New Work

1. Decide what is already true

Read DESIGN.md, representative code, tokens, components, and assets.

• Redesign: preserve product truth, content, function, constraints, and explicit brand commitments; replace the old visual world rather than polishing it. The old look is evidence of what the subject is, not authority over what it becomes.
• Established world: inherit it. A missing DESIGN.md does not erase a coherent identity already in code; document that identity instead of inventing a replacement.
• Incomplete brand: preserve confirmed assets and recognizable traits, then expand the system with the user for this surface.
• No visual authority: create a new world with the user.

A section, component, feature, or state inside an established surface inherits that surface. Never turn a local addition into a new identity exercise.

2. Ask what will change the work

Before implementation, get the user’s answer through the structured question tool when available. Ask two or three related questions; a precise request needs only a compact confirmation. Skip settled facts, not the confirmation: DESIGN.md settles the visual world, not this surface’s purpose or concept.

• Persuade: who must act, what they should believe, which real proof, content, or assets earn that belief.
• Operate: the task, information, important states, frequency, constraints.
• Read: the reader’s question, source material, structure, wayfinding.
• Experience: what leads, how exploration unfolds, which interaction or transition matters.

Across modes, ask what success looks like, what must remain untouched, and what would make a polished result feel wrong. Never ask for CSS values or canned aesthetic lanes.

3. Choose the right amount of invention

Extend an existing surface

Inherit its world and composition. Resolve only the new purpose, content, hierarchy, states, interaction, and how the addition joins the surrounding experience. No concept tournament, and no DESIGN.md change unless the user approves a durable system change.

Create a whole surface inside an established world

Keep the visual system fixed. Derive five to seven materially different structures from the content, task, and user behavior, ordered by resonance. For a genuinely open whole page, screen, or flow, run:

"<skill-base-dir>/scripts/impeccable" concept-seed --scope surface --mode <mode>

The script deals three of your structures; the dice pick which three reach the user, breaking the ranking rut while the user keeps a real choice. Present them on the decision page as full cards of equal salience, the dealt lead under kicker THE ROLL, with steer and re-roll; the user locks one. No canon card and no pick card at surface scope: the world is settled, so every card visualizes composition, not identity. With image generation and a comp-led default (.impeccable/config.json; the build-path paragraph below), each card declares a comp under .impeccable/mocks/decision/, generated after serving, in reading order, under visualize.md‘s comp discipline. Anchor each comp on the established identity: pass a screenshot of a representative existing page as a reference image (the harness image tool’s input image, or impeccable generate-image --ref) with a prompt that leads with the new surface’s structure and names DESIGN.md’s palette, type, and component character; prose paraphrases of a design system drift, pixel references do not. Without image generation, or under a code-led default, each card carries a wireframe schematic (impeccable serve-question --schema) the page draws itself. Locking a card is the approval and sets the build path: a locked comp builds comp-led with that comp as the approved comp, discharging visualize.md‘s three-option round with no second approval point; a locked wireframe builds code-led, its ambition carried by the direction contract. Never run the script for a local extension or a precisely specified narrow request; shape those directly.

