and the repository has gained 903 stars in a 30-day window.

What it is

Attention-span includes three interchangeable formatting styles for Claude Code, Codex, and other coding agents. The styles rearrange how the agent replies, rather than what the agent does. Answers come first, keywords are bolded, and long prose is broken into short lines that are quick to read. The author calls the styles "ADHD-friendly" and presents the pack as a way to pay attention to the answer instead of paying for tokens.

The numbers

The repository publishes a benchmark run against 12 coding tasks with hidden test suites, and quotes six figures from that run. Output length falls by "~43% shorter output on average (median 41%), and 50-71% on verbose answers". Pass rates come out at "both 97%, within noise" with the style on and off, so the shortening does not appear to cost task quality. The answer is in the first line of the reply 75 percent of the time with the style on, against 3 percent without. Deliverables are clean 88 percent of the time, against 12 percent with no style. The style file itself is roughly 650 input tokens, added once per session and cached after the first request, so the input overhead is small next to the output savings. The README links a full writeup at benchmarks/results/2026-08-11-benchmark.md for anyone who wants to rerun it.

The caveats worth naming

This is one developer's benchmark on twelve tasks, not a large external study, and the README does not name which Claude model was used or the specific harness behind the "hidden test suite" wording. The 97 percent pass rate matching between conditions is the interesting claim, and it is the one that should be independently reproduced before a team standardises on the style.

A working developer running Claude Code or Codex all day generates a large volume of output that has to be read, judged, and either kept or discarded. Cutting that volume by roughly half, when it does not make the run less correct, is a real change to the cost of that workflow: fewer tokens billed, less to skim, and the main answer visible without scrolling. The style file is a small piece of prompt configuration, not a fork of the tool, so trying it against a team's own tasks costs a session or two and can be reverted by removing one file. The interest, 903 stars in the past 30 days, suggests the audience it is aimed at already recognises the problem.

Source

Primary source: alexgreensh/attention-span on GitHub.