Follow-up to #134, which scoped validateProse to user-facing copy and
left the LLM-facing skill files alone. Bring those to the same bar,
phased so hardening repetition stays intact.
- Em dashes: 419 → 0 across SKILL.md and 35 reference files. Each
replacement picks the right relationship (colon, semicolon, period,
or parens) instead of letting the dash hide the choice.
- Closer cleanup: deleted or rewrote the "Remember:" sermonettes that
were pure adjective chants (bolder/quieter/clarify/delight/extract/
colorize/layout/typeset/audit/adapt). Survivors that load-bear an
instruction now hand off to /impeccable polish instead of summarizing.
- Opener taglines: rewrote the "[Verb] [object] to [outcome]" brochure
openers in 12 older files to lead with the failure mode, the
strongest claim, or a directive. Newer files (live, brand, product,
audit, critique, harden) kept their existing openers.
- data-driven: rephrased the two technical hits in live.md so the
validator can stay strict on this term.
- validateSkillProse: narrow validator scoped to source/skills/impeccable/.
Em-dash check + the small denylist of phrases with no technical
reading. Hardening repetition and structural-prose rules are
deliberately not enforced — those need human judgment.
Test failure on detectUrl is pre-existing (puppeteer needs --no-sandbox
when running as root); unrelated to these changes.
https://claude.ai/code/session_013zZY6rbB1bS8z3D63rX5hW
Co-authored-by: Claude <noreply@anthropic.com>
Site copy was being called out as AI slop (specifically the word
"load-bearing"). Five-pass cleanup with a build validator to keep it
from creeping back.
Pass 1 — mechanical purge:
- Remove "load-bearing" from impeccable.md, brand.md, live.md,
iterate-live.md
- Remove "highest-leverage" from critique.md, typeset.md, designing
- Remove em dashes from all 9 slop-page rule cards
- Replace "leverage" verb in personas.md
Pass 2 — rewrite the worst offenders:
- Changelog v2.0 "Data-driven skill rewrite" entry: drop "data-driven",
"frontier models", "collapses into monoculture", "biggest unlock",
"reflex defaults"; name the actual mechanism
- README opener: drop "deeper expertise and more control"; replace with
three concrete differentiators (7 reference files, 23 commands, 27
detection rules)
- Neo Mirai case study opener: action-first, name the image model used
Pass 3 — editorials:
- Fix negation pivot in distill.md ("simplicity is not about ... It is
about ...")
Pass 4 — homepage why-panels:
- Foundation lead: name the 7 reference files specifically
- Language lead: show the discipline mapping with real command names
- Production-codebases panel: drop "Impeccable isn't a sketchpad"
negation pivot
- Ships-code panel: replace "is native to that world" with "runs there"
Pass 5 — STYLE.md and validator:
- New STYLE.md at root: editorial brief with 12 principles and the
enforced denylist (each rule with a rationale and a suggested
replacement)
- scripts/build.js: validateNoEmDashes becomes validateProse. Adds 21
phrase rules with rationales, catches the \`--\` em-dash substitute,
expands target list to site/pages, site/content, README.md,
README.npm.md
- CLAUDE.md: replace the em-dash section with a STYLE.md pointer and
document the two-content-tree footgun (content/site/ vs site/content/
must be edited in lockstep until they're unified)
Co-authored-by: Claude <noreply@anthropic.com>
* feat(detector): flag italic-serif display heroes and uppercase eyebrow chips (#127)
Two new rules covering the structural tells of late-2025/early-2026
AI-generated marketing pages.
- italic-serif-display: oversized italic serif (Fraunces, Recoleta,
Newsreader, Playfair, Cormorant, Tiempos, ...) as the primary hero
headline. Anchored on h1 (or h2 at >= 48px) with font-style: italic
and a serif primary face.
- hero-eyebrow-chip: uppercase letter-spaced label sitting as the
previousElementSibling of a hero h1 (font-size >= 48px). Bounded
text length 2-30 chars, letter-spacing >= 1.6px, font-size <= 14px.
The pill-chip variant (background + border-radius: 999px) falls out
of the same gates for free.
Both follow the existing icon-tile-stack pattern: pure check function +
browser DOM adapter + jsdom adapter, wired into both element loops.
Two-column fixtures (4 flag / 6 pass each) drive the jsdom tests.
Skill copy in source/skills/impeccable/reference/typography.md and
critique.md calls out the patterns by name. The italic-serif rule's
description acknowledges that editorial/magazine register may legitimately
want the pattern -- judge by context.
Closes#127
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Add sandbox gotchas for Codex
* Trim verbose detector skill copy
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Paul Bakaus <paulbakaus@pauls-mbp-3.lan>
The seven-lane list in Phase C departure mode was acting as a menu:
the model ran "furthest from editorial" as its selection criterion and
converged on Swiss-grid / Terminal / Industrial-signage every time.
Replaced with a brand-voice derivation process (read personality
words, imagine physical experiences, derive visual directions).
Explicitly names the failure mode so the model can't fall into it.
Phase D family-pass labels are now open-ended nouns, not a fixed
vocabulary list that re-anchored the same categories.
Reinforced parameter generation: Phase C (both modes) now requires
naming 2-3 parameter knobs alongside each variant during planning,
not as an afterthought. The freeform bias paragraph aligns with
the budget table (2-3 for large compositions, not 1-2) and frames
0-param heroes as mistakes, not judgment calls.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Restructures live.md Step 4 into four phases: identity extraction
(Phase A, non-skippable, falls back to CSS variables and computed
styles when DESIGN.md is absent), default vs departure mode pick
(Phase B), variant planning by primary axis or aesthetic lane
(Phase C), and a family-level squint test before the sentence
pass (Phase D). Default mode preserves identity and varies
expression; departure mode only triggers on explicit signals
(PRODUCT.md anti-references calling out the current surface, or
the user prompt asking for departure).
