Communitygithub.com

unaverage-design

Studio-level design for anything visual, aiming for results that look like a top studio made them, not like generic AI design. For startups and marketing, B2B and B2C: websites, landing pages, product UI, app screens, dashboards, AI chat interfaces, pitch and sales decks, presentations, social posts, carousels, ads, emails, one-pagers, reports, infographics, posters, trade fair, logo and brand, HTML pages or artifacts. Recommends three ranked directions, lets you pick the effort (Quick, Standard, Studio), builds mockups to choose from and checks with measurement and independent reviewers. Writes English and German copy. Use whenever something visual is made or revised, also for 'does this look generic', 'looks like AI', 'make it premium', 'design check', or in German 'wirkt das generisch', 'sieht das nach KI aus', 'mach das hochwertiger', 'Landingpage', 'Präsentation'. Brand, CI and format skills take precedence for colors, fonts and logo. Not for plain text or video.

What is unaverage-design?

unaverage-design is a Claude Code agent skill that studio-level design for anything visual, aiming for results that look like a top studio made them, not like generic AI design. For startups and marketing, B2B and B2C: websites, landing pages, product UI, app screens, dashboards, AI chat interfaces, pitch and sales decks, presentations, social posts, carousels, ads, emails, one-pagers, reports, infographics, posters, trade fair, logo and brand, HTML pages or artifacts. Recommends three ranked directions, lets you pick the effort (Quick, Standard, Studio), builds mockups to choose from and checks with measurement and independent reviewers. Writes English and German copy. Use whenever something visual is made or revised, also for 'does this look generic', 'looks like AI', 'make it premium', 'design check', or in German 'wirkt das generisch', 'sieht das nach KI aus', 'mach das hochwertiger', 'Landingpage', 'Präsentation'. Brand, CI and format skills take precedence for colors, fonts and logo. Not for plain text or video.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/jakob-meyer/unaverage/tree/main/skills/unaverage-design

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

unaverage design

Generic AI design does not come from bad taste but from the statistical mean: the same fonts, gradients, card grids and hero sections, because they are the most likely. Outstanding work from Pentagram to Linear differs from it mainly in one point: every defining element demonstrably comes from the subject (product, name, material, data, people), and exactly one idea carries the whole while the frame stays familiar.

This skill turns that into a process with the user in the loop: understand, agree on the effort, ask for things from the user's world, show ranked directions as mockups, build in the target medium, measure, have it reviewed independently, deliver the best version. For small tasks it stays as light as a craft check.

It is based on research into design history, studios, product and web design, marketing impact, presentations, law and AI genericness, plus about 70 measured reference works (as of October 2026). Details are in references/, tools in scripts/, templates in templates/.

Principles

  1. Impact first. Every result looks finished, confident and desirable at first sight, as if from a studio, even when material is still missing. A strong anchor carries every surface; the dose follows the user's trade-off profile. Process, review and caution serve this impact and never replace it: a correctly reviewed but sparse draft is not a good result.
  2. One tangible idea from the product's world. Every work carries an image idea from an object, a situation or a concrete example from the product's world (the roast log of a coffee roastery, the consignment note of a freight forwarder, the hand signal for a pause at a dental practice), and one concrete example runs as a through-line across all surfaces. Abstract shapes (line, band, dot) and pure typography accompany the idea; they do not replace it.
  3. Subject before style. Every anchor, color and effect has an origin in the project. What has no origin is replaced by an element with an origin and the same job; if it has no job, it goes (principle 7). A style term is never the reason.
  4. Familiar frame, own expression. Navigation, reading order, forms and format conventions stay typical for the category. The distinctiveness sits in the anchor, type voice, color, image, tone and exactly one signature moment. The rest is deliberately calm.
  5. Generate wide, show few, the user decides. Internally several real candidates, externally at most three directions with a reasoned recommendation. Reviewers advise; they do not decide.
  6. Show, don't claim, and stay honest. No invented facts: numbers, customers, logos, quotes, people. The offer is shown: a product interface is built as a draft from the features in the brief, with consistent sample data from the through-line; that is design, not invention. Images prove the claim of their surface (the problem, the use, the result), not just the topic. Missing material gets an honest best-form stand-in that shows the claim; if there is none, a final-form slot instead of an arbitrary topic photo. Missing story copy is drafted from the brief. Placeholders are never the anchor, title or main figure; open facts, mandatory notices and stand-ins go into the review view and the wrap-up.
  7. Every means has a job. Every color, typeface, size and effect can be justified in one word (action, state, chapter, origin). Otherwise leave it out.
  8. Measure and look. No judgment from code alone and none from measurements alone. Every red measurement is confirmed on the image. Norms (contrast, line breaks, minimum type size) apply even when a reference violates them.
  9. Effort is an agreement. Scope, rounds and reviewers are named beforehand and not exceeded. At least two thirds of the time belongs to idea and design, not to review infrastructure; more reviewer calls showed no effect in tests. Anything beyond is named as an add-on before it happens.
  10. Best version, not the last. Every round is saved; the strongest one ships.

