Conversation
…t and craft into index-then-pick files - SKILL.md 559 to 331 lines: critical floor, flow skeleton, one Rotation block, one Picks block, fast mode, per-step Do lines - anti-patterns.md becomes a 120-line pre-emit list; audit reporting moves to verbs/audit.md - hero-enrichment.md and custom-craft.md become slim indexes over new references/enrichment/ and references/craft/ per-pick files - Core-15 sweep named in slop-test.md for component scope - pointer cleanup in macrostructures.md and component-cookbook.md; tokens.css header count fixed
…ls.md, no-vision study path - scripts/sloplint.mjs: zero-dep checker for the M-class gates, optional render tier, fixtures; calibrated to 0 FAIL on shipped examples - slop-test.md: every gate tagged M / M-R / R-J / J, numeric fallbacks for gates 35 and 44, posture clause on gate 23 - SKILL.md: critical floor, flow skeleton, script-first Step 7, harness-neutral phrasing, models.md dispatch, v1.2.0 - references/models.md: per-family quirks, loaded only off-Claude - study.md: capability check before image mode, fetch-tool neutral wording - color.md: Colour postures; responsive.md: mechanical-equivalent pointers
…tion contract - custom-theme.md: reflex check at two altitudes, spent defaults, a seven-entry slate, the deterministic draw with wildcards, the scene sentence, four colour postures, a five-block direction contract with a promise-by-promise finish review; loosened triggers; reworked worked examples - scripts/seed.mjs: reproducible draw (FNV-1a over brief + date + reroll) - references/direction-atlas.json: 41 design-history lineages dealt as wildcards
…ntinue
- scripts/variants/{core,serve,start,await}.mjs: zero-dep picker server
(loopback only, port walk-up, polling, atomic request queue, idle
self-shutdown), scaled-iframe grid + single view + keyboard flow,
chip.js overlay for generated routes, static compare fallback
- references/verbs/variants.md: context detection, per-framework route
recipes, shared ceremony + direction-plan table, divergence rules,
the poll ladder, pick-and-continue, riff
…-judge scoring, compare gallery Arms: Anthropic baseline, GLM on Together, Kimi on OpenRouter (DeepSeek and Qwen staged, disabled). Six briefs incl. metric-temptation, component-scope, and vision-probe probes. Floor-only vs full skill packs to measure prompt bloat. run-matrix orchestrates gen, shots, score, judge, report; the repair loop feeds sloplint FAILs back to the model.
…urfaced - one canonical top-of-file order: stamp, critique line, direction contract - theme-axes.md: the 20-row Rotation lookup now ships inside the skill - tier taxonomies unified by name; craft tiers keep construction letters - S1 reworked to stacked-numbered (the old side-column head was gate 54's own auto-fail); carnival + lumen examples brought up to gates 54/55/24; sloplint: mask/filter exemption on gate 22, precision fix + FAIL grade on gate 54, 15-line stamp window - variants playbook matches the scripts byte for byte (request ids, riff shape, fallback wording, flow order); draft-sweep exceptions written into slop-test, custom-theme, and SKILL - study.md: italic roles diagnosed as-is, built roman; preview-examples regenerated for v1.2
Full-pack matrix on Claude, GLM (Together), and the Kimi stand-in: zero mechanical FAILs everywhere after at most one repair round; the floor-only pack collapses on open models (96 FAILs on GLM), so the full skill is load-bearing, not bloat. No-vision probe: every model declined cleanly. Zero invented metrics. Rotation held per arm. judge.py: Together vision fallback chain + User-Agent fix; vision judge itself pending API credit or an OpenRouter key. gen-direct: 64K output ceiling (Claude needed 37K on one brief).
- five-themes cards open a complete full-page mockup per theme - before/after cards open the actual generated pages (before from hallmark-lab, after from the eval runs), copied into _proposals/pages - eval judge defaults to operator mode: no Anthropic key spend, the Claude Code operator scores from the screenshots; --provider together or anthropic for an automated pass
…ger, Arcade Each is a complete self-contained landing page for a fictional brand in the candidate theme's tokens, grounded in the reference sites (Aesop, Vitra, Teenage Engineering, The Row, Notion, Mercury, Panic). All seven pass sloplint clean (0 FAIL, 0 WARN). Linked from the theme cards via the Open-full-page overlay.
- scripts/lint-hook.mjs: reads the PostToolUse payload, lints .html/.css Hallmark artifacts on write, feeds FAILs back via additionalContext, never blocks (exit 0 always), no-ops on non-artifacts - scripts/install-hook.mjs: idempotent settings merge (--global/--print/ --remove), preserves other hooks and permissions - SKILL.md Step 7 note + README section; Claude-Code-only, degrades to the Step 7 sweep everywhere else
Lean specs (126-145 lines each, from the cobalt.md template) for specimen, midnight, brutal, garden, atelier, newsprint, terminal, manifesto, almanac, sport, studio, riso, bloom, coral, aurora, editorial. Each grounded in its real [data-theme] tokens block and theme-axes row: signature moves, anti-patterns, voice fixtures, macrostructure affinity, and honest neighbour-differentiation. Loaded eagerly by Step 3 only when that theme is picked, so zero cost to any build that picks a different theme. Also cleaned em/en dashes from the 4 pre-existing specs.
