Attention Span

A set of markdown files that change how a coding agent reports its work. Attention-kind is the flagship: the conclusion comes in the first line, each point is spaced out and marked with an arrow, the important words are bold so the bold alone carries the answer, and a rare technical term gets a five-word definition once. Spartan is the same shape with the warmth removed; Rundown opens with a TL;DR and shows state as a checklist. Around the three files sit a benchmark that measures the work and the output separately, a release workflow that fails when the version file, the markers inside the styles and the git tag drift apart, and four user-invoked skills generated from the style sources so they cannot drift from the wording. The styles install as a Claude Code plugin, work in the Claude apps as skills, and ship as plain markdown bodies that can be stripped of their frontmatter and dropped into another agent’s rules file.

Who built itThe repository is close to one person’s work: 61 of its 66 commits are the author’s, and the other five came from four accounts, each of which the changelog credits with a change that landed — the multi-agent install, the Spanish README, the style switcher command, and a rule about brevity. Nineteen commits carry a co-author trailer, eighteen of them for Claude Opus 4.8 (1M context) and one for Opus 5.

1. 01

   One markdown file becomes a plugin in a month

   The repository was created on 2026-08-04 and has 66 commits, 60 of them in August and six in the first six days of September. Seven releases came out of that month, one every few days at first: 0.2 on 2026-08-05, 0.3 the next day, 0.4 on 2026-08-10, 0.5 on 2026-08-11, 0.6 on 2026-08-15, 0.7 on 2026-08-21 and 0.8 on 2026-09-06. The first entry in that list is 0.2; there is no 0.1. The tags match the releases exactly, because the release workflow fails the release when the version file, the version marker inside each style and the tag disagree, and versioning is deliberately flat: the changelog says only the rightmost number bumps. Around that sit 1,150 stars, 37 forks and two open issues, and five accounts have committed — the author on 61, four others with one or two each. Nineteen commits carry a co-author trailer, eighteen of them naming Claude Opus 4.8 with a one-million-token context window.

2. 02

   What the rewrite actually looks like

   The README argues with two side-by-side answers to the same question. Asked which database a new social app should use, the unstyled answer runs 430 words and reaches its recommendation after a paragraph of framing; the Attention-kind answer is 94 words in five arrow-marked points, the first of which is the recommendation. Spartan does the same to a question about cutting one of three priorities, 310 words down to 168, and Rundown turns a hiring summary into a TL;DR, a four-line checklist, one blocker and four numbered next moves. What the files ask for is narrower than “be brief”: answer first, say the least that fully answers, expand only where cutting the expansion would cost the reader, define a rare term once, never restate a point. The changelog records sharpening rather than invention. 0.2 split answers from deliverables, so an explanation or a decision stays lean while a document runs as long as the work needs, and stated that brevity trims the reply and never the internal reasoning. 0.5 added deliverable purity — ask for a commit message or an email and you get that and nothing around it — plus a rule that a warning is never trimmed to save space, and one that the bold and the TL;DR alone must carry the whole answer.

3. 03

   The benchmark was rebuilt after somebody took it apart

   Five days after the first benchmark, a reader opened an issue arguing it measured compliance rather than quality: the model generating the answers was the model judging them, the sample was 12 questions times two runs, and the two metrics that moved — answer-first from 63% to 96%, skimmability from 2.7 to 4.8 — restated the style’s own instructions as a rubric, so of course they improved. The author agreed in the reply, said the benchmarks could have been better and rebuilt them. The version dated 2026-08-11 is built to be attacked: the work is checked by hidden test suites where pass or fail is ground truth, the solutions are generated hermetically with the ambient style off so the arms differ by one file, and the headline numbers use no model as a judge at all. On 12 coding tasks the two arms pass the same number of runs, 35 of 36; on 24 questions the styled answers are 43% shorter on average and 50 to 71% shorter on the verbose ones, while a two-line answer shrank 7% — the saving scales with how much the answer would otherwise over-explain. Instead of a reading-grade score, which sees only sentence length, it counts words before the first emphasized point: about 6 against about 40, with the answer in the first line 75% of the time against 3%. The writeup ends by listing its own limits, including that work-equivalence is shown only on tasks that have hidden tests.

4. 04

   Every rule arrived with a user attached

   Most of the rules in these files came from people running them daily. Issue 6 reported a blocking question buried in paragraph four of seven, where a yes-or-no answer preceded a step with a side effect; 0.7 shipped the fix as a placement rule — a question you must wait on is the last block with nothing after it, and line one carries it when the reply has other content — and the same release added a placement checker and six side-effect scenarios in which real Claude lands the ask last six times out of six while a control that buries the question fails six times out of six. Issue 9 pointed out that the checklist folded unknown into not started, and 0.8 gave it a fourth state; issue 8 asked for the next-move choices to be numbered so they could be picked by number, also shipped in 0.8. A contributor’s pull request, titled “brevity governs the reply, not the work”, came out of a session on a slow data pipeline where the agent kept stopping early and offering to write up claims it had not verified, having turned the write-up into the finish line. Issue 5 asked for the styles as skills for the Claude apps, and 0.8 shipped those as well, generated from the style files. Two community threads are still open: a request to skip the whole TL;DR-plus-next-move scaffolding on a single-fact answer, opened the day of the last push, and a pull request carrying four more README translations.

5. 05

   The distribution story, and what it declines to claim

   The shipped artifact is a markdown file, so distribution is mostly about where the file lands. The one-step route is a Claude Code plugin that brings the three styles, the switcher command and the four skills; the manual route curls one file into an output-styles directory and names it in settings. The plugin manifest is versioned in step with the styles, and the release workflow refuses to publish while manifest, version file, in-file markers and tag disagree. The trade-offs are written down. The style body costs about 650 tokens of input, loaded once per session and cached after the first request, and the README argues the measured output saving dwarfs it in a few replies. Styles apply to the main conversation only, because subagents run their own prompt, and the files keep the coding instructions intact — the claim the hidden-test arm checks. The skills are user-invoked only, so they cost no passive context until you type one, and a script generates them from the style sources so the two sets cannot drift. The README also declines a pitch a reader might expect: if cutting token spend is the goal, the bigger cost is the work the agent does rather than how it talks, and two sister tools by the same author are named for that job. One point in the benchmark thread is left uncovered by the material: AGPL-3.0 on a prompt written in markdown — the recorded reply stops before it.