Create or replace the visual world

1. Name the product’s unique mechanism in one sentence, the audience’s real scene, its cultural home, and what this first surface must prove. Note the page this category always ships and its predictable opposite; both are the rut, kept out of the seven-candidate list. A brief that paints its own picture, a product name, a titled artifact, a governing metaphor, adds its literal reading to the rut: spend at most one candidate on it and derive the rest from elsewhere in the audience’s world.
2. From that cultural world, list seven concrete visual systems, artifacts, places, or rituals the audience knows by heart, each with one line on why it resonates and can carry the mechanism, ordered by resonance. The audience’s world includes its graphic and screen traditions, not only its physical objects: the notation, publications, identity programs, data graphics, and interfaces it reads daily. A nameable abstract system (a school of poster, a documentation standard) is as concrete a candidate as any artifact. What would this thing look like as a physical object; what did its world look like before the web? Near-duplicates count once. When more than three of the seven share one material family, the derivation stopped at the subject’s most obvious artifact; dig until the list spans at least three families.
3. Turn that material into complete directions: each joins a reusable visual world to a concrete first-surface experience.
4. Run "<skill-base-dir>/scripts/impeccable" concept-seed --scope direction --mode <mode> and follow what it prints. No substitute, no skip: on a new or replacement world, writing artifact code before this script has run and its assignment is acknowledged is a contract violation, whatever the harness, the model, or the time pressure; the roll is what keeps every run from converging on the category default. The script assigns the direction to build and deals catalog challengers. Fuse each challenger before judging it: the challenger supplies the form and its system grammar, the product supplies every fact, clarity wins conflicts. Weigh fused challengers against the assigned direction on exactly two axes, audience identification and product clarity. Losing to strong grounded material is a valid outcome; beating a thin or tool-monoculture list is the point. Close with a verdict per challenger, decided before any borrowing: wins (beats the assigned direction on both axes; becomes the build candidate), competitive (holds one axis; stays a full alternate), or declined (loses both). A declined challenger is not spent: name the one discipline of its system the assigned direction lacks, and raise the assigned direction to match before presenting it. A donation transfers ambition and system discipline (a palette’s total commitment, a grid’s density courage, a form’s structural honesty), never the challenger’s clothes; a lifted motif is a costume note, not a raise, and one world owns the page. Write each raise into the presented direction as its own line, named for its donor; a raise nobody can read did not happen.
5. Present one direction, fully committed and already raised by the hand it beat, raises visible as named lines: world, first viewport, visitor path, signature interaction, cross-surface reach, honest risk. Route each challenger by verdict: winning and competitive challengers are full alternates with their QUALITY BAR cards and one-line case; declined challengers render demoted, compact and quiet, each carrying its verdict and what the direction kept from it, never full-size, never silently dropped, still adoptable on request. The verdict informs the user’s choice, never pre-empts it; the demoted row is the hand’s proof of judgment. A hand holds at most three full-card challengers: when the roll deals more, the three strongest join and the rest wait in the re-roll pool, noted in one line; dropping a challenger from the hand itself takes a named product-truth failure, disclosed. Add one card for your own top-ranked grounded candidate when it is not the assigned direction, kicker IMPECCABLE’S PICK, same anatomy as every card, with an honest risk line naming its familiarity when true: the strongest grounded direction is often where most runs in this category land, and the user deciding that trade is the point of showing it. Familiar and effective is a legitimate destination, not a failure of nerve; the pick card and the standing exit serve it at two depths. One pick card, never two, never a ranked list: a lineup of your candidates hands selection back to a taste function and invites the safest card. The pick never takes the lead position; when the dice assign your top candidate there is no pick card, and the assigned card notes it topped your list. Add re-roll with an optional one-line steer, in three registers: plain (a fresh hand, same spread), safer (your remaining conventional grounded candidates plus the canon against named competitors), bolder (foreign forms only, at full commitment). The register is the user’s steering on the familiar-to-bold axis, never yours to pre-select; when the answer carries one, re-run the seed with --register <value> and the next --reroll round, and follow what it prints. A user saying “bolder” or “safer” while a direction round is open means these registers, never the bolder or harden commands. The two channels share this structure and differ only in richness: cards and boards on the decision page, names and one-liners through the structured tool, whose option list carries the assigned direction, the pick, the winning and competitive challengers, and the standing exit last; declined challengers fold into the assigned option’s description as their kept lines, so the raise survives the text channel.

The standing exit: every direction round offers one quiet, permanent alternative, the category standard, played straight. It is the user’s door, never yours: never recommend it, never weigh it against the roll, never let it soften the dealt directions; the counterweights bind the unchosen default, not the chosen one. When the user takes it (the canon action, a safer-steer, or plain words asking for the familiar or competitor-like path), convention becomes the commitment: ask once for two or three products this should sit alongside, make their craft level the bar, and execute the canon at full fidelity, without irony or smuggled quirk. Record a standing preference as a brand commitment in PRODUCT.md. Re-roll eliminates every direction already shown, grounded and challenger alike; after two consecutive re-rolls, ask what quality is missing. Re-roll on your own only on named factual grounds, when the assigned direction cannot carry the product’s truth or task; taste is never grounds. The user may re-roll freely, and a user- or brief-pinned direction beats the roll, always. Resolve collisions field by field: preserve every user- or brief-pinned constraint. In dimensions the brief leaves open, the assignment still binds through its topology, controls, state vocabulary, and ritual; when only its materials conflict with a pinned visual direction or PRODUCT.md brand commitment, translate that material expression and name the translation in the presented direction. A look mismatch is not grounds to re-roll. Present the decision visually: write an options payload with the assigned direction leading, its raised lines included; the pick card when one exists; the dealt challengers as alternates with their QUALITY BAR cards, verdicts, and kept lines; re-roll with its safer and bolder registers; steer; canon enabled; and buildPath carrying the recorded default with toggle: true whenever image generation exists (details in the build-path paragraph below). A degraded roll with no challengers still uses the page, as a single text-only card with re-roll. Give every card the same anatomy: thesis, palette, materials, first viewport, honest risk, and the challengers’ case lines (--schema prints the exact shape); the page renders identity from these fields, demotes declined challengers to their row on its own, and a challenger’s catalog image rides as labeled inspiration, never the promise of the build. Author canonCard too: the category standard as one honest card, same anatomy; the page keeps it subordinate, and the counterweights still bind you. Run "<skill-base-dir>/scripts/impeccable" serve-question --start --payload <file> (--schema first for the payload shape). It daemonizes, prints the page URL and a key, and exits; open that URL for the user, in-app browser first, then the system opener, then showing the URL. Collect the choice with --wait --key <key>, repeating while it exits 3; the ANSWER prints as JSON. An ANSWER of {"optionId":"reroll"} keeps the server alive and the page open on a loading hand: rerun concept-seed with the same --scope and --mode plus --from <seed-key> --reroll <n> (1 on the first re-roll, counting up), build the next payload, deliver it with --update --key <same key> --payload <file>, then return to --wait on that key. Never --start a second server or fall back to chat here: either strands the open tab on a hand that never arrives. Exit 4 means the page closed unanswered: re-present once through the structured question tool, and with no answer there either, proceed unattended with the assigned direction and state the assumptions. A harness that can leave a shell blocked in the background may run the script without --start and let it auto-open and block. Never predict the fallback: run the script, and only exit code 2 from starting it routes the decision to the structured tool; that exit is the fallback, never an error to retry.