…> 24) - tokens.css: 4 [data-theme] blocks (full token set; Grid neutrals tinted to pass gate 22; Ledger/Arcade color-scheme dark; Arcade second accent) - main.js: THEMES / ARCHETYPES / THEME_GENRES / COPY entries for all 4 - theme-axes.md: 4 rotation rows; SKILL.md Step 2.6 cluster list; genre files (Grid->editorial, Ledger->modern-minimal, Field+Arcade->playful) - count strings bumped 20 -> 24 across skill, site, README; stale 57->58 gate counts fixed in marketing copy - Arcade gate-47 carve-out in slop-test + anti-patterns (pixel/scanline allowed; fake OS chrome still banned) - site loads the 4 new themes' fonts; all 24 render in the live picker
- scripts/imagegen.mjs: opt-in generated-still hook via Together's images
endpoint (FLUX.1.1-pro default, gemini-3-pro-image optional), cache by
prompt hash, provenance sidecar. No key -> graceful exit 3, falls back
to CSS-art / placeholder (verified). Wired into tier-e-generated.md +
SKILL Step 4 as opt-in, never auto-fires.
- site/examples/{grid,field,ledger,arcade}-01: canonical builds from the
proposal mockups, all sloplint 0-FAIL, referenced by each theme spec.
Lean specs (131-150 lines) grounded in each new [data-theme] block and its mockup: signature moves, anti-patterns, voice fixtures, and honest neighbour-differentiation (Ledger vs Cobalt/Midnight; Arcade vs Terminal/Hum). Arcade spec states the gate-47 carve-out inline (scanlines, starfield, pixel type, cursor are the sanctioned retro signal; fake OS chrome stays banned) and keeps body copy in the readable sans. Catalog token names throughout; page-decorative tokens noted as page-added.
Runs 'claude -p "/hallmark <brief>"' with the skill copied into a scratch project, parses the stream-json transcript for conformance (skill loaded via stamp, files written, reference-read count + load-order, cost), writes into runs/<brief>/<arm>-cli/. Skills do not auto-trigger headless (measured), so it invokes by name: a conformance harness, not an auto-trigger test. Claude arm uses the subscription (no key); GLM/Kimi arms Anthropic-shaped and disabled until their auth tokens exist. Plumbing validated against the real binary here (spawn, stream-json parse both event shapes, run.json); a real run needs a logged-in terminal.
…nce) + smoothness (grafts, section-zoom, decisions) + thumbnails + scoped injection
- scripts: core.mjs picker gains PNG-thumbnail grid (falls back to the
scaled iframe), a Graft button + G key; serve.mjs adds /thumb + /inject
routes and a graft action on /api/pick; new thumbs.mjs (dual-engine,
zero-install); buildInjectJs for dev-only Vite/Astro/SvelteKit preview
injection. All v1 behavior preserved (smoke-tested end to end + browser).
- verbs/variants.md v2 (280->423): progressive-first flow, parallel default
+ analysis-once + shared head, sketch-depth drafts by default (--full),
optional --fast-drafts, compositional grafts, section-zoom after the pick,
decisions.md log, thumbnail step, Vite/Astro/SvelteKit injection recipes.
- eval/variants-bench.{md,mjs}: speed-bench doc + a mechanical timing harness.
- site/_proposals/variants-demo.html: client-only progressive-picker demo
(gradient text removed to stay on-brand); variants-vs.md head-to-head.
Verified in the browser: progressive fill (building -> thumbnail), labeled
counter, guarded arrows, Pick/Riff/Graft, live single-view iframe.
…, roadmap refresh Skill and package to 1.3.0; preview-example version strings to v1.3.0; README variants section rewritten for v2 (progressive reveal, parallel + sketch-default, thumbnails, compositional grafts, section-zoom, scoped injection); ROADMAP marks the five v1.3 items shipped.
- main.js: 4 stray '57 gates' marketing strings -> 58 - all-themes.html: visible H1 '20 themes' -> '24 themes' - gen-cli.mjs header: correct arm ids (claude/glm/kimi) + token env names (ZAI_AUTH_TOKEN / MOONSHOT_AUTH_TOKEN) - ledger spec: radius claim '8px tight' -> '12px card / 8px controls' - arcade spec: note the starfield/scanline vars are page-local decorative tokens (define in :root per arcade-01), not catalog tokens - tier-e stamp example uses the real imagegen default; hyphens
The example was ported from the proposal mockup with oklch(x% 0 0) neutrals; the shipped Grid theme tokens carry a faint cool cast (0.003-0.010 at hue 255). Bring the example's palette in line: 7 FAILs -> sloplint clean.