Working with other rules

  • Brand and CI take precedence for colors, fonts, logo, tone and mandatory notices. The direction then works through anchor, composition, image rule, density and copy. If a CI requirement is on a generic list (for example Inter as the house typeface or violet as the brand color), it still applies and is listed in the wrap-up as "intentional, CI".
  • Format and production skills (PPTX, banner, video) govern technology, sizes and export. Video goes to a video skill.
  • An existing design brief or a given direction applies. It is translated into parameters, not renegotiated.
  • User rules from CLAUDE.md or memory are gates (for example writing rules). In design copy a single, deliberately set dash is allowed when it is a reasoned design decision and no user rule forbids it; a dash cadence is an AI tell (copy.md 4).

Language and market

  • Talk in the user's language. Messages, questions, the contract and the wrap-up use the language of the user's latest message. Terms the user sees are translated; the German ones (labels of the direction question, contract fields, the wrap-up line "not run or substituted") are in markets/german.md 1.
  • Copy follows the brief. The output language is the one the brief asks for; if it says nothing, the language of the supplied copy, the existing site or the stated market; only then the user's language. Record output language, market and spelling variant (en-US, en-GB, de-DE, de-AT, de-CH) in contract field 4. Several languages: case X6 (use-cases/README.md 7).
  • Write copy in the output language from the start. Never draft in English and translate; draft the headline variants in the output language.
  • Load market modules when they apply:
ModuleLoad when
markets/german.mdany output copy is German (whole file); the user writes German but the copy is not German (section 1 only)
markets/eu-dach-compliance.mdpublisher or audience in the EU, Germany, Austria or Switzerland, and a legal trigger fires (compliance.md 1)
  • No module for the output language: apply copy.md, set lang, run the font check with the nearest character set and list "spelling and typography not checked by a native speaker" in the wrap-up (copy.md 8).
  • House rules stay house rules. Writing rules from CLAUDE.md or memory apply to notes, messages and the wrap-up and, where they say so, to copy style (for example no dashes); they never change the spelling of the output copy. The same holds for the conventions of this skill's own texts (for example US spelling, dashes named by code point).

How market modules are built and loaded: markets/README.md.

Process

StagePurposeQuickStandardStudio
0 Read and classifycontext, use case, register, material, tools, language and market, silent groundworksilentsilentsilent, with competitor scan
1 Kickoffone bundled message: understanding, effort, questions, plus the directions when the user already named thingsone line, 0 to 2 questions3 to 5 questions; top 3 ranked here only when the user already named things, otherwise with the mockups5 to 8 questions; top 3 (optional 4th card) here only when the user already named things, otherwise with the mockups
2 Design contractthe brief on one page, the reviewers' only context5 lines internalinternalconfirmed
3 Directionscandidates, set, mockups, choiceskipped3 mockups (none with a given direction)3 to 4 mockups plus blind comparison (none with a given direction)
4 Buildidea and anchor per surface, material, system, copy, build, gates after every renderdirectone version2 build-outs of the key artifact
5 Reviewgates, generic check, measurement profile, reviewers, dispositiongates plus R1 short reviewplus R1, R2plus R1 to R4
6 Wrap-upresult with evidence, deviations, open items, not run or substitutedyesyesyes
7 Log and learndirection log, skill improvementsyesyesyes