When image generation exists, every card also declares a comp path under .impeccable/mocks/decision/, the canon card included. Where the harness sandboxes its shell, start the page through the least-sandboxed command path it offers: a sandboxed shell cannot bind the board’s port, and the first-attempt failure costs a retry every session. Serve the page first, then produce the comps; the page shimmer-waits per slot and the user may answer before they land. Each card’s image is that direction’s north-star comp at full fidelity under visualize.md‘s comp discipline: the requested surface’s first viewport, structure-led prompt, real product name and real content, no invented commercial claims, in that card’s own palette, type character, and material world, committed all the way. Generation takes the same time at any fidelity, so an unfinished draft pays comp cost for draft quality; fairness between cards is equal fidelity in each card’s own grammar, one surface, one aspect, never shared unfinishedness. The frame’s aspect is the surface’s own: portrait at device viewport for a native app or mobile-first surface, landscape for desktop web; the decision page adapts to either, and a phone screen comped landscape is a broken frame, not a neutral default. Produce in reading order, the assigned card, then the pick, then the full-card hand, then canon, each file written with its prompt sidecar the moment it is done, so a re-roll’s spend front-loads onto the cards read first; declined challengers get no comp, their catalog thumb is their face. With parallel subagents, fan out one agent per card: each spawn is the shipped asset producer with a single-comp packet, that card’s fields, PRODUCT.md, the shared frame, and the card’s declared path, up to four in flight. Regenerate inline any slot still empty when its agent returns; drop without ceremony any slot still empty when the user answers. No other supervision is owed. Without parallel subagents, generate in the main thread after serving, same order, and let the harness’s own generation display carry the progress; the wait for the answer follows the last file. The chosen card’s comp is not spent by the choice: comp-led, it enters the comp round as compositional option one; code-led, it returns at the finish review as the critique reference, what the image dared that the build did not. Unchosen comps stay in .impeccable/mocks/decision/ as the round’s spent hand; they carry no approval and imply none. With no image generation, cards carry their identity in palette chips and facts, and that page is complete, not a lesser version; the page then also demotes every challenger’s catalog art to a labeled thumbnail on its own, because salience must encode the verdict, never the accident of which cards have images.

The execution contract, comp-led or code-led, is a workflow preference, not a per-surface decision; no round asks it. The recorded default rides every round and the page’s toggle handles the exception. Read the default from .impeccable/config.json (buildPath), the gitignored .impeccable/config.local.json winning where one machine differs from the team’s committed value; with neither, comp-led is the default whenever image generation exists. Author every direction and surface payload with buildPath: { "value": <default>, "toggle": true }; the page renders a footer toggle with the trade stated beside it, and the ANSWER returns buildPath plus buildPathFlipped. A flipped value binds that session only and is never written back, with one exception, the only question this preference ever earns inside a round (init records it up front on projects that get the chance): when buildPathFlipped comes back true on a project that records no buildPath at all, ask once after the round closes whether to keep it as the standing default. Either answer writes .impeccable/config.json; the answer picks the value, never whether to record one. Yes writes the flipped value; “no, just this once” writes the value they flipped away from, the standing default they just confirmed by declining. Ask on the flip, never on the untouched default: a user who left the toggle alone told you nothing. A declined offer nothing writes down is an offer the next session makes again. When the user asks in words to change the standing default, update the file without asking. Comp-led: the chosen card’s comp is law, generated before building when it does not yet exist, and the finish review audits the build against it; boldest composition on the table, fix rounds expected, and the comp is non-optional, no silent skipping. Code-led: no comp of this page and no apology for it; the QUALITY BAR boards still calibrate finish, and the ambition moves into the written contract, the FIRST VIEWPORT block plus a named signature interaction and motion grammar, which the finish reviewer audits in behavior; code-led is not a discount on commitment. A code-led round still declares each card’s comp path as a flip reserve: when the user flips the toggle to comp mid-round, --wait returns once with BUILD PATH FLIPPED while the page shimmers the slots; generate each open card’s comp into its declared path then, lead first, and wait again. The flip back is free, and a comp that already rendered rides at the finish review as the critique reference. Without image generation there is no toggle and no choice: code-led is the only path, stated in one line rather than asked. The old two-card execution-contract round is retired; followup: true remains the general mechanism for delivering any later round over the same table via --update.

Catalog worlds are working systems, not mood references. When one survives, carry its palette and material, type and composition, topology, controls and state, and responsive rules into the product. When the source is itself an interface language, commit to its native grammar across navigation, content, controls, and states. Open the QUALITY BAR board and hero for the world you build the moment the choice lands, even if you viewed another card earlier; the ANSWER line names the chosen card’s images (when the harness only reads files or runs sandboxed, download them into the workspace and open the relative path; sandboxed viewers reject absolute paths outside it). They set the craft level the build must reach, a rendered reference’s finish, commitment, and art direction, never the composition; your surface serves this product.