The picker's single view, the on-page chip, the injection overlay, and the static compare page all wrapped or barred the page they were showing. Replace all four chromes with one shared HmDock factory in core.mjs: a compact glass bar pinned to the bottom-right corner (grid glyph, chevrons, direct number jump with a dimmed pulse for still-generating directions, small-caps title, Pick, and an expandable tray with tuple meta, Riff, Graft, and key hints). The page itself now fills the viewport everywhere. - chip/inject mount the dock in a shadow root so page CSS cannot touch it, and read /api/state to learn titles, meta, and the live direction count, so a riffed direction 4 is reachable even with a stale data-of (serve.mjs gains CORS on /api/state for the routes-mode cross-origin read) - chip gains Graft parity and an optional data-title attribute - single view cycles ready directions only; compare page swaps the fake Pick for an honest 'pick N' chip that copies the chat reply to the clipboard - variants.md updated to the dock (chip contract, picker walkthrough, riff note, static-compare shape, poll-ladder wording); site demo mirrors it Verified in the browser on a scratch run: grid, full-bleed single view, tray, ready-only arrow cycling, pick queued and drained through await.mjs, chip on a direct page, live count refresh, static compare, mobile.
Hallmark stops targeting open-weight models and goes deep on the three terminals where design work happens: Claude Code, Codex CLI, and OpenCode. Same core everywhere (SKILL.md standard), plus a per-harness adapter each harness loads before Step 0. - references/harnesses/codex.md: $hallmark invocation, workspace-write sandbox + network-off reality (picker server asks first, --static fallback), no hooks/subagents (sequential variants + progressive reveal), AGENTS.md as a Step 0 signal, batch-per-direction patches - references/harnesses/opencode.md: /hallmark + skill-tool invocation, allow/ask/deny permission model, parallel variants via its subagents, formatter/LSP interplay, compatibility-path discovery - references/models.md deleted; its harness-agnostic rules live on in both adapters. SKILL.md dispatch now routes by harness, not by model family - agents/openai.yaml: Codex interface + implicit-invocation policy - scripts/install.mjs: detects installed harnesses and installs a lean 1.5MB payload (SKILL.md, references, scripts, agents, LICENSE - never site/ or eval/) to ~/.claude/skills, ~/.agents/skills (+ refreshes the stale legacy ~/.codex/skills copy), and leaves OpenCode to its compatibility discovery; --dry-run / --remove / idempotent - README: 'Open models' section replaced by 'One skill, three harnesses' with the install matrix; ROADMAP and eval/README reframed (Tier B kept as archive only); version 1.4.0 Ran the installer live: all three locations now serve 1.4.0 with the adapters in place (the Codex location previously held a 27MB v1.0.0 copy).
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Unify every version string: skill frontmatter and preview block, package.json, the four preview-examples samples, the eval direction note, the ROADMAP opener, and the site badges/captions (which were still stamped v1.1/v1.3). The PR is retitled to match and sits as a draft on the renamed v1.2 branch.
…ng dead links - Seven theme specs (arcade, cobalt, field, garden, grid, ledger, riso) drop 'canonical build to mirror' for 'reference build: match its register (tokens, voice, motion feel), never its composition; reusing its section order, hero geometry, or grid is a gate-32-grade repeat' - Hum's .btn block becomes a recipe: press mechanics mandatory, skin free; the verbatim-copy mandate is gone. Spring easing scoped to micro press/pop only, and hero-discipline's bounce ban now says exactly that - imagery-kit.md reframed: composition playbook today, hosted assets later; hard rule that no URL outside the (currently empty) manifest is ever emitted; the wash example now shows the constructed gradient equivalent - hero-enrichment's Kit-led route renamed Composed-abstract (built in-page)
Every Hallmark hero shared one geometry (70-90% viewport, bottom-heavy 1.3x padding). hero-discipline.md now opens with a posture table - Settled (the old geometry, still the default), Banner, Poster, Ledge, Corridor, Stage - each with its own min-height envelope, padding rule, alignment allowance, H1-H9 pairings, and genre affinity. Mobile collapses to content height. - hero-discipline.md loads on EVERY page hero now, not only enriched ones (SKILL.md Step 3 Always list), and the posture is picked at Step 2 with nav/footer, rotated (no non-Settled repeat back-to-back, Settled never three in a row), logged (hero: Stage/H2), stamped, and in the Picks block - gate 6: the 100vh/100dvh ban is scoped outside a stamped Poster; the centred-element ceiling binds unchanged on all postures - gate 44a: the padding check reads the stamped posture (Ledge inverts to top-heavy with base-anchored content, an anchored Poster waives it); 44b fold-fit at 1280x800 stays posture-independent - typography.md: define --text-display-s (referenced by the headline brackets but never declared): clamp(2.2rem, 4vw + 0.75rem, 4.2rem) No script changes; absent stamp field = Settled, so every existing build stays legal (canonical example dirs verified 0-FAIL).
The round-1 plan included per-model tic bans for the harness adapters but the item fell out of the execution order; landing it now. codex.md gains a Model-tics-to-refuse section (ghost cards, sketchy SVG filler, feTurbulence grain washes as filler, diagonal-stripe wallpaper, emoji bullets), and opencode.md gets the symptom-keyed version since it runs many models, with the standing line that the Critical floor and gates outrank any model's taste. Refuse-by-default framing: the brief's own words can earn a pattern back, the model's taste cannot. The grain line points at the forthcoming texture.md so budgeted texture stays sanctioned while filler grain stays refused.