Adds reflex-reject aesthetic lanes to brand.md as a parallel to
the existing font reflex-reject list. Editorial-typographic is
the first entry. Expands SKILL.md's category-reflex check to two
altitudes (theme + palette from category, then aesthetic family
from category + anti-references).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a configurable lookup path for PRODUCT.md / DESIGN.md / DESIGN.json so
they don't have to live at the project root. Resolution order (first match
wins):
1. process.env.IMPECCABLE_CONTEXT_DIR (absolute or relative to cwd)
2. cwd, when canonical or legacy files are at the root (back-compat)
3. Auto-fallback subdirs of cwd: .agents/context/ then docs/
4. cwd as a default "no context found" location
Existing layouts (PRODUCT.md / DESIGN.md at repo root) keep working unchanged
- step 2 preserves the current behaviour. The auto-fallback covers the two
most common conventions seen in the wild (.agents/context/ for AGENTS.md
auto-import setups, docs/ for the request in the issue) without needing any
configuration.
Changes:
- load-context.mjs: export resolveContextDir() and use it inside
loadContext(); add contextDir to the JSON output
- live-server.mjs: import resolveContextDir and read PRODUCT.md /
DESIGN.md / DESIGN.json from the resolved dir instead of process.cwd()
- SKILL.md: short note on the env var and fallback dirs in Setup -> Context
- tests/load-context.test.mjs: 19 cases covering env var, fallbacks,
legacy migration scope, and back-compat
Legacy .impeccable.md -> PRODUCT.md auto-migration stays scoped to cwd root;
fallback dirs are read-only as far as auto-rename is concerned.
Closes#119
* fix(live): switch live-poll to execFileSync, validate ids strictly
live-poll.mjs built the live-accept invocation with execSync and string
interpolation of event.id and event.variantId. Both fields originate in
the browser; validateEvent only checked truthiness, so shell metacharacters
in either field would land in the shell-parsed command.
Real exploitability is gated by the per-session token (loopback only,
unguessable UUID), so risk is low. The construction itself is structurally
unsafe though, and the fix is small.
- live-poll.mjs: execSync(string) → execFileSync('node', argv). Drops the
hand-rolled single-quote wrap for --param-values; execFileSync passes
each arg as a discrete argv slot, no shell parsing.
- live-server.mjs validateEvent: tighten id and variantId to match the
actual generator shapes (8 hex chars and 1-3 digit numeric strings).
Defense in depth so any value reaching downstream code is inert by
construction.
- live-server.test.mjs: add three regression tests covering accept/discard
rejection of shell-metachar ids and non-numeric variantIds. Update the
three existing fixture ids to match the new pattern.
Reported in #122.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore: refresh pnpm-lock.yaml to match package.json
Cloudflare Pages runs pnpm install --frozen-lockfile and was failing on
ERR_PNPM_OUTDATED_LOCKFILE: the lockfile was missing entries for
@ai-sdk/anthropic, @ai-sdk/openai, @anthropic-ai/claude-agent-sdk,
@anthropic-ai/sdk, @google/genai, ai, modern-screenshot, zod, and had
stale specifiers for jsdom, marked, playwright, wrangler, puppeteer.
Drift was introduced when package.json was last edited without a lockfile
regen. Running pnpm install --lockfile-only resolves it; verified with
pnpm install --frozen-lockfile (clean install succeeds).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Paul Bakaus <paulbakaus@pauls-mbp-3.lan>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cursor Bugbot review on 8660d3a flagged a real corruption bug:
> Multi-line self-closing div breaks depth tracking in expandReplaceRange.
> The forward div-depth walk applies openRe / selfCloseRe / closeRe
> per-line. A multi-line `<div\n className="spacer"\n/>` causes openRe
> to match the opener line but selfCloseRe fails on both lines because
> `/<div\b[^>]*\/\s*>/` requires the full tag on one line. Depth is
> permanently over-counted by 1, so the walk overshoots.
Trace on the JSX-marker-inside-wrapper layout:
- Inside the wrapped element, a multi-line `<div … />` increments depth
at the `<div` line and never decrements.
- Forward walk's depth never returns to 0 → end stays at block.end (the
inner marker comment) → replace range stops there.
- Wrapper's outer `</div>` is left orphaned in the file after
accept/discard, breaking the JSX. Worse: an unrelated subsequent
`<div className="next-card">…</div>` sibling gets its `</div>`
mis-counted as the wrapper close, and the depth walk corrupts further.
Fix: rewrite the forward walk on JOINED text instead of per-line. A
single regex `/<div\b[^>]*?(\/?)>|<\/div\s*>/g` spans newlines (because
`[^>]` matches `\n`), so it correctly identifies multi-line opens,
closes, AND self-closes. Convert the match offset back to a file line
index to set `end`. Walk-back logic for the wrapper opener is
unchanged.
Test coverage:
- New `expandReplaceRange handles multi-line self-closing <div />` test
in live-accept.test.mjs constructs the exact Bugbot scenario: a
multi-line `<div\n className="spacer"\n/>` inside the picked
element AND an unrelated `<div className="next-card">After</div>`
sibling right after. Asserts the discard removes ALL impeccable
markers / wrapper attrs, preserves the next-card sibling intact, and
the multi-line `<div />` survives inside the restored content.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two more Cursor Bugbot findings on commit 11dfad81:
1. `filterByText` short-text returned the wrong sentinel value.
