The v2.1 ablation sweep (n=10 × 4 brand niches × 3 providers, anchored to
commit 54c3a502, ~544 cells) confirmed these four rules carry no weight in
the skill:
- skill-typo-no-all-caps-body — duplicate of brand-ban-all-caps-body; brand
version is more specific (reserves caps for labels + headings)
- skill-typo-codex-hero-ceiling-repeat — the codex-block restatement of
skill-typo-hero-ceiling didn't add reinforcement on top of the universal
rule
- skill-typo-scale-ratio — duplicate of brand-typo-modular-scale; same
signal, brand version carries the clamp() / fluid implementation detail
- skill-typo-font-count — models don't reach for ≥4 font families in any
niche we test, so the rule has no measurable effect
Each deletion is the Agent A / B / C / D Phase-2 audit recommendation;
none of the four ever validated under either prose state.
Adds EMPIRICAL_VALIDATION.md naming the seven cross-provider winners as the
trustworthy core, and documents the systemic findings (self-priming, detector
saturation, vocabulary anchoring) so future skill edits can avoid the same
traps.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(live-inject): preserve the character after an insertAfter anchor
insertTag()'s insertAfter branch sliced the post-anchor remainder by
prefix.length. When the anchor was not already followed by a newline,
prefix is one character longer than the anchor (the appended '\n'), so
content.slice(prefix.length) dropped the first real character after the
anchor — e.g. `<head>X...` lost the `X` during live-mode injection (#227).
Slice the remainder from the original anchor offset instead. The
insertBefore branch and the already-followed-by-newline case are
unchanged. Add a regression test for both the no-trailing-newline and
newline cases, and regenerate the tracked per-agent bundles so the fix
ships everywhere.
Fixes#227. Root-cause analysis from the issue reporter.
* Fix live inject CRLF insertAfter handling
---------
Co-authored-by: Paul Bakaus <paul.bakaus@gmail.com>
Derive a Gecko-compatible manifest at build time and package
extension-firefox.zip alongside the Chrome zip:
- background service worker is declared as an event-page `scripts`
entry (top-level listeners + in-memory Map run unchanged on Gecko)
- browser_specific_settings.gecko with id, strict_min_version 140.0,
and data_collection_permissions (required by AMO; honored on 140+)
- packZip helper parameterized over cwd/excludes; `*.DS_Store` strips
junk at every depth and .DS_Store is excluded from the staging copy
- guard against a missing background.service_worker shape
CI now builds the extension and runs a pinned `web-ext@8 lint` over
the staged Firefox tree (innerHTML warnings are non-blocking); the
unpacked staging dir is excluded from the uploaded artifact. The
release script attaches both zips and points to AMO.
Bumps the extension to v1.2.0 with a changelog entry.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Paul Bakaus <paul.bakaus@gmail.com>
Mirrors the 5 prose changes in skill/SKILL.src.md + skill/reference/brand.md
out to every harness directory (`.claude`, `.gemini`, `.cursor`, `.codex`,
`.agents`, etc.) so the staged skill that workers / agents read matches the
source. Auto-generated by `bun run build:skills`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase-2 ablation audit caught these rules causing the exact behavior they
ban via the literal examples in their own prose. Verified: OpenAI samples
under skill-on produced "fake theater", "vendor theater", "heatmap theater"
as verbatim copies of the 'X theater' example. Same pattern for the
restrained-on-cream example, the aphoristic-cadence template, and the
"reserve uppercase for…" enumeration.
- skill-ban-codex-x-theater: drop the 3 syntactic templates + 3 example
phrases ("Productivity theater" etc.)
- brand-imagery-required: drop the niche enumeration that cued
"imagery not required elsewhere"
- skill-typo-no-all-caps-body: drop the "Reserve uppercase for labels /
eyebrows / badges" enumeration that primed uppercase usage
- brand-color-no-converge: drop the "restrained-on-cream" example that
was priming cream-heavy palettes
- skill-copy-no-aphoristic-cadence: drop the literal cadence template
("serious statement, then punchy short negation") that named the
rhythm it bans
Ablation re-run pending in impeccable-evals to measure impact.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Improve CI test coverage
* Stabilize live E2E harness
* Shard live E2E CI
* Cache live E2E CI dependencies
* Stabilize live E2E smoke CI
* Update generated live browser bundles
* Tighten live E2E smoke runtime
* Prevent live E2E smoke hangs
* Stabilize live E2E CI coverage
* Fix stale accept DOM cleanup
* Regenerate live browser outputs
The verbs/labels/icons were copied three ways: live-browser.js (ICONS + ACTIONS),
VISUAL_ACTIONS in live-event-validation.mjs, and the marketing demo. Collapse
them to one source, skill/scripts/live-vocabulary.mjs (LIVE_COMMANDS + derived
VISUAL_ACTIONS).