The grain/halftone/misregistration/scanline recipes scattered across the riso, carnival, and arcade specs (plus the assets.md feTurbulence recipe) generalize into references/texture.md: five named artifacts with tokened budgets (--texture-opacity caps, one artifact family per page, never animated, pointer-events none), an affinity table (Riso/Carnival/Arcade/ faint-Newsprint earn it; Specimen/Atelier/modern-minimal stay banned), and a custom-ritual hook (B.6b: print-lineage draws may take ONE artifact). Rides along: e7's knob drops its scroll-linked-parallax option (it contradicted the page-wide parallax ban), hero-discipline's grain note stops claiming Specimen/Atelier carry texture (they ban it) and now names the real carriers, and the starfield ban line gains the scoped Arcade carve-out cross-reference so the two files stop arguing. Specs point at the shared recipes; SKILL.md Step 3 loads texture.md conditionally.
motion.md banned scroll-scrubbed animation 'unless there is a specific reason' while tier-c sanctioned animation-timeline outright - a live contradiction with no vocabulary between them. references/ scroll-choreography.md now IS the specific reason: four named patterns (pinned chapter, scrub-linked headline, progress spine, sticky diptych) with Baseline-2025 recipes, each inside @supports + reduced-motion + min-width guards so the base page is always the finished page. Two correctness rules are baked in: the animation shorthand resets animation-timeline (declare after), and pinned-pane children cannot drive their own view() (steps ride a named timeline on the tall track). Budget: one pattern per page (sloplint F8 will count), never on body text, linear easing for scrub, hero orchestration untouched, JS fallback is the IO reveal-once only. motion.md reroutes to the vocabulary, tier-c's decision row defers to it, macros 14/16 and f2 point at it, and SKILL.md loads it conditionally.
references/dark-mode.md: per-token re-derivation table (elevation gets LIGHTER, accent chroma drops as lightness rises, hue never moves, body weight -50), the exact dual-block emission shape ([data-mode=dark] plus a prefers-color-scheme auto block guarded by :root:not([data-mode]), with the no-flash toggle note and color-scheme), and the dark discipline: shadows become lightness or inset rings (glow stays a THEME carve-out, named), cards sit lighter than canvas, one heading colour, never CSS invert on raster, kit multiply assets skip dark. Both modes run the full gate sweep and the stamp records modes: light+dark. color.md's recipe shrinks to a pointer and its 12-16 vs 12-18 paper disagreement unifies at 12-18. Dual-mode stays strictly opt-in: a dark theme is not a dual-mode page. SKILL.md gains the Step 6 bullet and the conditional load entry.
references/data-viz.md: form heuristic first (one value = stat tile never a one-bar chart, > 7 categories = a table, parts-of-whole = stacked bar never pie/donut/gauge), hand-built SVG construction specs (bars <= 24px baseline-zero, 2px lines, 1px solid gridlines never dashed, ONE y-axis never dual), colour adapted to the one-accent discipline (lightness ramp on the accent for multi-series, fixed order never cycled; labels never wear the data colour), honesty riding gate 46 (zero baselines, annotated zooms, internally consistent sample data), the dark note, and the dashboard bar: one figure, one chart, one table - the six-gauge KPI wall is the tell this file kills. Stat-Led, Ledger, Sport, and Almanac point at it; SKILL.md loads it conditionally.
references/brand-truth.md: when a brief names a real company, the brand's colours and fonts are captured mechanically (homepage + same-origin CSS under study.md's remote-safety rules, meta theme-color, names-only font CSS) into .hallmark/brand-spec.md with provenance and a derived-not- guessed attestation, and theming anchors on the spec. Memory hex is a named failure. The spec caches like preflight.json (90-day staleness, design.md still outranks it), the no-fetch path asks once and otherwise builds free of brand claims, and real logos are never redrawn or hotlinked from strangers. Wired as pre-flight signal source 0.5, into the Step 1 brand-colour signal, the conditional load list, and custom B.1's anchor derivation.
references/finish.md joins the every-build list: heading measure by size (ch caps ON the element, from ~20ch at display down to ~45ch at xl, body envelope untouched), typeset punctuation (curly quotes, true ellipsis, non-breaking unit pairs, tabular-nums on comparing numbers), text-wrap balance/pretty, hanging punctuation on pull-quotes, one filled primary per page, the button height budget threaded through the existing 44px floor and 48px coarse-pointer rules via the hit-target expansion recipe, icon discipline (one set, natural grid, no coloured coins, first-line alignment), concentric radii (inner = outer - padding), the mobile-runs- larger idiom (body-adjacent UI steps DOWN at 60rem; display clamps exempt; 16px mobile floor), and motion finish (exits 60-75% of enters, stagger bands 30-50/60-100ms). Companions: typography.md gains the measure table at its sizing section, motion.md's exit ratio widens from ~75% to the 60-75% band so it stops disagreeing with microinteractions.md, responsive.md points at the mobile-larger idiom. The F-tier checks land separately.