When the trimmed snippet was shorter than 8 chars, the function
returned `candidates.slice()` (all candidates). The caller then
sees `filtered.length > 1` and fires `element_ambiguous` — exactly
the opposite of the documented short-text fallback ("caller falls
back to first-match," which corresponds to `filtered.length === 0`).
So any picker event with a short textContent on a page with multiple
matching siblings spuriously errored.
Fix: return `[]` for short text, matching the JSDoc.
2. `endLine` in the wrap output was wrong for multi-line picked elements.
`wrapperLines.length` counts ARRAY elements, but one element is a
`\n`-joined multi-line string (originalIndented). The actual
wrapper-region row count is `wrapperLines.length + (originalLines.length
- 1)`. Reporting `endLine = startLine + wrapperLines.length` placed
the boundary inside the wrapper for any multi-line pick, giving
downstream agents an incorrect range.
Fix: add the originalLines offset (matching what `insertLine` already
does after the prior commit).
Test coverage:
- `short --text falls back to first-match instead of erroneously firing
element_ambiguous` covers fix#1.
- `returns endLine that includes the multi-line original content offset`
covers fix#2 by wrapping a 5-line <section> in a real HTML file and
asserting the reported endLine points at the variants-end marker (and
the next line is </main>, proving no rows were missed).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Sanity-check on the live-inject unwrap path turned up a real round-trip
bug on HTML files that ship a `<meta http-equiv="Content-Security-Policy"
content="..." />` tag (the leading space before `/>` is the canonical
self-closing form).
Trace:
- The tag-finder regex (`<meta\s+([^>]*?)\/?>`) captures any whitespace
between the last attribute and the closing `/>` as part of `attrs`.
- patchCspMeta did `attrs.replace(content, newContent) + ' ' + marker`,
appending the marker AFTER that captured trailing whitespace. Result:
`...content="..." data-...="..."` — a double space inside attrs and
the original space-before-slash gone.
- revertCspMeta then strips the marker via `\s*${origAttr.full}`, which
greedily eats both spaces — so the round trip leaves `"/>` with no
space, even though the original was `" />`.
Fix: split off the trailing whitespace from `attrs` before patching,
splice the marker into the attribute body with a single leading space,
and re-append the original trailing whitespace. The marker-removal
regex then consumes exactly one space and the trailing space rides
through unchanged.
Test coverage:
- New `round-trips through CSP-meta patch and revert` test in
live-inject.test.mjs covers the canonical Vite shape (CSP meta with
` />`).
- Plus a `round-trips with insertAfter` test for symmetry — the existing
suite only covered insertBefore.
- Existing 4 round-trip tests (HTML, JSX layout, multi-file, column-0)
all still pass byte-for-byte.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Companion to the prior outer-indent fix. live-wrap.mjs's
`originalLines.map(l => indent + ' ' + l.trimStart())` calls
`trimStart()` on every line, which strips ALL leading whitespace and
collapses multi-line picked elements to a uniform indent. So a 6/8/6
shape like
<aside className="card">
<h1 className="hero-title">Hero</h1>
</aside>
was being reindented to 10/10/10 inside the wrapper, and on
accept/discard the round-trip restored 6/6/6 — the <h1> ended up at
its parent's depth instead of nested inside it.
Fix: extract `minLeadingSpaces(lines)` and strip only the COMMON
minimum across the picked lines before reindenting under the wrapper.
That mirrors how `deindentContent` on the accept side already works,
so wrap+accept now form a clean round-trip.
Test coverage:
- Expanded the indent regression test in live-accept.test.mjs to
also assert the inner `<h1>` at 8-space indent and the closing
`</aside>` at 6 — proving the relative depth survives wrap and
discard end-to-end.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cursor Bugbot caught this on PR #118 review:
> JSX discard/accept restores content with wrong indentation. In the JSX
> path, `indent` is captured from `lines[block.start]` — the marker comment
> line inside the wrapper div, which is indented 2 extra spaces relative
> to the original element. But `expandReplaceRange` expands the replacement
> to include the outer `<div data-impeccable-variants>` wrapper, which sits
> at the original element's indent level. `deindentContent(original, indent)`
> restores content to the marker's deeper indent, so all restored lines end
> up 2 spaces deeper than the original element was.
I'd actually noticed the symptom during the live testing session ("some
odd indentation in card-2 after discard") and dismissed it as cosmetic.
Bugbot's analysis matches exactly.
Fix: anchor the deindent base on `replaceRange.start` instead of
`block.start`. For HTML the two are identical (markers sit outside the
wrapper), so HTML is unchanged. For JSX `replaceRange.start` is the
outer `<div>` at the original element's indent — correct base.
Also dropped a duplicate `expandReplaceRange` call in handleAccept that
the earlier edit left orphaned.
Test coverage:
- Two new regression tests in live-accept.test.mjs:
- `discard restores JSX content at the original indent` runs the
real wrap CLI and asserts the restored <aside> opener lands at
its original 6-space indent (was 8 before the fix).
- `accept (no carbonize, raw HTML) restores at the original indent
on JSX` exercises the same anchor on the accept path.
- Inner-element indent loss inside the wrapped content (`<h1>` ending
up at the same indent as its parent `<aside>`) is a separate,
pre-existing wrap behavior — left for a follow-up; explicitly
noted in the test comments.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Same alpha-string trap pattern as the recent detectPageTheme fix, on a
different code path. resolveCanvasBackground walks parents looking for
an opaque background; on a page that doesn't set its own bg the loop
runs out and fell through to:
return getComputedStyle(document.body).backgroundColor
|| getComputedStyle(document.documentElement).backgroundColor
|| '#ffffff';
`getComputedStyle(body).backgroundColor` for a default-bg page returns
the literal string "rgba(0, 0, 0, 0)" — non-empty, truthy — so the `||`
chain short-circuits to transparent-black instead of falling through to
'#ffffff'. modern-screenshot then composites the capture onto a black
canvas; the WebGL shader overlay flashes solid black until the shader
finishes loading.
