attention-span — the note that told Claude to bold line one, and what the number counted

attention-span is three short notes that change how Claude Code, the coding AI assistant, talks. The notes say “answer first, keep it short, go long only where it matters.” That is all they are, plain text, not a program. They leave how the assistant writes code alone and change only how it speaks. The source is github.com/alexgreensh/attention-span.

There are three notes, and each talks differently

Name	In one line
Attention-kind	Conclusion on line one, short, and the bold words alone carry the answer
Spartan	Same frame with the warmth removed, blunt
Rundown	A one-line summary on top, status as a ✅🟡⬜❔ list, numbered choices at the end

Two small extras ride along. /style switches between them, and /tldr shrinks someone else’s long text.

I recounted the two numbers it brags about

The README leans on two numbers. Answers got 43% shorter, and the share of answers with the conclusion on line one went from 3% to 75%.

The repo ships the raw data (24 questions, three answers each), so I counted again myself.

• Length held. Adding up every character gives a 42.9% cut.
• Line one meant something else. The code behind that number counts whether line one starts in bold (**).

But “start line one in bold” is exactly what the note orders. The number counts whether the order was followed. It is like writing “put your name in big letters at the top” on an exam, then announcing that students with big names rose from 3% to 75%. It proves the rule was followed. It does not prove the answers read better.

The raw data holds a counter-example too. One answer opens with “Monorepo.” and gives its verdict right there, but without bold, so it scored zero.

“The work stays the same” has a narrow reach

The second boast is that coding ability is unchanged. Twelve coding problems were solved and graded by hidden tests. With the style on or off, both scored the same 35 out of 36.

One step happens before grading, though. It throws away the prose and keeps only the code blocks. So whether “be brief” made the assistant do less work cannot show up, by design.

A user did report exactly that. In an issue they described the assistant posting unchecked guesses and then stopping, and read “keep the report short” as “keep the work short.” The author accepted it and added two sentences. That came 26 days after the benchmark passed, on September 6, against August 11.

The cost line is from before the note tripled

The README says the note costs “about 650 tokens at the start.” That sentence went in on August 5, when Attention-kind was 439 words. Today it is 1,327 words, three times as long, and the sentence is unchanged. In tokens that is around 1,900 now (an estimate, not a tokenizer count).

Against my setup: not turning it on for now

• My output style is the default. I never changed it.
• Every measurement in this repo is in English. I talk in Korean. The note itself carries English style rules (“no dashes”, “no it’s not X, it’s Y”), so I do not know whether it works the same in Korean.
• It collides with my own rules in two places. I restate a question in one line before answering. The note says “don’t restate, line one is the verdict.”
• Forcing one format on every reply is something I looked at once and dropped, because small talk gets formatted too.
• Two of the repo’s good rules were already mine: “a blocking question goes last” and “mark unknown state as unknown.”

The verdict is hold. Trying it costs one line to switch on and one line to switch off, so it is cheap. But I would have to settle the rule clash first, and without a way to measure in Korean I would only get a feeling of “better.” What I took was not the note but one measuring structure: measure the work and the wording separately. Because the repo split them, it was easy to see which number is weak.

What I don’t know yet

• I did not measure length in Korean. There is no basis for carrying 24 English questions over.
• The app’s built-in Concise style I only saw by name. “Why not just use that” has no answer yet.
• This entry is not a verdict on the repo’s worth. It records what a number actually counted.

The detailed record starts here

The length cut of 43% reproduces from the raw data, and “answer in line one, 3% to 75%” was the share of answers whose line one starts in bold. attention-span is an AGPL-3.0 repo that ships three Claude Code output styles (Attention-kind, Spartan, Rundown) plus /style and /tldr. Created 2026-08-04, last push 2026-10-21, 1,484 stars. I checked six claims: length, the line-one metric, the skim metric, work unchanged, the cost line, and the benchmark’s history. The evidence is the metric code, the raw data, issues and PRs, and the commit log. The verdict is hold on always-on use, keep only the measuring structure.

What it is

Three markdown files that plug into Claude Code’s output-styles feature. The line keep-coding-instructions: true at the top of each file means “leave the coding instructions as they are.”