…move layout-and-space.md gains § Surfaces directly above Depth: whitespace, then a hairline seam, then a tinted well (paper-2/3, no border - the tint IS the boundary), then a card, reserved for independently interactive or fundamentally different content. Never straight to cards, never white cards on a tinted page, and Depth now reads as the ladder's dressed top rung. Gate 9 names the ladder as its fix vocabulary: different regions climb different rungs. The F7 monotony check lands with the F-tier.
sloplint gains F1-F9, a tier that reports beside the gates without ever counting in the 58: straight quotes and three-period ellipses caught on a new position-preserving masked-prose view (code blocks, comments, and tags space-padded so line numbers stay exact; entity-encoded quotes match too, and doc.text now decodes quot/apos/hellip/mdash for future text checks), missing text-wrap balance on display heads, more than one filled-primary button, icon-in-coloured-circle coins, display-size numerics without tabular figures, one container recipe stamped across six-plus selectors, a second scroll-choreography pattern, and fake-human avatar hotlinks (the tier's one FAIL) with person-keyword stock URLs as the WARN half. gateSortKey groups F ids after the numbered gates; component scope inherits F4/F5/F6/F9. Verified: a fixture trips all nine with correct grades and lines; the canonical examples stay 0-FAIL (their handful of new F-WARNs is the tier working); the legacy sweep holds at exactly the 190-FAIL baseline.
…ore fake people site/imagery/ now ships (serving at usehallmark.com/imagery/ on deploy): 16 photoreal placeholder avatars (768px, varied ages/skin/hair on neutral studio grounds, natural texture, no smooth-mesh AI look; provenance JSON sidecars committed; each paired deterministically with a unisex name so gate 19 never meets another Jane Doe), 4 watercolor washes and 4 single- colour brush gestures on white ground (composited via multiply - no faked alpha; the category skips dark themes), 4 hand-built tileable texture SVGs, 12 hand-built currentColor marks in four geometric families (inline them - img cannot inherit currentColor; wordmarks are a composition of mark + display-token text, since baked-text SVGs cannot use page fonts), and 3 app-screen frames built on Hallmark tokens with their HTML sources committed beside them for one-command re-render. Total 1.7MB, ~$1.06 of Together spend, every generated file QC'd in the browser (two off-brief brushes were regenerated single-colour and pixel-verified). imagery-kit.md's manifest is now the real allowlist with provenance columns; the categories table gains avatar/mark/screens rows and the white-ground multiply contract; the placeholder ladder's tier 1 points at the real kit and tier 4 loses person keywords (F9's territory); the portrait row of the image-need table routes to kit avatars.
…gh, usage capture gen-cli.mjs learns --model and --effort (forwarded to the claude call), a bare no-skill control arm (plain brief prompt, no skill copy), a $3 per-cell budget guard, and --setting-sources project on BOTH arms so the user-scope hallmark install can neither leak into the control nor version-skew the skill arm; stampPresent on a bare cell doubles as the leak detector. The result event's usage lands in run.json (tokens in/out/cacheRead, turns, stop subtype) with the camelCase modelUsage fallback, and cells are named arm-effort-cli so effort variants coexist.
…minal Running the harness from inside a Claude Code session leaks ANTHROPIC_BASE_URL and the CLAUDE_CODE_* host vars into the spawned claude, which then reports not-logged-in even in an authenticated terminal. The child env now drops them alongside the API key. Also removes the six dead cells that run recorded: headless subscription auth needs a terminal login this environment cannot perform, so the A/B results were produced with in-session subagents instead (disclosed in the PR).
Two briefs (voice-cloning studio, pre-traction B2B SaaS) built twice each, once with no skill and once running the Hallmark flow, same model and same subscription. Artifacts, run records, and the skill's own state files land under eval/runs so the comparison is auditable rather than asserted: the two control pages carry 95 sloplint failures between them (off-scale spacing, unowned colour literals, pure #fff surfaces, layout-property transitions, no provenance stamp), the two skill pages carry none, at roughly a third more tokens and a tenth more wall-clock. Both arms respected the pre-traction brief's honesty constraint, so that half of the test was a wash and each run.json records it that way, along with the runner substitution and the effort setting actually used.
The eval run outputs, the internal design proposals, and a scratch QC page for the imagery kit were all riding along in the PR while being linked from nothing. They come out of tracking: eval/runs (the four A/B cells from this session plus the older model-comparison artifacts, all of it evidence rather than product), site/_proposals (44 files of before/after and theme-exploration pages, preserved outside the repo), and site/imagery/_qc.html (a scratch grid I used to eyeball the generated assets). The eval harness itself stays so the tests are re-runnable, and .gitignore now covers the pycache and the QC pages that slipped past it. The four theme examples added this round stay: arcade-01, field-01, grid-01, and ledger-01 are the worked examples their own theme docs link to, not stray test output.