Fix: drop the buggy fallback. The while-loop already covered <body> and
<html>; if neither is opaque the only sensible answer is the browser's
default canvas color (white).
Test coverage:
- New tests/live-browser-regression.test.mjs pins the anti-pattern
with a static-source check (live-browser.js is an IIFE with no module
exports, so this is the cheapest reliable regression guard). Also
pins the equivalent guard for detectPageTheme's readOpaque helper
added in the prior commit.
- Wired the new test file into `bun run test`'s explicit list.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
While driving the new live loop end-to-end against the repeated-aside
fixture, --text disambiguation silently fell back to first-match instead
of landing on the picked card.
Root cause: `el.textContent` concatenates child text nodes without
inserting whitespace, so `<h1>Hero Two</h1><p>Second card body copy.</p>`
reads as "Hero TwoSecond card body copy." — but the source has whitespace
between </h1> and <p>. The single-space normalization on both sides
missed the join boundary; substring comparison failed; filterByText
returned [] and the caller fell through to first-match.
Fix: filterByText now compares both single-space AND no-whitespace
normalizations on each side, accepting the candidate if EITHER matches.
Bumped the minimum-target-length threshold from 6 to 8 to compensate
for the slightly looser comparison.
Plus two doc clarifications surfaced during the same session:
- live.md now warns that variant CSS using bare `:scope { ... }` styles
the variant wrapper div, not the picked element. Always use a
descendant combinator (`:scope > .card`, `:scope .hero-title`, etc.) —
the fake test agent's CSS is the canonical template.
- live.md documents the agent-side abort path. Aborting an in-flight
generate via `live-accept --discard` only mutates source — the browser
bar stays in GENERATING forever. Use `live-poll --reply EVENT_ID error
"msg"` instead so the browser receives the error SSE and resets.
Test coverage:
- New unit test in live-wrap.test.mjs covering the textContent-without-
inter-element-whitespace shape (three identical <aside> branches each
with <h1> + <p>, picks the second by --text).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Closes#114.
Three orthogonal bugs that surfaced together when live mode picked an
element inside a Vite React/TSX component with sibling branches:
1. JSX wrapper insertion produced invalid TSX
- Replacing a single picked JSX child with [comment, <div>, comment]
yields three adjacent siblings, which oxc rejects with "Adjacent
JSX elements must be wrapped in an enclosing tag."
- A Fragment `<></>` solves the adjacency case but breaks
`cloneElement`-using parents (Radix `asChild`, Headless UI, etc.)
with "Invalid prop supplied to React.Fragment."
- Fix: keep the wrapper `<div data-impeccable-variants="ID">` as the
single JSX-slot child and tuck both marker comments INSIDE it.
accept/discard now expands its replacement range to include the
wrapper's `<div>` open/close lines via div-depth tracking.
2. carbonize produced nested template literals in TSX `<style>`
- extractCss captured `{` / `` `} `` lines from the agent's existing
`<style>{`…`}</style>` template, then handleAccept re-wrapped with
another pair, producing `<style>{`{`@scope…`}`}</style>` which oxc
rejects with "Expected `}` but found `@`".
- Fix: extractCss now strips a leading `{` and trailing `` `} ``
wherever they appear in the captured content (own line OR attached
to the first/last CSS line), so re-wrapping always yields exactly
one `{` ` … ` `}` pair.
3. Ambiguous source matching for repeated JSX branches
- `findElement` returned the first substring match. Multiple
`<aside className="card">` siblings all matched the same query, so
wrap silently landed on the first regardless of which one the user
picked.
- Fix: live-wrap accepts `--text TEXT` (the picked element's
textContent), collects ALL candidates via `findAllElements`, and
narrows by a tag-stripped, JSX-expression-stripped substring match.
Returns `element_ambiguous + candidates[]` when multiple branches
match equally; falls back to first-match when source uses dynamic
content (`<h1>{title}</h1>`) so existing flows aren't broken.
- The fake e2e agent now forwards `event.element.textContent` to
wrap, and live.md tells the agent to do the same.
Test coverage:
- New `vite8-react-tsx-repeated-aside` e2e fixture: three identical
`<aside>` branches, picks the second card's <h1>, runs the full
wrap → Go → cycle → accept → carbonize cycle on a real Vite + TSX
dev server, asserts that Hero One and Hero Three survive untouched
(proving wrap landed on the correct branch).
- Six new unit tests across live-wrap.test.mjs and live-accept.test.mjs
covering the Fragment-replacement design, both leading/trailing
template-literal placements, --text disambiguation, the dynamic-
content fallback, and the element_ambiguous error shape.
- New `runtime.assertSourceContains` fixture hook so other regression
fixtures can assert sibling-branch survivability cheaply.
All 186 unit + static-fixture tests pass; all 21 live e2e fixtures
(20 prior + new TSX) pass with no console errors.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Closes#113.
Picker chrome could become unclickable inside Radix Dialog portals, and
clicking it dismissed the host dialog. Three orthogonal issues surfaced
during manual verification:
1. Modal-aware chrome
- Add `defangOutsideHandlers` and apply it to bar, picker, params
panel, annotation overlay, global bar, and design panel host.