Every direction the roll can land on must already be viable: every relationship and claim it visualizes true, a real palette and component family, a distinctive composition with one product-specific experience, workable at full-surface scale within the available assets, tools, and performance budget. A candidate that fails on truth is replaced before the roll, never rescued by it. Truth binds claims, not demonstrations: in greenfield work, author whatever illustrative material the concept needs at full fidelity, label it synthetic wherever a visitor could mistake it for the real thing, and hand the user the list of what to replace with real material. What stays uninventable are commercial and factual claims: prices, customers, benchmarks, endpoints, capabilities the product does not have. Refusing a bold direction because its demonstration data does not exist yet is the timidity reflex wearing honesty’s clothes.

For Persuade, the opening must make the offer intelligible and desirable, expose a clear action, and demonstrate something only this product can prove. Conversion lives inside the form’s own vocabulary: a hook that lands in one line, a visible primary action, a legible reading order. A committed form that hides the offer or the action has not finished translating. For Operate, expression may never obscure the task, state, or familiar affordance. For Read, comprehension and wayfinding remain intact. For Experience, the work itself leads from the first viewport.

4. Commit the world

Pick a color strategy before picking colors: Restrained (neutrals plus one accent; the default when the visitor came to operate or read), Committed (one saturated color carries 30-60% of the surface), Full palette (3-4 named roles), or Drenched (the surface IS the color). Persuade and Experience surfaces have permission for the bolder strategies; take them when the brief allows. Color commits at page scale: fields that own whole regions, not accents scattered over a neutral ground. Dark or light is never a default: write one sentence of physical scene (who uses this, where, under what light) and let it force the answer.

Choose faces like objects from the subject’s world, in the mode’s register. Operate and Read surfaces are well served by system stacks and workhorse UI faces; Persuade and Experience surfaces want faces with a point of view, and these training-data defaults mean you stopped looking: Fraunces, Playfair Display, Cormorant, Lora, Crimson, Newsreader, Syne, Space Grotesk, Space Mono, IBM Plex, Inter-as-display, DM Sans, DM Serif, Outfit, Plus Jakarta Sans, Instrument Sans. Naming one of these faces anyway requires a reason no other face could satisfy, and a subject association is never that reason: books wanting a serif, bookshops wanting hand-lettering, and tech wanting a mono are the associations the list exists to break.

Calibration: AI-generated interfaces cluster around a few looks regardless of subject: warm cream ground, high-contrast serif display, and a terracotta or signal-red accent; near-black with one neon accent and glowing edges; broadsheet-editorial hairlines, italic display serif, and small tracked mono labels. All are legitimate when the brief calls for them. Where the brief leaves the aesthetic free, landing in one means the self-check failed: if someone could guess your aesthetic from the category alone, or from category-plus-avoidance, rework until neither answer is obvious. Energy is not the enemy of trust: a brief’s negative constraints (no gamification, no hype) rule out those devices, not exuberance, and adjectives describing the product’s behavior (quiet support, calm coaching) do not dictate the surface’s energy. A bookish, warm, or child-facing subject does not soften the calibration: book cloth, thread, jackets, endpapers, and shelf ephemera span the whole saturated spectrum, and cream paper is the smallest corner of that world; landing on cream plus serif for a book subject is the default wearing the subject’s clothes. A brief-pinned world pins the world, not its softest rendition: the pinned world’s full material range stays in play, and a rendition matching what any model ships for that world failed the self-check at execution rather than selection.

Your measured rendition prior: warm, bookish, family, and child-facing subjects come out as cream grounds, serif display with italic accents, and lamplight, even when the assigned direction never asked for them. Treat that first palette as already spent. Before writing code, reread your OWN-WORLD block: when it says cream, paper, parchment, ivory, or lamplight for a Persuade surface the brief did not pin, the rendition failed and you rework it from the world’s saturated materials first. The same subject renders as bookcloth, thread, jacket, and endpaper color on other models; nothing about the subject requires your default.

5. Record the decision

Before code, record the chosen direction as a development-only contract under ## Direction contract in the relevant surface brief. A direction contract is durable route or artifact strategy, so create or update the brief even when no other surface strategy needs persistence. Use six short blocks, roughly 150 words; do not spend tool calls counting words. THESIS: the one idea this surface owns and the category-default arrangement it refuses. OWN-WORLD: the palette and component language, specific enough to be recognizable with all content removed. STORY: what the visitor understands, believes, and does. FIRST VIEWPORT: the exact composition, what is where and at what scale, and where the primary action sits. FORM: the chosen form, its position on your ordered list, and the seed key the script printed. Close with one more line, FINISH: the run’s exit condition, verbatim “unreviewed and undocumented is unfinished; this build ends with the finish review, the verdict, DESIGN.md, and every shipping raster carrying its provenance”. The surface brief is the reminder later agents reload across edits and sessions: a page that looks complete with the FINISH line undischarged is not done, it is abandoned at the finish line. If a block reads like a mood, the direction is not decided yet; the finishing review audits the render against this contract.

Never copy the direction contract into implementation source or any browser-delivered artifact. This includes HTML or framework comments, hidden DOM, <template> elements, data-* attributes, rendered JSX or TSX output, serialized props or state, React Server Component payloads, client bundles, metadata or JSON-LD, accessibility-only text, and files served beside the artifact. A compiler or optimizer removing development metadata is not a safety boundary. Reviewers and documenters receive the contract from the surface brief.