Hallmark was telling good models what to build. This splits it into a floor that binds and a library that offers, and bans the eyebrow outright. The eyebrow. Gate 54 goes from geometry to existence: short inert type before a heading that announces it is banned in every arrangement, and no brief earns it back. The detector inverts from CSS-first to DOM-first, so it now sees flex and block stacks, heads declared inside a media query, class-less headings, and the word "overline". Small type before a heading survives only when it carries a fact, a destination, or an identifier, which keeps captions, field labels, datelines, unit labels and linked chips legal. references/section-entry.md is the replacement: twelve ways into a section, ten of them pointing at material the skill already had. Themes stay themes. A written contract in theme-axes.md says a theme carries material and never a page plan, and all 24 are trimmed to it: 4,070 lines down to 2,861. Lumen loses a mandatory hero apparatus and a fixed section sequence; hum loses 45 lines of global layout rules filed under a theme. Fake brands, invented people and shippable taglines come out. Genre files stop mandating nav and footer archetypes they disagreed with the themes about. Tiers. All 58 gates gain one: 34 carry a Floor clause nothing ships through, 18 are Reflex gates a build with an argument can waive on the record, 6 never block. Fifteen gates split so a real floor stops dragging a taste threshold behind it. Gates 42 and 43 stop ordering a rotation through the catalog and instead ask that the default be named rather than defaulted into. A waiver declares itself in the stamp with guard evidence the linter re-derives; three per artifact, one per gate, Floor never waivable, and an over-cap waiver lets its gate fire so waiving twelve is worse than waiving none. The catalogs stop being a mandatory pick. Step 2 asks for a designed page shape stated in plain words; the named shapes are there when a starting point helps. Anti-repetition moves from "rotate the codes" to "differ from your last three and say how", which constrains sameness without constraining invention. Contradictions. Display line-height had three ranges and now has one owner. The outlier face was capped at two slots by one line and uncapped by the next. Gate 39 hard-coded 44px against finish.md's 32-40px allowance. Gate 14 banned padding animation that floating-nav.md prescribes. Gate 57's theme list had drifted to 20 of 24. F3 and gate 51 fought over overflow-wrap; break-word settles it. Evidence. 201 eyebrows removed across 38 example pages and 32 across the archived test builds, each replaced with a real opening rather than deleted; aggregate failures fell while doing it. The marketing site drops its own section labels and moves hero labels below the lede. A cold build through the revised skill lands 0 FAIL / 0 WARN, names its own structure instead of picking a code, and argues with a Reflex gate rather than waiving it.
…s are true again Four things the last pass left open. Reflex findings no longer masquerade as Floor failures. They report at a new REFLEX grade which still blocks, so nothing drifts silently, but "0 FAIL" is now a true statement about the floor even while a build owes an argument above it. Across every committed page that distinction turns 230 reported failures into 56 real Floor findings and 174 arguable ones, which is the number that was hiding. The waiver cap is gone. Three-per-artifact was the restraint the tier existed to remove: a colour-field poster that wants a black stage, achromatic neutrals, a long measure and one face doing everything is right, and it was being blocked for being expressive. Guards, reason quality and the log are what keep waivers honest. Past three the build gets a NOTE saying what it now is, a house style rather than a set of exceptions, and pointing at design.md. The Floor has two admission criteria instead of one embarrassed exception. Door 1 is "no admired page breaks this on purpose". Door 2 is a named house rule where admired pages do break it but a language model reaches for it unprompted on every section. Gate 54 is the only member, listed in a table that any future addition has to join explicitly. The apology is gone; the reasoning is in the file. Finish findings gained a record form. `answered <check> · <reason>` parses, so a build that reasons its way past gate 38 leaves something the next run can read instead of prose in a comment. Also: the skill stops mandating em and en dashes in page copy, which had it arguing with its own house rule. The Picks block and the Step 5 preview merge into one recital, since the same decisions were being serialised four times. The font catalog and the per-tone voice samples become skip-unless-needed, roughly 200 lines a build. Two waiver bugs found by a cold build are fixed: "intentional design choice" was passing the boilerplate test, and a malformed two-field waiver was being ignored in silence. The 18 hero captures and 14 gallery thumbnails are regenerated against the live pages, so the README and the homepage stop advertising eyebrows the skill bans.
…ould not have reached for Every build now derives its own visual world. Silence used to route to the catalog; it routes to the ritual instead, and the 24 themes become a fast path you ask for by name, by --fast, by an existing design.md, or by component scope. custom-theme.md is promoted and renamed to direction.md, trimmed from 336 lines to 258, and the draw runs on every build rather than on a branch that almost never fired. Nine themes prescribed a display face that is a training-data default, one of them (Source Serif, on Field) already on our own ban list. All nine are re-faced: Specimen to Bodoni Moda, Studio to Erode, Atelier to Sentient, Newsprint to Zilla Slab over Spectral, Sport to Saira Condensed, Editorial to Bricolage Grotesque over EB Garamond, Cobalt to Familjen Grotesk, Hum to Chillax, Field to Literata. Atelier moving off a Didone also settles the Specimen collision the rejection table had been arbitrating by hand. The catalog keeps the saturated faces but now says plainly that they are the first answer, and treats them the way R.2 treats any spent default. Repairs the flip exposed. Gate 57 tested for the presence of a catalog name, so a derived build stamped theme: custom and sailed past the exact failure the gate exists to catch; it now tests the absence of studied-DNA. macrostructures.md still ordered you to pick a shape before writing code. Colour posture was scoped to custom builds only. Two gate exceptions in the playful genre were scoped to a theme name that a derived build cannot claim, and are now scoped to the declared property. The tier map had a worse bug than any of those. Thirteen gates carry a Floor clause and a Reflex clause under one number, and the linter held one tier per number, so six Floor clauses were quietly waivable and seven Reflex clauses were quietly unwaivable. tierOf() now resolves the exact clause first, and gates 7, 24, 38a, 47 and 55 emit the clause they actually found. Gate 47 was also reporting WARN for a Floor violation that reads "fail, always". That surfaced real defects in the corpus: seven italic emphasis words inside roman headlines, the top AI tell, sitting in our own showcase pages. They are now roman accent spans. Two improvised absolutes in the contact sheet use the token that was already declared beside them. The study verb's output page had six hand-written numerals above headings, missed by the eyebrow sweep; the list is a real ol with a counter now. Corpus: 56 findings down to 41, zero gate 54, Floor enforced honestly. Read-set cost of the flip: +303 lines a run (+19.3%), down from +400 (+25.4%). That is short of the +200 target and I am not closing the gap by gutting the palette construction or the font guidance, which is where the remaining weight sits and what makes a derived run worth taking.