- Sets `pointer-events: auto !important` on interactive chrome so
Radix's `body { pointer-events: none }` modal scroll-lock can't
silence our UI.
- Stops `pointerdown` / `mousedown` / `focusin` propagation at the
chrome boundary so DismissableLayer / FocusScope outside-handlers
never fire for clicks that land on us.
2. detectPageTheme: misread transparent body as black
- `getComputedStyle(body).backgroundColor` returns `rgba(0,0,0,0)`
when no bg is set; the prior regex captured (0,0,0) and ignored
alpha, calling every default-bg page "dark."
- Honor alpha, walk body → html, fall back to
`prefers-color-scheme` only when both are transparent.
3. Exit X invisible on host pages with `button { padding: ... }`
- Every other chrome button sets padding inline; exitBtn didn't.
Host resets like `button { padding: 0.5rem 1rem }` (in the new
fixture, common in the wild) inflated the 24x24 button into 56x40
and pushed the SVG into a non-rendering region — DevTools showed
the right styles, the X just didn't paint.
- Pin `padding: 0` + `box-sizing: border-box`, match the toggle
icon spec (14 / stroke 1.5 / textDim → text on hover).
4. Toast no longer obscures the global bar
- Position the toast above globalBarEl's actual rect instead of a
fixed bottom: 16px that overlapped the bar's bottom: 14px.
Test coverage: new `vite8-react-radix-dialog` fixture exercises the
full pick → Go → cycle → Accept loop with `@radix-ui/react-dialog`
+ `Portal` + `Overlay` + `Content`. Without the fix, clicking Go
dismisses the dialog and unmounts the picked element. All 20 live
e2e fixtures pass; all 180 unit tests pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The configure row's text input filled its background with translucent
magenta (BP.accentSoft) on focus. Composited against the dark bar surface
this produced a murky purple where the browser's default placeholder
gray washed out — flagged in a real session as "godawful styling, gray
text on dark magenta really hurts my eyes". Fix: focus state shows an
accent-colored border only, no fill; placeholder color is set explicitly
to BP.textDim via a one-shot stylesheet so it reads in both themes.
tests/live-e2e/agent.mjs: runAgentLoop's wrapTarget now accepts either a
static {classes,tag,elementId} (test fixture mode) OR a function that
derives the target from each generate event (real-use mode where the
picked element is unknown ahead of time).
tools/live-loop.mjs: standalone runner that attaches the LLM agent to a
running live-server. Used as a test-harness shortcut for validating live
mode out of band; in production the user's coding agent (Claude Code,
Cursor, etc.) plays this role directly via the live skill spec.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
startVariantObserver's "dominated" check only matched when the variant
wrapper was added directly as a mutation's addedNode. SvelteKit (and any
framework whose HMR replaces a whole subtree on edit) adds the wrapper as
a descendant of an added <main> or similar — the observer ignored those
mutations and the session stayed in GENERATING forever even with all 3
variants present in the DOM.
Surfaced by the LLM-agent E2E run on vite8-sveltekit. The fake-agent path
masked the issue because its splice timing happened before Vite's reload
finalized; the slower LLM call shifted timing into the failure window.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Aligning to the design system is now non-optional, drift gets named by
root cause (missing token / one-off / conceptual), and a new Information
Architecture & Flow dimension covers the user-flow shape that polish
previously left to chance. Folds the missing pieces from the deprecated
normalize skill into the v3.0.1 changelog bullet rather than a new bump.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
CSP meta-tag auto-patch (live-inject.mjs)
When the user's HTML carries <meta http-equiv="Content-Security-Policy">,
the cross-origin load of /live.js and the SSE/POST stream back to
localhost:PORT are both blocked. Insert: append http://localhost:PORT to
script-src and connect-src, plus blob: to img-src (the shader overlay),
stash the original content value as a base64 data-impeccable-csp-original
attribute. Remove: decode the marker and restore the original verbatim.
Header-based CSP (Next/Nuxt/SvelteKit configs) intentionally untouched —
those flow through the existing detect-csp.mjs reference path.
JSX-aware accept (live-accept.mjs)
- Carbonize stash now emits style={{ display: 'contents' }} for JSX targets
instead of style="display: contents" (HTML form). React 19 was throwing
"Failed to set indexed property [0] on CSSStyleDeclaration" on the
string form because it iterated chars onto the style object.
- extractCss now matches </style> anywhere on a line, not just at line
start. Previously a JSX template-literal close like `}</style> would
leak the backtick + brace into the carbonize stash, breaking JSX.
- Carbonize stash wraps the CSS body in {` … `} for JSX targets so curly
braces in CSS rules don't get parsed as JSX expressions.
Conditional-render UX (live-browser.js)
- Drop the 2s-then-window.location.reload() fallback in the SSE 'done'
handler. That reload was masking a real failure mode: when the picked
element lives inside conditional render (closed modal, hidden tab,
other-route), Fast Refresh remounts the parent and state resets, so
the variants land in source but never reach the DOM. Reload also reset
state to default, leaving the user stuck.
- Replace with a 6s contextual toast: "Variants ready. If the picked
element isn't visible, retrace the path that revealed it — they'll
appear automatically." The MutationObserver stays armed and
auto-transitions to CYCLING once the variants finally mount.
- Pick-time heads-up: when the picked element is inside [role="dialog"],
[data-state="open"], a multi-tab tabpanel, or an aria-expanded
collapsible, fire a brief upfront toast so the user knows what to
expect if state resets during generation.