On a new or replacement world, DESIGN.md is written at finish, from the built world, by the shipped documenter (section 7); a rulebook written before the build gets defended against reality instead of describing it, and hands the design-system detector an unstable target. A new world shipped with no DESIGN.md is still an incomplete run. An ordinary extension does not rewrite DESIGN.md.

Read the existing surface brief before updating it:

"<skill-base-dir>/scripts/impeccable" surface-brief read <primary-target>

"<skill-base-dir>/scripts/impeccable" surface-brief write <primary-target> <body-file> [related-target ...]

After writing, read the brief once more and verify that all six contract blocks and the seed key are present before building.

Keep the brief small: scope and visitor mode; audience, job, action/task, proof/content, and constraints; chosen direction and memorable moment; unresolved decisions. Do not copy global product truth or DESIGN.md tokens into it.

On a comp-led build, whenever any image generation is available (a harness-native tool or the API fallback impeccable context reports), the locked direction is visualized before it is built, never skipped: load visualize.md and follow it, three compositional options put before the user for approval, the chosen card’s decision comp plus two variations. This step is proven to produce the most compositional and ambitious work. On a code-led build the comp round is skipped by contract, never by drift: the ambition it would have carried lives in the direction contract’s FIRST VIEWPORT block and named signature interaction, and the finish reviewer audits those promises in behavior.

For shape, return the selected direction to shape.md and stop before persistence or implementation.

6. Build with full commitment

Build the assigned direction, not a safer interpretation of it. The form supplies structure, reading order, component conventions, and native motion; the product supplies every fact. Commit every atom: nav, buttons, inputs, and links are rebuilt in the form’s vocabulary, and a stock component inside a committed form is a lapse. Land the first build fully committed; the passes that follow exist to make the committed thing clear and effective, never to dilute it. In unattended work, the safe rendition is the known risk.

Comp-led: the comp is a measured contract

When an approved comp exists, it is a spatial contract, not a mood board: only the user can downgrade its authority, in explicit words. Models systematically believe their HTML, CSS, and SVG recreation of an image succeeded when it did not, so the build runs as a state machine on disk whose gates measure the screen against the comp instead of asking you to remember it. Start it once, and let it tell you what is next:

"<skill-base-dir>/scripts/impeccable" build-phase start --direction <seed key> --kind <assigned|pick|challenger|canon> right after the direction choice (this is also the choice ping; the roll’s output names the exact command), or start --comp <approved comp> when a surface round already locked one.

Then, in order, each closed by "<skill-base-dir>/scripts/impeccable" build-phase advance (every verb below runs as "<skill-base-dir>/scripts/impeccable" <verb>; exit 2 means the gate failed and printed why; fix that and advance again; write nothing for a later phase while an earlier gate is open):

0. comps. The comp round from visualize.md: three compositional comps of the requested surface at its own viewport under .impeccable/mocks/, each with a prompt sidecar, put in front of the user; the chosen one’s sidecar gets "approved": true. The gate counts them and reads the approval; a start --comp skips this phase because it already happened. The comp-led path is a frontier-tier job: it asks the builder to hold a measured layout, place plates at their boxes, and act on numeric readings across a dozen attempts. Smaller or faster models produce a recognisable page and stall under the hero gate; if the model in hand is one of those, say so before the direction round and take the code-led path, or expect the run to end at the hero with its readings unmet.

1. spec. Measure the comp: impeccable comp-spec --comp <comp> --grid writes a coordinate grid over the comp; open it, name every salient region by grid span in a regions file (text and control regions snap to the largest ink mass inside their span, so a headline named B1:E4 measures as the headline and not the column beside it; snap: false keeps the span, and an explicit box is taken as drawn) (kind plate / image / texture for anything painted: every illustration, photograph, figure, product object, and material texture; text / control / chrome for what code draws; every region carries a note saying what the comp shows there, which the plate prompt and the gate messages read), and run impeccable comp-spec --comp <comp> --regions <file>. The spec carries each region’s box, sampled palette, and medium; impeccable comp-spec --print is the build’s reference from here on. Type is measured, not guessed: impeccable font-match --measure <text region> reads the comp’s cap height, width class, and weight off the pixels, and impeccable font-match --rank <region> --text "..." takes its candidates from a fingerprint index of the Google Fonts catalog (the nearest faces to the crop’s shape) plus any names you pass with --candidates, renders them at that cap height with the region’s words, and ranks them by fingerprint distance (its USE line is the CSS; its proof sheet shows the comp over the top three); with no browser resolvable it records the catalog’s nearest face and says the size is estimated, which is still the choice to build on. Do not install a browser to rank, and never write a chosen face into the spec by hand: the gate accepts only what font-match wrote. The spec gate refuses to close until the lead text region is measured and ranked. A region note that describes painted material (a diagram, drawing, photograph, texture) under a code kind is refused at the spec: reclassify it as a plate, or reword the note if code really draws it. The script refuses a regions file that leaves comp ink unnamed (callouts, a parts table, a notes block): what is never named can never be missing, so everything the comp shows gets a region. It also refuses a text / control / chrome region larger than a quarter of the comp: that is a column, not an element, and a column scored as one region hides the plates, tables, and notes inside it. Name each element inside it (container: true only when it truly is one undivided element). Anything drawn is a plate: an inline SVG past an icon’s budget (a diagram, notation, leader lines with arrows, a “quick approximation” of the artwork) is refused at the hero; icon-sized SVG (under 64px, a few paths) is fine, and a chart the page draws from data at runtime is a chart, not an illustration. Callout lines and arrows that annotate a drawing belong to that drawing’s plate, with only their labels set as text. A crop of the comp is never a plate (the plates gate refuses a file that is a resample of the comp region: the comp’s grain, its neighbours’ edges, and its resolution would ship as the artwork); the crop is the reference the plate is generated from. A plate region’s box has to hold its whole artwork with a margin: the spec measures the artwork’s contact with the box edges and refuses a box that cuts through it (bleed: true only when the page really crops it there), because a plate placed with object-fit: cover on such a box shows the artwork minus the side the box lost. Anything not in the spec does not exist on the page: no borders, rules, containers, or chrome the comp does not show. Only three concessions exist: fonts (the closest obtainable face), icon glyphs (close enough, exact if the user chose an icon library; this covers the pictogram only, never a control’s chrome, so a chevron, an arrow, a dropdown’s border and fill, a button’s shape are the comp’s), and genuine defects in the comp such as spelling errors.

