A 10,000-character skill test passed at 2,304 while Claude Code's listing neared a 30,000 budget, a team says
A DEV Community post says Claude Code 2.1.296's skill listing budget is 30,000 characters, three times the 10,000 one team's commit test enforced. Japanese characters count double there, and the listing already stood near 28,100.
The Engineer · Build desk
Key figures: about 1,900 characters of the 30,000 skill listing budget remain, by the post's count; the team's commit test would allow 7,696 more characters; the whole listing already stood at about 28,100 characters, an unconfirmed figure.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Listing budget left, about, by the post's count | 1,900 characters | 26 | |
| Room left under test cap | 7,696 characters | 15 | |
| Whole listing, about | 28,100 characters | unconfirmed | 19 |
What happened
- The team wrote its rule and its test from the documentation without measuring, and says they are off in the unit, the part of the listing covered and the number compared against.
- The test's total is a plain sum of JavaScript string lengths, which count UTF-16 code units, so a Japanese character adds one to it.
- The test reads every skill folder, git-ignored or not, so it sums 2,304 characters on the team's machine and 1,692 on a fresh clone.
- The cap was 2,000 characters when the checks arrived on 2026-09-07 and rose to 10,000 on 2026-09-27, when a code comment cited 1M windows for the main session and every subagent.
Why it matters
- Who's exposed exposure Once the listing overflows, Claude Code drops descriptions from the least-invoked skills first, so a team near the limit loses keywords on its rarely used skills before its busy ones.
- What it limits constraint The commit gate stays green until descriptions total 10,000 characters. That leaves it room for 7,696 more than today's 2,304, against about 1,900 characters of space left in the listing.
- What it costs cost Any estimate built on /skill-doctor's figure for the Japanese descriptions is short by 898 tokens, and the listing goes in front of the model on every turn.
The 10,000 figure traces to a unit slip. The repo's first cap was 2,000 characters, annotated as 1% of a 200K window, and the second was 10,000, set when every session ran a 1M window [7][8]. Both equal 1% of the window size read as a character count [16]. The passage the post quotes from the skills page states the 1% rule without a character count [11]. The post's measurement on 2.1.296 is 30,000 for a 1M window, three times the repo's figure [18][21].
Scope is the second miss. The skills page says the listing always contains every skill name, and the test sums descriptions only [14][1]. Even if every one of the 2,304 characters counted double, descriptions would come to 4,608, about 16% of the 28,100 listing [23][19]. Most of the listing is text the test never reads.
The per-entry cap has the same shape. The page says each entry's combined text is capped at 1,536 characters, and the test checks the description alone [12][13].
Counting is the third. A JavaScript length gives a Japanese character one unit, and Claude Code's budget gives it two [6][18]. The token estimate shows a similar gap from another side: /skill-doctor priced the eight Japanese descriptions at 610 tokens against 1,508 actual, about 2.5 times too low [20][24].
The 94% belongs to the whole listing. 28,100 of 30,000 is 93.7% [19][18][22]. The test's 2,304 is 23% of its own 10,000 ceiling [1][17].
These are one team's numbers on one version. For the 30,000 to apply elsewhere, a repo needs the behaviour the post saw on 2.1.296 and a 1M window [18]. For the doubling to matter, its descriptions need characters that count as two, and this team's are Japanese [4]. The post calls that "the part we'd most like other people to check on their own machines" [4].
What to watch
- Whether Anthropic's skills page adds a character count and the double-width rule to its '1% of the model's context window' wording.
- Whether other teams reproduce the 30,000 budget and the double count for Japanese on 2.1.296 and on later versions.
- Whether the team replaces its 10,000-character description sum with a count of what the client actually sends.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap
- Not scored
- Incentives
- Not scored
- Confidence50
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The team's commit test caps the combined length of its skill descriptions at 10,000 characters, and the total sits at 2,304.
- [2]
Every turn, Claude Code puts a listing of skill names and descriptions in front of the model. The listing has a budget, and when it overflows some descriptions are dropped.
- [3]
The team's written rule and its commit test were written from the documentation, not from a measurement, and the post says they are off in three ways: the unit, the part of the listing covered, and the number compared against.
- [4]
The team's skill descriptions are written in Japanese. The post says this mattered more than anything else and is 'the part we'd most like other people to check on their own machines'.
- [5]
The team's rules file says all skill descriptions together must stay within 10,000 characters, reasoning that the listing budget is 1% of the model's context window and that its main session and subagents all run with a 1M-token window.
- [6]
The test's total is a plain sum of each description's JavaScript string length, which counts UTF-16 code units, so a Japanese character counts as one unit.
- [7]
The checks arrived on 2026-09-07 in commit bbf0dc95. The first cap was 2,000 characters, with a code comment calling it 1% of a 200K window, and the nine skills then totaled 1,885 characters.
- [8]
On 2026-09-27, commit 567a49c7 raised the cap to 10,000 with the owner's approval; the new comment says the main session and every subagent run on 1M windows.
- [9]
The committed total is 1,692 characters across eight skills, with descriptions between 145 and 293 characters.
- [10]
The working tree has a ninth skill folder that git ignores, with a 612-character description mostly in English. The test lists every folder, so it sums 2,304 on the team's machine and 1,692 on a fresh clone.
- [11]
The skills page, fetched 2026-10-13 at 19:11 UTC, says the budget scales at 1% of the model's context window and that when the listing overflows, Claude Code drops descriptions starting with the skills invoked least.
- [12]
The skills page also says each entry's combined text is capped at 1,536 characters regardless of budget.
- [13]
A second test checks that each description is between 40 and 1,536 characters long.
- [14]
The skills page says the listing always contains every skill name, but with many skills Claude Code drops some descriptions to fit the character budget, which removes keywords needed to match a request.
- [15]
The test's 10,000 cap leaves room for 7,696 more characters than the current 2,304.
- [16]
Both caps equal 1% of the window size read as a character count: 1% of 200,000 is 2,000 and 1% of 1,000,000 is 10,000.
- [17]
The test's total of 2,304 is about 23% of its own 10,000 ceiling.
- [18]
On Claude Code 2.1.296 the real listing budget for a 1M-token window is 30,000 characters, and a Japanese character counts as two against it.
- [19]
By that count, the whole listing the team's agent received was already at about 28,100.
- [20]
/skill-doctor priced the team's eight Japanese descriptions at 610 tokens when they actually added 1,508.
- [21]
The post's measured budget is three times the repo's cap.
- [22]
The whole listing stood at about 93.7% of the 30,000 budget, roughly 94%.
- [23]
Even if all 2,304 description characters counted double, they would come to at most 4,608, about 16% of the 28,100 listing.
- [24]
The actual token count for the Japanese descriptions was about 2.5 times /skill-doctor's figure.
- [25]
/skill-doctor's figure for the Japanese descriptions was short by 898 tokens.
- [26]
By the post's count, about 1,900 characters of the 30,000 budget remained.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Context window budgetsFollow
- Multilingual tokenizationFollow
- Claude Code SkillsFollow
Entities
- dev.toFollow
- Claude CodeFollow
- /skill-doctorFollow