- live-event-validation.mjs imports VISUAL_ACTIONS from it.
- live-server.mjs serializes LIVE_COMMANDS into window.__IMPECCABLE_VOCAB__ when
it serves /live.js, next to the token/port. live-browser.js (served raw, can't
import at runtime) builds its ICONS + ACTIONS from that injected vocab instead
of an inline copy — byte-identical icons, zero behaviour change.
- site/components/LiveDemoPalette.astro imports the same module at build time, so
the demo and the real picker can no longer drift.
Adds a /live.js test asserting the injected vocab deep-equals the canonical list.
Harness skill dirs refreshed via build. (Pre-existing, unrelated: `bun run
build:site` fails on an htmlparser2 import in the CLI detector.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Rewrite the hero around the why (the missing design vocabulary) instead of the
live-mode how: "The missing design vocabulary for agents." The live demo now
opens the picker's command palette and picks a verb before generating, which is
the move that makes the live approach unique and was previously skipped.
- Demo palette mirrors the real action picker (live-browser.js): same 12 verbs,
the same SVG icons, a 4-col icon-over-label grid, selected chip on a kinpaku
wash with its icon recolored. Light + dark covered.
- Shared <LiveDemoPalette> component renders the grid from one list, so the hero
and /live-mode no longer hand-copy the markup. /live-mode lands on "delight",
the hero on "colorize" (via data-demo-pick); pages without a palette filter the
switcher beats out of the shared timeline.
- Trim the opening beats so the cursor clicks the element at ~1.3s (was ~2s), and
slow the palette browse so the vocabulary is readable.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The changelog + FAQ pages share changelog-faq-kinpaku.css, which paints its
card/panel/pill backgrounds with literal dark oklch(8% …) values that don't
flip with the theme. The only light override was dead twice over: scoped to a
.changelog-faq-kinpaku wrapper that's never on the body (these pages use
.kinpaku-surface), and naming the wrong elements (.faq-item/.changelog-entry
instead of the cf-prefixed classes). So the cards stayed near-black on the
light page.
Replace that dead block with correct html.light .cf-* rules: flip the card
surfaces (cf-entry, cf-stats, cf-faq-item, cf-entry--current), the before/after
frame + before tag (cf-ba-shot, cf-ba-tag--before) to the shared light card
treatment, and drop the bright-kinpaku accents (cf-version, cf-stat-num,
cf-current-badge, cf-ba-tag--after, cf-faq-question) to --ks-kinpaku-ink so
they stay legible on paper, matching the .cf-eyebrow. Answer-body tokens
already flip, so they're untouched. Dark mode unaffected.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Each command section now reads in its own category accent (--spread-accent)
across the palette: the kicker, the skill name (big title + active list
item), and the divider all take it, while the slash before /impeccable
drops to the muted namespace ink. The per-section divider accents are also
made vibrant — refine/simplify/harden were muddy kinpaku-pale/-deep/oxide;
now create/refine/simplify = kinpaku gold, evaluate/harden = patina
verdigris, system = neutral, matching the established category color
grouping. Light mode re-applies the accents over the shared eyebrow :is()
rule; dark mode picks them up from the base rules.
Also includes light-mode readability fixes for the live-demo G-bar (brand
mark, active tool chip, control chips, pin-note text/caret) and a
specificity fix so the hotel-hero demo text stays light on its photo.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Default is "auto", which inherits from the OS via prefers-color-scheme
and follows it live. Clicking cycles auto → light → dark → auto; the
explicit choice persists in localStorage while auto stores nothing.
The toggle shows the active preference (half-circle / sun / moon) keyed
on a new data-theme-pref attribute, so "auto" is its own visible state
rather than collapsing into whatever the OS resolved to.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Paul BakausandClaude Opus 4.7
github-actions[bot]
Abdul WahabandGitHub
Gabriel GrantandGitHub