2. plates. Every raster region ships as a plate: an illustration, photo, or figure regenerated at asset resolution from its comp crop, UI text removed, at its plate path (isolated ink, figures, or objects use native transparent PNG so they sit on the page’s own ground; photos and textures stay opaque); a texture (paper, cloth, grain) is a clean patch of the comp region mirror-tiled to size, generated only when no clean patch exists. impeccable comp-spec --crop <id> writes the reference; save impeccable comp-spec --plate-prompt <id> --background transparent to a prompt file for a cutout, or use --background opaque otherwise. Prefer the harness-native image tool with that crop and prompt, then impeccable embed-prompt <plate> --prompt-file <prompt.txt>. The API fallback is impeccable generate-image --ref <crop.png> --prompt-file <prompt.txt> --out <plate.png> --size <WxH> --quality high --background transparent (use --background opaque for full-frame assets); create the output directory first. Verify actual alpha, white foregrounds, fine edges, and clear holes on light and dark grounds; do not chroma-key native output. The plates gate scores the assets against the comp; also inspect placement and scale visually. With parallel subagents, spawn the shipped asset producer (impeccable-asset-producer; impeccable_asset_producer in codex; /impeccable-asset-producer in Cursor; on GitHub Copilot say “Use the impeccable-asset-producer agent”) with the spec path and let it produce them all; without subagents, produce them here. A crop of the comp is a reference, never a shipping pixel. The gate checks every plate exists, is at least 1.5x the region’s size, and reads as the region. Page code waits for this gate: a page written before its plates exist is a page that draws its material in CSS. A single-file deliverable changes nothing here: the plate is produced the same way and inlined as a data URI. --force exists for one case only, the user downgrading the comp’s authority in words you quote in --reason; the script refuses every other reason.

3. hero. impeccable build-phase scaffold first: it writes the measured layout as CSS custom properties (.impeccable/build/scaffold/layout.css: --r-<id>-x/y/w/h in % of the comp, plus cap height, font-size, family, and weight where measured) and a reference page (hero-reference.html) with every region at its box and every plate placed. Bind the numbers to your own semantic structure, an element per region; the reference is a check on positions, never the page, and overlapping boxes are overlapping boxes. Then build only the first viewport, at the comp’s own dimensions, the comp’s words copied verbatim (the user approved that comp with those words; rewording is a stated decision after the hero passes, never a silent one inside it), every text region sized from its measured cap height and set in its ranked face, plates first: place every plate at its spec box (object-fit: cover, an <img>, a background image, or an inlined data URI named for it) before any text or control, capture into .impeccable/review/hero-repro.png, run impeccable build-phase record hero once so you see the plate regions read as match before any text exists, then lay the semantic layer over the plates from the spec’s palette and boxes and advance. The gate first refuses while any plate is unreferenced by the source, then runs impeccable comp-diff, writes .impeccable/review/diff/hero/ (side-by-side, heatmap, one paired crop per region, report.json), and passes at 72% overall with no hard veto outstanding (a missing region, a contradicted plate or text block, an SVG illustration, a clipped plate, invented ink block at any score); above the bar, the numeric readings become advisories printed with the pass, and the polish pass before responsive is where they get fixed: the gate also reads each text region’s cap height, line count, weight, ink colour, and position against the comp, each chrome strip’s height off its rule, and the frame for ink where the comp is calm (a kicker, an extra nav item, a divider), and says each miss as a number (“cap height 78px in the build, 103px in the comp”); those numbers are the edit. When it fails, open the region crops it lists, in order, before editing: a region scored missing needs its material, contradicted needs its structure re-derived from the spec box, drift is where size and spacing edits belong; the gate refuses a third attempt that only nudges values on the same region. This is where the run’s ambition is won or lost, and a retry here costs minutes where a rebuild verdict at the finish costs the run.