Hydration race (live-browser.js)
- SvelteKit (and any framework that hydrates after HTML parse) was
failing post-Vite-page-reload because init() ran resumeSession()
before the variant wrapper hydrated into the DOM. The OLD reload
fallback masked this by triggering a second reload whose hydration
benefited from warm cache. Without that, fix it properly: install a
scout MutationObserver in init() that retries resumeSession() once
[data-impeccable-variants] lands in the DOM.
Shader overlay (live-browser.js)
- WebGL fallback in showShaderOverlay used Object.assign(img.style,
canvas.style, …), which throws on modern Chromium because
CSSStyleDeclaration's indexed properties are not writable. Use
cssText to copy positioning instead.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
typography.md pointed at SKILL.md's `<font_selection_procedure>` and
`<reflex_fonts_to_reject>` XML tags, which were removed in the v3
consolidation and moved into brand.md as the "Font selection procedure"
and "Reflex-reject list" sections. Agents loading typography.md via the
craft flow were chasing content that no longer existed. Now points at
brand.md with correct section names.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
colorize.md: the brand-register paragraph claimed "a dominant color can own
the page" and "accent rate stays ≤10%" in the same breath. SKILL.md scopes
the ≤10% rule to Restrained only; Committed / Full palette / Drenched
exceed it on purpose, and brand.md explicitly encourages those strategies.
Rewritten to defer to the color-strategy ladder.
critique.md: two cross-references still pointed at "Step 4" / "Step 5"
after those headers were renamed to "Ask the User" / "Recommended Actions".
Swapped the references to the new names so the flow is self-consistent.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Merged ten tactical items from ehmo/typecraft-guide-skill into the typography
reference at the upstream author's request: dark-mode weight/tracking/leading
compensation, font-display: optional vs swap, preload-critical-weight-only,
variable fonts for 3+ weights, clamp() max-to-min ratio bound, container/
font-size coupling to preserve measure, text-wrap: balance / pretty,
font-optical-sizing: auto, quantified ALL-CAPS tracking (5-12%), and the
paragraph-rhythm rule (space OR indent, never both).
Skipped: platform-specific tables (iOS/Android/print), confidence markers,
severity-graded report format, academic sources, and the punctuation
subsection (em-dash prescription conflicts with the project copy rule).
Attribution lives in NOTICE.md, not inside the skill content.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Runtime fix in live-browser.js: the 2s static-server fallback in handleAccept
now swaps the outer wrapper with the `[data-impeccable-variant="N"]` div itself
(+ display:contents), matching what live-accept.mjs writes to disk. Scope rules
anchored on the variant attribute keep matching on the non-HMR path, so the
accepted design no longer flashes unstyled until reload. Propagated to all
harness script copies.
/designing:
- §03 Polish redesigned as drenched magenta masthead: commands live in the band,
three title/description columns beneath on cream.
- §04 Maintain redesigned as architectural poster diptych: extract + document
vizzes become the hero element, caption below.
- §05 Interop section removed.
- §05 (was §06) "Pick a register" renamed to "Brand, or product." with a
two-lane hairline-divided layout and an auto-selected framing in the sub.
Live mode status: BETA → ALPHA across the periodic table, magazine spread,
and docs callout, reflecting real-world-testing readiness.
Skill bootstrap: removed the `<post-update-cleanup>` block from source/SKILL.md
(the source repo is the origin; running cleanup-deprecated here would touch
legitimate source). CLAUDE.md and AGENTS.md now document the skip.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Section 7 no longer reads as default-zero: composition-sized targets,
freeform bias toward 1–2 dials on non-tiny surfaces, hard cap of four.
Cross-link freeform to §7 in the action loader; sync all harness copies.
Made-with: Cursor
Unify the design-system panel's data shape around DESIGN.md frontmatter
as the primary source of truth; the sidecar carries only what Stitch's
frontmatter schema can't (extensions + live component HTML + narrative).
Also fix a long-standing build bug that destroyed per-project config.
Shape changes:
- Server /design-system.json now returns { parsed, sidecar, hasMd,
hasSidecar, mdNewerThanJson, parseError?, sidecarError? }. No more
mode switching; both layers ship when present and the panel merges.
- Panel consolidates renderSidecarVisual + renderParsedMdVisual into a
single renderDesignVisual that merges frontmatter primitives with
sidecar extensions.colorMeta / typographyMeta. Helpers for color,
typography, radii model-building. Parsed-md narrative synthesis
survives as a fallback when no sidecar.
- DESIGN.json rewritten at schemaVersion 2: extensions.{colorMeta,
typographyMeta, shadows, motion, breakpoints} + components (with
refersTo pointing back to frontmatter component keys) + narrative.
Token primitives no longer duplicated in the sidecar.
Build fix:
- scripts/build.js:634 wiped .claude/skills/ (and every other harness
dir) on each rebuild, then recopied from dist. After commit b0feed0
unbundled per-project config.json from dist, the sync destroyed the
user's live-mode config on every build without replacing it.
- Added stashPerProjectArtifacts / restorePerProjectArtifacts in
scripts/lib/utils.js. Hoisted PER_PROJECT_SCRIPT_ARTIFACTS to a
module-level export so build.js and readSourceFiles share one
source of truth. Build now preserves config.json across the sync.
Verified in browser: panel renders 10 colors, 9 typography roles, 3
shadows, 6 grouped components, 9 rules, 25 do/don't items, all merged
correctly from frontmatter + v2 sidecar with zero console errors.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two loopholes in the critique procedure let the model shortcut past its own rigor: "SHOULD delegate" with a broad "if sub-agents are not available" escape, and "Browser visualization (when available)" framing that made the [Human] detector-overlay tab read as optional color. Both get rationalized away under context pressure even though the isolation is what makes the combined score honest and the overlay is the user-facing deliverable.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adopt the Stitch google-labs-code/design.md format's two-layer shape:
YAML frontmatter carries machine-readable tokens, prose body covers the
six canonical sections. The sidecar DESIGN.json now extends frontmatter
rather than duplicating it.