.work is a two-column grid holding three items: the counter, the name, and the description. With no placement the description auto-flowed into row 2 column 1, so every project blurb rendered 96px wide and wrapped to seven lines. Pinning it to column 2 puts it under the name where it belongs. The row gap was a second bug hiding behind the first. .work__name already declares an 8px bottom margin for exactly this pair, but the grid's 32px gap stacked on top of it, so the two spacing systems fought and the name sat 40px above its own description. The gap is now column-only, in both the base rule and the narrow-viewport override, and the margin the author wrote is the one that governs. Checked at 1280px and 375px: name and description share a left edge, 8px apart, no horizontal overflow. Still 0 FAIL.
Two mechanisms Phase 1 could not supply, both conditional, neither able to block. Step 5.5 renders the first viewport before composing it in CSS. It fires only when TOGETHER_API_KEY is present (signal 9); with no key the step does not exist and nothing else changes. Two comps at 1280x800 built from the direction contract rather than the brief, varied on one compositional axis so the comparison means something, one approval point, then build. The prompt carries the scene sentence, the paper and accent as plain colour words, the display class rather than a face name, and one negative clause pushing off the category default, because a prompt that reads like the brief returns the category default in pixels and that is what R.1 already spent its budget refusing. The comp gets no vote on the Floor. An image generator will hand back eyebrows, a gradient ramp and a fake browser window because it has read none of the gates. Same rule as an archive exemplar: take the composition, not the compliance. Reproduction is the part that makes the step worth having. A model reliably believes its recreation succeeded when it did not, and the belief is what stops it looking, so the built hero is screenshot at the comp's own dimensions and answered in writing on mass, interval and entry before anything below the fold exists. Diverging from the comp deliberately is fine. Diverging without noticing is the failure. The medium inventory sits under it: every region gets a written implementation medium before any of it is built, and the gate is that a CSS gradient is not a texture. Writing CSS against woven cloth or printed grain is how a committed comp silently becomes a flat page with the same section order. Step 7 gains a reviewer with no inherited transcript. The thread that walks the gates is the thread that formed every opinion it is grading and reads the page as the thing it meant rather than the thing it made, which is the weakest reviewer available and the one every build has used. The new one gets the artifact, the screenshots, the contract and the Floor list, and nothing about why. It answers the question no gate can reach: does the direction contract describe the page in front of it. Every gate is local, so a page can pass all fifty-eight while quietly not being what its own stamp claims. Findings are then scored resolved, partial or unresolved rather than re-reviewed, unresolved is a legitimate outcome, and two rounds is the ceiling. Off Claude Code the review runs in-thread and says so, because the user is entitled to know which one they got. Read-set cost for a run with no key and no subagents: zero. comp.md loads on signal 9 only, and the reviewer brief lives in slop-test.md, which already loaded at Step 7.
The subagent brief told each draft to sweep the Floor tier, which contains the eyebrow ban, while loading neither section-entry.md (the twelve other ways into a section) nor hero-discipline.md (the hero rules, and a sketch-depth draft is mostly hero) nor the genre file nor the direction ritual. Three drafts were being asked to clear a bar with the manual withheld. All five are in the load list now. All three directions derive, the same as any other build. This also makes the paper-band spread satisfiable for the first time: the catalog holds exactly two mid-band themes, both cool and both grotesk, so a catalog-only run could not fill the mid slot twice without colliding on a second axis, and with the last-3 rotation it could not fill it a third time at all. The rule was stated as non-negotiable and was quietly impossible. A derived system lands wherever the band needs it. A theme the user names for one direction still costs that direction its ritual and not the others. Sections 4 and 5 were in the reverse of execution order: the flow says serve the skeleton picker first and generate into it, section 5 said "do this BEFORE the directions finish", and it sat after section 4. Swapped, renumbered, and the five cross-references repointed. Smaller things. Promotion said "fix every FAIL or waive it on the record", which is the pre-tier vocabulary; FAIL is the Floor and is never waivable, REFLEX is the tier a waiver exists for. The port was assumed at 4180 and every file rewritten when it turned out otherwise, while start.mjs prints the real one on every start. SKILL.md advertised three full directions while the verb has defaulted to sketch depth with --full as the override. The manifest example still described a catalog pick. Forty lines of Vite, Astro and SvelteKit dev-config move to variants-injection.md, read only once the user has agreed to the config edit, and one restated picker description collapses to a pointer.
…ping Signal 8 used to mean "derive, but measure the reflex first": the build called out for a consensus triple and argued with it. That put two design systems in one page and neither one won. It now means the opposite. A connected archive owns the brief, the structure, the system and the code; Hallmark skips Steps 1 through 6 and enters at Step 7 over whatever landed. The sweep infers the genre nobody picked, fixes the material half in place, and prints the structural findings it cannot repair without throwing away the page. Gate 20 learns a second stamp so a swept page can record the sweep instead of claiming a build it did not do. The standalone path is untouched. No archive, no signal 8, and the flow runs exactly as before.