4. sections. Build the rest of the surface inside the spec’s system: the same corner language, line weights, and palette, and nothing the comp never shows. Where the comp does not cover a region, it inherits the recorded system.

5. motion. The signature interaction, reveals, and motion, orchestrated once rather than scattered.

6. responsive. The other viewports, and the first viewport at common desktop widths (1280 to 1600), not only at the comp’s exact size: fluid columns, no fixed-pixel grid that wraps a hundred pixels narrower. Capture desktop.png (1440 wide, full page) and mobile.png (390 wide) into .impeccable/review/; the gate diffs the desktop capture against the comp and refuses a first viewport that only held at the comp’s width. A comp’d surface that is mobile-first was comped portrait; the plates were produced for that frame.

Code-led

No comp and no apology for it: the ambition lives in the direction contract’s FIRST VIEWPORT block and the named signature interaction, and the finish reviewer audits those promises in behavior. The chosen decision comp rides to the finish review as the critique reference.

Both paths

• The first viewport is a thesis, not a header. Demonstrate the mechanism immediately, at the scale the form has in life; do not trap the concept inside a standard hero or card shell. The memory test: if someone left after one viewport, what would they describe an hour later? If the honest answer is a mood, the concept has not committed yet.
• Prove, don’t claim. Show the subject doing its job: the interface at work, the mechanism dramatized, specifics a competitor could not copy-paste. Demonstration data is design material: author it at full fidelity and label it synthetic; claims stay uninventable.
• Author the assets; never substitute chrome. Great surfaces live on carefully made content: names, entries, copy, covers, thumbnails, textures. In greenfield work every blank the ask round left open is yours to author at production fidelity; content is authorable, claims are labelable, no section is omittable. Gradients, glass, generic icon tiles, and many-vertex clip-path polygons where an authored asset belongs are the gap wearing chrome; the detector flags the last two.
• Build the form’s web leverage. When the chosen world names a technique (canvas, WebGL, view transitions, generative motion), build the technique itself, not a static imitation of it.
• Pace the scroll like a studio. Vary density, scale, image, motion, and quiet inside one grammar; a dense passage earns a quiet one, and the page ends anchored by a real close. One spacing rhythm throughout, with more space above a heading than below it.
• Use real, verified imagery when the brief implies it. Search for the subject’s physical object rather than the category; one decisive photo beats five mediocre ones. Verify stock URLs resolve.
• Author motion as material. Give the page the form’s native motion once, orchestrated, rather than scattered hover effects. Bound expensive effects and keep content visible by default.

Preserve semantics, accessibility, performance, responsiveness, project conventions, and working behavior.

7. Inspect and finish

Inspect the surface’s target sizes in one batched screenshot round: desktop and mobile on the web; on a native platform (ios / android / adaptive), the shipped device classes per OS, captured from the simulator or emulator the way the platform reference’s Verifying the build section describes. When the harness reports the user’s actual viewport (an in-app browser’s size, a named resolution), add that width to the set: the width that breaks is the one the user sees first. Critique the render against the user’s request and the direction contract, fix material gaps, and confirm with one final round; two rounds is the ceiling, and fixes batch between them rather than earning per-tweak screenshots. On a comp-led build, run "<skill-base-dir>/scripts/impeccable" comp-diff --comp <approved comp> --build .impeccable/review/desktop.png --spec .impeccable/build/spec.json --out-dir .impeccable/review/diff/final and read its region rows and paired crops as the critique: the side-by-side is the view the build thread never has on its own, and a region it scores missing or contradicted is a fix whatever the page looks like from memory. Never judge fidelity from one full-page thumbnail; it hides exactly the failures that matter. On a Persuade surface, verify the mode did its job: a first-time visitor should know what this is, why it matters, and what to do within seconds, in the form’s own vocabulary.

A capture is evidence only when it is valid, and you validate before you send. Settle or disable entrance motion first: an element hidden by animation timing reads as a missing element and gets fixed into a regression. Capture full-page shots from the document top. Capture the comp comparison at the comp’s own pixel dimensions. Then open every file once and confirm it shows what its name claims: no black or blank regions, no wrong section behind a right filename, no half-loaded state. A malformed capture sent onward costs the whole round; the reviewer answers it with disposition: recapture and nothing it reviewed binds.

After the second inspection round the build thread’s polishing is over: no further defect hunts, micro-edit scripts, or rebuilds here; whatever remains ships through the handoffs, where a fresh context does the finding better and cheaper. On the web, where this harness runs no design hook, run "<skill-base-dir>/scripts/impeccable" detect --json on the changed targets once here, fix what is mechanical, and pass the remaining findings to the reviewer; a hookless web build that skips this ships every tell the hook exists to catch. A native platform skips the detector entirely: it reads HTML and CSS and has no verdict on native code, so the reviewer’s floor check is the only slop gate and the input packet says so. Capture the screenshots into .impeccable/review/, one file per captured viewport (on the web, desktop.png and mobile.png, plus user-<width>.png whenever the user’s viewport joined the inspected set; on native, one per device class, such as phone.png and tablet.png, suffixed per OS on adaptive), creating that directory when the harness does not; the paths you pass the reviewer are its spec, every viewport you inspected is named required in the packet, and that directory is where it looks when a passed path is missing.