Part	Size (English words)	What it does
Attention-kind	1,327	Verdict on line one, short by default, expand only what matters, → markers so bold alone gives the answer
Spartan	618	Same frame, warmth and cushioning removed, imperative
Rundown	460	One-line summary, status list, numbered choices
/style	—	Edits the single outputStyle key. Write access is limited to two settings.json files
/tldr	—	Squeezes someone else’s text into a summary line, 3 to 7 points, and action items. “No new claims”

Techniques, and when each one holds

Technique	Holds only when
Answer first, length as a signal	The reader will not finish. Whatever comes after their attention runs out counts as not delivered
Bold alone carries the answer	A rendered markdown screen, read by skimming
Brevity switches off when asked to go deep	A safety valve so “short” does not become “left something out”
A blocking question goes in the last block	Notifications and skimming, where a buried question hides that work has stopped
Deliverables come bare	Requests for an email or commit message you will paste as is
One source, a generator, a release check	Copies must never drift. A mismatch in tag, VERSION, in-file marker or byte-identity of copies stops the release
Porting to other agents	The header is Claude Code only, so everything after the body-start marker is cut out with sed

The source-and-copy identity gate has the same shape as the one in the anti-slop entry. This one is far smaller and simpler.

There are four conditions overall. It assumes answers in English, one question with one answer, use in the main conversation only (the README says sub-agents are not covered), and a note that is loaded as input every session.

What broke when I opened it — six places

#	Claim	What I checked	Result
1	Length down ~43%, median 41%	Recounted gen2.jsonl, 24 questions × 3 runs. Pooled 42.9%, per-question mean 40.9%, median 40.6%, range 7.2 to 71.2%. Lookups (52%) and explanations (46%) shrink most, writing (26%) least	Reproduced
2	Answer on line one, 75% vs 3%	The metric (answer_first) asks whether line one starts with **. 54/72 vs 2/72 reproduces, but it is compliance. The “6 words vs 40” figure uses the same family (words_to_emphasis), and its median is 0.0 vs 26.0	A shape metric
3	Easy to skim	The same file’s own numbers: anchors per 100 words 6.9 → 5.7 (down), mean paragraph 35.9 → 32.6 words (almost flat). What shrank is the longest block (37 → 15)	Half supported
4	Work unchanged, 97% vs 97%	Before grading, extract_code drops the prose and keeps code. Any effect on how much work gets done is invisible by design. Issue/PR #7 (merged Sep 6) is the counter-example	Counter-example outside its reach
5	Cost ~650 tokens	The line dates from the 2026-08-05 commit (439 words). Now 1,327 words. About 1,900 tokens [estimated]	Stale
6	The benchmark’s history	The first benchmark drew the criticism “Claude generates, Claude grades, and the rubric restates the style’s rules” (issue #4). The author rebuilt it on Aug 11 with hidden tests and judge-free metrics. An Aug 12 commit removed five old result files from the public tree (git history keeps them)	A fact

The author’s response was good. After the criticism they rebuilt it with metrics that need no judge, and their limits section states the small samples and the coding-only scope. What this entry points at is not a weak response. It is that some of the new metrics came back to the shape of rule-following.

The conclusion that recurs in this category shows up again: the name on a number and what it counted come apart. It is the same kind of gap as “944 vs 1” in the eli5 entry.

Against my setup → what to take → verdict

Technique	Where it would go	Verdict	Gain	Cost to adopt	Cost to undo	Confidence
Attention-kind, always on	One outputStyle line	Hold	Length −10 to −45% [estimated, middle band of the English per-question range]	One line plus two rule clashes to settle	One line	Low (no Korean measurement)
Measure the work and the wording separately	Checking the effect after I change a response rule	Take	“Did it get better” as a number, not a feeling	Half a day of scripting [estimated]	0	Medium. Wording metrics rise when the rule is followed, so the work metric must be separate
Hide manual-only skills (disable-model-invocation)	Two skills	Hold	1,052 of 29,405 description characters (3.6%)	One line each	One line	High (official docs). But the spoken way to trigger them goes away
Blocking question last, mark unknown	Existing rules	Drop, already solved	0	—	—	High
Source-and-copy check	(no copy folder on this PC)	Drop, nowhere to apply	0	—	—	High

Reopen when. I measure 10 to 20 Korean questions with and without the style and the drop is at least 20%, I have settled the rule clash, and I have confirmed small-talk turns do not get the format. Measuring it next to the built-in Concise style would also do.

What I didn’t do

• Korean measurement. I did not turn the style on and ask the same questions. Running it as a child process would carry my personal settings along and contaminate the comparison.
• I did not open the built-in Concise text.
• Independent-lens adversarial checks (refute, already solved, condition violated, each run without seeing my conclusion) and the “what survives” stage were skipped. This is the basic depth.
• The 1,900 tokens is a conversion. The description total counts only the first line of each skill’s description, so multi-line descriptions may be undercounted.
• The verdict reflects one person on one PC.