技能 2026-10-07 08:57 UTC
skill.color-expert
由提交人 claude 提交,主要更新类型为:问题缺陷修复与稳定性提升。提交说明:Fix skill accuracy, split the index, and give the eval set a stated rule。解决了已知运行时异常与边界错误,提升系统稳定性。
Commit提交信息原文
Fix skill accuracy, split the index, and give the eval set a stated rule
Five review findings, addressed.
1. SKILL.md accuracy. Repair a spliced sentence in Deep References
("GenCol ... thresholds)or paper." was originally "GenColor paper").
CSS Color 4 defines 148 named colors over 138 unique values, not 147;
drop the conflicting "140 CSS names" in the colordx reference. CIELAB
is 1976, so Lewis 1931 anticipates it by decades, not 30 years — which
is what the underlying reference file already said.
2. Counts. The repo quoted four different file totals (140+, 144, 173,
113) and understated words, tools and video sources by a third to a
half. Every number in README.md now comes from scripts/stats.sh;
per-section file counts and the two layer line-counts are gone, since
they rot on every commit. MAINTENANCE.md already listed "stale counts"
under Signs of Drift and now points at the script.
3. Index size. references/INDEX.md was 123 KB (~30K tokens) for a file
SKILL.md tells the agent to consult, because the summary column ran to
1,352 characters and a third of the bytes were alignment padding. Split
into a compact map (32 KB, one line per file, all 181 reachable) plus
per-directory indexes carrying the full summaries verbatim, with online
tools and the PDF list moved out to their own files. No content lost;
every link checked. Adds a layer budget to MAINTENANCE.md so it does
not grow back.
4. Trigger evals. Six cases were labelled should_trigger:false while
SKILL.md explicitly claims them (picking colors for a landing page,
token structure, chart color assignment, interpolation space), so the
set encoded no consistent rule. evals/TRIGGERING.md now states the
test — trigger when a color-quality judgment is at stake, not when
color is incidental — and every case carries a kind and a why, with
the arguable ones recorded so they are not silently re-flipped.
5. SECURITY.md. It told scanners the references are curated and "not raw
scrapes"; the huevaluechroma and colorandcontrast files are raw
extractions with page markers and broken line wrapping. Said plainly,
with a note that the scraped pages were public.
Also fixes a genuinely broken link in jaffer-color-name-dictionaries.md
and files huevaluechroma/links.md, the one reference never indexed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HyXKLpstbYZPQ3e17xsP54