Then spawn the shipped finish reviewer, impeccable-finish-reviewer (impeccable_finish_reviewer in codex; /impeccable-finish-reviewer in Cursor; on GitHub Copilot say “Use the impeccable-finish-reviewer agent”), with the original request, confirmed answers, the artifact path, the screenshot paths, the direction contract, existing hook findings, the QUALITY BAR card and approved comp paths (a code-led build has no approved comp; the chosen decision comp rides in that slot as the critique reference, named as such), on a comp-led build the build state (.impeccable/build/state.json), the spec, and the diff directories (.impeccable/review/diff/hero/ and .impeccable/review/diff/final/, whose side-by-side, heatmap, region pairs, and report.json are the fidelity evidence), the craft-floor reference path, and on a native platform the platform reference path(s), ios.md / android.md, both on adaptive, plus one line saying no detector ran, so the reviewer judges in the platform’s conventions rather than the web’s. The reviewer has no browser; screenshots you fail to pass are checks it cannot run. Never read the shipped agents’ definition files before spawning; the harness loads them at spawn, and you owe only the input packet. Wait on any agent with one long timeout rather than a loop of short polls, and spend the wait on the next independent step. Verify the return carries the five contract sections (a recapture return carries one, its recapture list); on an empty or thrashed return, respawn once with the same inputs. This review never runs inside the build thread and never inherits it: spawn the reviewer fresh, with no forked conversation history (fork_turns: 0 in codex); a reviewer that inherits your transcript inherits your framing, your optimism, and your abstractions, and everything it needs travels in the inputs above. Only a harness with no subagent capability at all substitutes a fresh in-thread pass after stepping fully out of the build context, run from degraded/finish-reviewer.md, and a substituted or failed-and-replaced review is disclosed in one line at finish, never silently.

Act on the disposition word; there are exactly four. recapture: the evidence failed, not the build. Recapture what the return names under the capture-validity rules, then run a full review over the new evidence. A review conducted on invalid evidence binds nothing, and a verdict pass may never follow it. rebuild: fidelity failed wholesale, not in patches. Skip the fix batch and execute the rebuild immediately: re-derive the named regions, produce the named assets, and send the result back for a fresh full review, never a verdict pass; a rebuild replaces regions wholesale, so the whole matrix runs again over the recaptures. Tell the user what is happening rather than asking permission to fix a failure. Consult the user only on a second rebuild directive, both verdicts on the table, or when rebuilding would discard content the user approved. ship: nothing is owed; report the verdict at its scope and continue to the documenter. fix: apply the material fixes in one batch, rebuild once, and recapture the same viewports over the same files. A recapture measures positions, loading, and overflow; it cannot measure whether a fix reached the quality the finding named, so send the recaptured screenshots back to the same reviewer for a verdict scoring every material fix resolved, partial, or unresolved (through the harness’s agent continuation; without one, run the scoring fresh from degraded/finish-reviewer.md‘s Verdict Pass). Fixes scored partial or unresolved get another batch, recapture, and verdict. Two rounds is the budget an unattended run ends at; an attended session’s ceiling belongs to the user, so when the second verdict still lists open items, put the table in front of them and let them choose between shipping as it stands and funding another round. Whoever decides, stop the moment a round resolves nothing, and the reviewer’s findings are the only list you work from, never your own re-opened hunt. Do not run a second detector.

A rebuild and a fix round share one asset rule: a raster either round creates or replaces is still asset work under visualize.md‘s Produce section and keeps its provenance like every build raster, and a raster the round abandons is deleted in the same batch. Before either round’s result goes back for review or verdict, run "<skill-base-dir>/scripts/impeccable" embed-prompt --scan <asset-dir...> over the directories the artifact’s rasters ship from and clear every file it reports by embedding what it is missing: the exact generation prompt for a produced raster, the origin for a sourced, stock, or pre-existing one. The scan only reads; deletion is reserved for rasters the round abandoned, never for a file the scan flagged.

Report the final verdict under the reviewer’s own disposition word and at its actual scope. A verdict pass scores the listed fixes and nothing else: “the reviewer scored all three fixes resolved” is a claim it supports, “no material issues remain” is not. A table with open material findings is never announced as a pass, never softened, and never dressed as whole-surface approval when only a fix list was scored. When the user answers a ship with evidence against it, their own screenshot, a named mismatch with the comp, that evidence outranks every capture you made: put their material in the packet and spawn a fresh reviewer for a new full review. Patching inline and self-certifying is how a rejected page ships twice.

After the last correction, spawn the shipped documenter, impeccable-documenter (impeccable_documenter in codex), with the project root, artifact path, direction contract, PRODUCT.md, document.md, and write boundary. Without subagents, load degraded/documenter.md and document.md before writing. Verify the outcome: new worlds and approved system changes require token-bearing DESIGN.md and .impeccable/design.json, not prose alone. Ordinary extensions compare the finished build against the incumbent system, preserve its files, and report the evidence checked; report pre-existing drift without repairing it unasked. Recheck after later edits. Finish only when review and documentation are complete.