- document.md: new frontmatter section, Step 2b staging, sidecar scoped
to extensions, Do's/Don'ts now cite PRODUCT.md anti-references by name,
OKLCH vs hex presented as project posture not mandate.
- design-parser.mjs: tiny YAML-subset reader, exposes model.frontmatter,
schemaVersion bumped to 2, prose-scraping fallback intact.
- live-browser.js: "basic view" CTA copy reflects frontmatter-first model.
- DESIGN.md: add frontmatter with 10 colors, 9 type roles, 7 components;
OKLCH values direct per The OKLCH-Only Rule.
- tests/design-parser.test.mjs: coverage for no-frontmatter, Stitch-shape,
missing-terminator, comment handling.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Config drift was a real tripwire for projects with static generators: new
HTML files get added, never make it into config.files, silently skip
injection. Two additions.
config.files entries now accept glob patterns (**, *, ?) expanded via
fs.globSync in live-inject. Multi-page projects can write
["public/**/*.html"] once and never maintain the list again. New optional
exclude field filters out matched files (email templates, demo fixtures).
HARD_EXCLUDES of node_modules and .git are enforced regardless of user
config so vendor trees can never receive a tracking script.
live.mjs now runs a drift scan after inject: walks common page-source
roots (public/, src/, app/, pages/) and reports HTML files not covered
by the resolved inject targets. Respects user excludes so intentional
omissions aren't flagged. Output JSON carries configDrift: { orphans,
orphanCount, hint } or null. live.md documents the agent flow for
surfacing drift to users without auto-mutating the config.
Unbundle config.json from the distributable skill: it's a per-project
artifact, not skill code. readSourceFiles now skips any PER_PROJECT_ARTIFACTS
during source scan so build output to .claude/ .cursor/ etc never ships
one project's inject targets to another's install. The per-harness
copies stay gitignored via the existing **/skills/impeccable/scripts/config.json
rule; each consuming project writes its own on first /impeccable live.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Without comments or strokes the screenshot is pure visual anchoring,
biasing the model toward the existing rendering and fighting the
three-distinct-directions brief. Local blob still drives the shader
overlay; upload and screenshotPath are gated on annotation presence.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a coarse-controls ("Tune") popover that slides out from behind the
contextual bar via clip-path, showing 2-4 per-variant knobs (range / steps /
toggle) driven by a `data-impeccable-params` manifest. Range/toggle drive
CSS custom properties on the variant wrapper; steps toggle a data
attribute. Values reset per variant; on accept, current values are passed
through live-poll to live-accept as an `impeccable-param-values` comment
for the carbonize cleanup step to bake in.
Other live-UI work in this change:
- Theme-aware palette (barPaletteForTheme) now drives the contextual bar,
action picker, and tune popover. Dark sand on light pages, paper on
dark. Detection has a localStorage dev override for QA.
- Action picker chips get inline SVG icons (wand / bars / funnel / sparkle /
type ramp / circles / grid / devices / curve / star / bolt) stacked
above the label; selection state recolors via currentColor.
- Accept button switched to saturated site magenta with paper text.
- Cycle dots reworked: solid accent for active, neutral for arrived,
hairline ring for pending. No more magenta-on-gray noise.
- Tune chip sits in the cycling row with a count pill badge; open state
uses accentSoft bg + accent text (no ad-hoc white border).
- Popover uses the bar's palette with a deeper surface (surfaceDeep),
lives behind the bar via z-index so a 6px overlap reads as tucked under
it, and animates with clip-path inset() for reliable slide behavior.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Default browser focus-visible ring was a heavy blue outline that
clashed with the dark capsule. Inject a palette-aware inner box-shadow
ring (accentSoft + accent) scoped to the global bar buttons via a
one-time <style> tag. Keeps keyboard focus visible without the
visual noise.
Previously Escape while picking just hid the highlight and set state
to IDLE, leaving the global bar's Pick button visually active. Next
pick attempt fell into a broken state where the button looked on but
no picker ran. Route through togglePick() so the flag, the button,
the UI, and the state all flip together.
Small copy fix on Why panel 04: Figma stamp now reads "last touched
Q3 2025" (was 2024; should reflect closer to the current calendar).
Previous hardening emitted a 7-line todo array and a 10-line ASCII-bar
stderr banner on every accept event, both printed to the agent's
transcript. Per-event overhead added up fast on multi-variant sessions.
Keep the three-layer defence but shrink the per-event noise:
- todo is now a single string: "REQUIRED before next poll: carbonize
cleanup in FILE. See reference/live.md ..."
- stderr is one line with the same pointer.
- reference/live.md keeps the full five-step checklist (loaded once
per session, so its verbosity is a fixed cost — no repetition tax).
Attention signal still triple-redundant: stdout todo, stderr line,
reference section.
After an LLM-triggered session where the carbonize cleanup got skipped
entirely (the instruction was buried as a single bullet among four
cases in live.md, and `_acceptResult.handled: true` felt like a
"done" signal), add three redundant reinforcements:
1. live-accept.mjs now emits a `todo` array on the event payload when
carbonize is true, listing all five cleanup steps plus a pointer to
reference/live.md. The agent reads this as part of the event JSON.
2. live-poll.mjs prints a loud multi-line stderr banner on every
carbonize=true event. Even agents that parse only stdout JSON see
stderr output and can't treat the event as handled without action.