…measuring the old skill b1 names no theme, so it must derive, and it did: direction.md was the first reference read, the stamp carries a real seed and a five-line direction contract, and the derived world (the IPA vowel quadrilateral, plotted, on a speech-lab plate) reaches the actual hero rather than stopping at the stamp. Twelve reference reads, zero waste, in the right order. R.1's rejection is legible in the genre note: it refused the radial blooms the atmospheric genre would have allowed, because the reflex check had already spent that move. b7 passes --fast and must not derive, and did not. It read two theme files and never opened direction.md. No seed, no contract, no slate, no draw line. The posture line still emits, which is the color.md repair working: posture is universal now, not a derived-run field. Both artifacts lint clean at 0 FAIL, 0 REFLEX, 0 WARN. Three harness bugs, all of them the flip's doing. stampPresent scanned a fixed filename list of index.html, styles.css and tokens.css, and a build that named its stylesheet page.css scored skillLoaded=false with a correct stamp sitting in it; its regex also demanded Hallmark immediately after the comment open, so "/* Formant Studio · Hallmark derived system" missed too. It now scans every emitted css and html and matches the marker anywhere in a leading comment. The reference-read budget was ten, set when a run read one theme file. A derived run reads direction.md and theme-axes.md in its place, so the floor moved up by one, and a clean derived build measures twelve. Fourteen leaves room for two conditionals and still catches the thing the check is for, which is defensive pre-loading of a forty-file tree. The cost cap was three dollars, hardcoded. A derived build on Opus lands near three and stopped mid Step-7, which scored a complete artifact as error_max_budget_usd and never exercised the gate sweep at all. It is --budget-usd now, defaulting to five. b7 also joins the brief set: the suite had no coverage of the catalog fast path after that path stopped being the default.
…e bug The runs archive copied a fixed filename list, index.html, tokens.css, styles.css, and stampPresent scans that archive. A swept build whose stamp lives in page.css scored skillLoaded=false with a correct stamp on disk, which is the same stale assumption already fixed in stampPresent and in the reference budget, one function lower. The archive now takes every top-level html and css the build produced. The three affected run.json files are rescored from the real artifacts: only the buggy cell changed (inspo false to true); the bare control stays false, which is the leak check holding. Also archives the second three-arm run: bare 56 FAIL with the same six numbered eyebrows, both Hallmark arms clean, and the control now provably offered zero MCP tools.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Hallmark v1.2, one release
Everything on this branch ships together as v1.2: the catalog grows to 24 themes, variants becomes a fast progressive loop with a corner dock, the skill goes deep on the three terminals that matter, and a round of anti-sameness improvements lands across the rulebook.
Themes and rules
Variants v2 + the corner dock
Progressive reveal, parallel sketch-depth drafts, thumbnails, compositional grafts, section-zoom, a decisions log, scoped Vite/Astro/SvelteKit preview injection. Every control surface is now one bottom-right corner dock over the full-bleed page (picker single view, on-page chip, injection overlay, static compare), shadow-DOM isolated, state-fed so riffed directions stay reachable.
Tooling
Verification
Mechanical checks green across the board: all examples sloplint-clean, hook fixtures pass, installer idempotent, corner dock verified end-to-end in the browser, consistency scripts assert the axes/fingerprint/routing invariants. Open items needing a logged-in terminal: Tier A conformance runs, the variants speed bench, one live image generation.
Round 2: the capability layers
finish.md, every build): heading measure by size, typeset punctuation, text-wrap balance/pretty, one filled primary, button height budget, icon discipline, concentric radii, mobile-runs-larger, motion finish. Backed by a new F-tier in sloplint (F1-F9): nine mechanical polish checks that report beside the gates without ever counting in the 58 (fake-avatar hotlinks are the tier's one FAIL).scroll-choreography.md): four earned patterns (pinned chapter, scrub-linked headline, progress spine, sticky diptych) with Baseline-2025 progressive-enhancement recipes, one per page, replacing a blanket ban that contradicted the craft tier.texture.md): grain, halftone, misregistration, scanline, and deckle as budgeted, tokened recipes earned by theme lineage.dark-mode.md): parity by per-token re-derivation, dual-block emission on request, dark shadow discipline.data-viz.md): form heuristic, hand-built SVG mark specs, one-accent series colour, one y-axis never dual, honesty tied to gate 46.brand-truth.md): real brands get fetched values recorded to a provenance spec, never memory hex.site/imagery/(16 QC'd photoreal placeholder avatars with paired unisex names, washes, brush gestures, marks, textures, three token-true app-screen frames with sources), a filled manifest as the URL allowlist, and the fake-avatar services banned by check.Verification: the full example sweep holds at the round-1 baseline, an F-tier fixture trips all nine checks with exact grades and lines, every new reference is wired into SKILL.md's load lists, and the four documented contradictions (scroll ban vs craft tier, e7 parallax knob, hero-discipline's texture misattribution, color.md's dark-band disagreement) are resolved.