3. reference/live.md pulls the carbonize branch out of the "Handle
accept" bullet list into a dedicated "Required after accept
(carbonize)" section with a numbered five-step checklist, a
rationale paragraph for why skipping is bad, and an explicit
"do not poll again until the file is clean" instruction.
The three layers are deliberately redundant: a future LLM that ignores
the reference text should still be caught by the stderr banner or the
in-event todo, and vice versa.
The inline pre-restore wasn't actually fixing a timing issue — the
fix was the fonts.ready + load retries. Since live.js's own
top-level block runs before DOMContentLoaded and we can do the same
retries there, we don't need an inline script injected into every
user page. Simpler HTML, single source of truth.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
scrollTo(y) clamps to the current document.scrollHeight, which is
several hundred pixels short of the final value until async-loaded
fonts swap in (Cormorant Garamond italic grew consulting-section
layout by ~585px in the logs). The initial synchronous scroll was
clamping to ~6165 even though the Go-time target was 6749.5.
Retry on document.fonts.ready and on the window load event, both of
which fire once the document reaches its final height.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
external live.js is fetched, so by the time it runs the browser has
already queued its reload-scroll animation and history.scrollRestoration
='manual' has no effect. Inject a tiny inline synchronous <script> into
the same block live-inject writes, BEFORE the external live.js tag. It
sets scrollRestoration='manual' and does window.scrollTo(0, savedY)
during HTML parse — before the browser can animate anywhere.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
startScrollLock calls stopScrollLock at the top as a reset. I had
clearScrollY() inside stopScrollLock, so every Go sequence was:
writeScrollY(6749.5) → startScrollLock → stopScrollLock → clearScrollY
— the persisted value was wiped right after being written, so resume
after reload read null and locked to 0.
Move clearScrollY to the three genuine session-end sites (hideBar
error path, confirmed/accept, cleanup/discard). stopScrollLock no
longer touches persistent storage.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three concrete bugs from the diagnostic logs:
1. saveSession was writing scrollY alongside state, so every call during
resumeSession clobbered the Go-time value with whatever the browser
had left us at (typically 0). Move scrollY to its own localStorage
key, touched only at Go and on user-scroll reanchor.
2. history.scrollRestoration='manual' was being set inside init() at
DOMContentLoaded — by then the browser has already started animating
its restore, especially with scroll-behavior: smooth on html. Apply
it at script parse time, and apply the saved scrollY immediately
there too, before the browser's animation starts.
3. Corrections only fired on MutationObserver. A programmatic smooth
scroll (browser restore animation, or another script calling
scrollIntoView) produces zero DOM mutations — so we never caught it
walking scrollY from 0 up to 4800+ in the recorded session. Snap
back on every scroll event, gated by a 250ms user-gesture window so
we don't fight real user scrolls.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Log target-Y at Go, every mutation that triggers a correction (with the
mutation type + added nodes), every correct-or-noop (with from/to/delta),
every reanchor, and every external scroll event >5px. Lets us see which
step is actually moving the page during wrap / variant insert.
Element-based scroll tracking broke every time: Bun's HMR destroys the
target element, the browser's scroll anchoring picks a different nearby
element (e.g. the #downloads CTA) as its new anchor, and the page jumps
to wherever that surviving element is. My element-based correction then
computes against a replaced DOM node with stale / wrong geometry.
The primitive the user actually cares about is window.scrollY — they
want the page to stay where it is, regardless of which element survives
the patch. Pin scrollY directly: capture it at session start, restore it
on every mutation inside the wrapper, re-anchor on user scroll, store it
in saveSession for reload-resume.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two things were wrong. First, I capped large corrections — which was
backwards: a huge delta is exactly when we most need to restore (it
means the browser's own scroll anchoring drifted, which is what makes
the page 'jump to Get Started' when Bun's HMR destroys and re-inserts
our target). Remove the cap so any delta is corrected.
Second, the browser's built-in scroll anchoring was competing with us:
when Bun destroys our target element, the browser picks the nearest
surviving element (like a CTA anchor in another section) as its new
scroll anchor and scrolls to keep THAT stable. Disable overflow-anchor
on html and body for the duration of the session so we own scroll
entirely; restore the original values on stopScrollLock.
Kept the user-scroll grace window (400ms): wheel / touch / arrow keys
re-anchor and suppress corrections, so momentum scrolls don't get fought.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Watching document.body caught every mutation on the page — shader
animations, Bun HMR indicators, tooltips, anything — and fired a
correction on each one, which fought the user when they tried to scroll
mid-session. Now the observer only responds to mutations inside the
session's wrapper. On user scroll intent (wheel / touchstart / touchmove
/ arrow & page keys), cancel any pending rAF correction and re-anchor
to the element's new position, so momentum scrolls don't get yanked
back by a stale correction.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The reload-only fix missed the primary case: Bun's HTML loader hot-patches
the DOM in place rather than doing a full page load, so the resume
codepath never ran and the browser's scroll drifted wherever Bun's patch
left it. Likewise variant cycling (taller → shorter) and agent-driven
variant inserts both mutate layout without a reload.
Add a scroll lock: on Go (and on resume after a true reload), capture the
selected element's viewport-top and install a MutationObserver on body
that re-measures the target and corrects scroll on every batch of DOM
mutations. The target is re-resolved each pass via sessionId + visible
variant, so it survives DOM swaps that invalidate `selectedElement`.
Scroll intent events (wheel, touchstart, arrow/page keys) re-anchor to
the new position so we never fight a user who scrolls during a session.
Also set `scrollRestoration = 'manual'` at init so true reloads don't
land the user somewhere odd before our correction runs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Dan DocaandGitHub
vivshaw