Stage 0: Read and classify

Without a message to the user:

  1. Check the context before asking: prompt and attachments, project folder (CI, logo, fonts, photos, screenshots, data, earlier designs, STATUS or HANDOFF), existing website or product, user rules, other skills, direction log.
  2. Determine the use case through the case matrix in use-cases/README.md and the profile in the group file: sizes, safe areas, output format, typical AI failure, starter set of directions. Minimum type size lives only in use-cases/README.md 4. For decks the deck type (talk, reading or both, then hybrid), for screens the register. Determine output language, market and spelling variant (Language and market above).
  3. Set the register: Operate (product UI), Persuade (website, display ad, hero), Read (report, reading deck, document), Feed (social, social ads), Distance (trade fair, poster, stage). The register limits what a direction may do, for example a type scale of at most 2.5 in the Operate register (consumer apps: directions/README.md 2.1, last row).
  4. Triage for the effort level (see Effort levels) and the direction situation: given, partial (tight CI) or open; goal clear or unclear. The same brand with a logged direction the user chose or confirmed (level Standard or Studio, no special mode) counts as given when that direction has a fit of at least 1 for this case. A new direction on request goes the normal way; a new campaign platform, meaning a new idea across several surfaces, or a fit of 0 counts as partial with the log as its CI. At triage Quick M0 stays and the logged direction replaces the short calculation.
  5. Material check: what is there, what is promised or can be requested, what must never be invented; images on the user's website are listed as candidates for the image-source question by address only (contract field 7), their rights unconfirmed.
  6. Tool check (dialogue.md 7.1): shell, Node 22.12 or newer with puppeteer-core, Chrome, Python with pillow and numpy (find the Python command once and use it as $PY), Swift, network, subagents, question tool, image connectors (they usually cost quota, use only after approval). From this the operating mode (see Environment and tools). A missing tool is never silent.
  7. Silent groundwork for the kickoff (from Standard): a short world inventory of the project and the category default R0. Check whether the user has shown what the look should come from: at least two things they like the look of or want the work to draw on, in words such as "like our ...", "we love ...", "use our ...", or in attachments sent for that purpose. Things that only describe the product, its features, its process, its name or its location do not count, however tangible, and neither do things found on the site, nor brands, sites and named works, which give only a principle (dialogue.md 4.2). In doubt the user has not. For the kickoff form and for Q16 and Q17, "named things" means this check; every single object, document or place from the brief, Q16 or Q17 still enters the world and counts. If the user has and the goal is clear, if the user asks for directions at once, or in the modes Autonomous and No user available, also compute the ranking at prompt level (unknown material counts as requestable, noted as an assumption) and one candidate with a project name per direction in the set (exploration.md 2, 4, 6.4; directions/README.md 11). Otherwise ranking and candidates wait for the answers (stage 1, M1c). In Studio a competitor scan first. After the answers the ranking is computed or recomputed; if the set or the recommendation changes, say so in one sentence before the mockups.

Stage 1: Kickoff

One single message, everything bundled. Form by situation:

SituationFirst message
small task (triage Quick)scope in one sentence, at most two questions (usually an asset request, else logo and color), start at once
direction giventranslate the given direction into parameters, name the effort in one line and have both confirmed; if it sits on a known AI look, say so once, implement it and propose a signature element of its own, taken from the subject
tight CIlike the two rows below, but the directions stay within the CI and differ in idea and anchor, not in color
direction open, goal clear, no things named yeteffort, questions on things and facts, assumptions; no directions yet (M1c). The directions follow from the answers, as mockups in the next message
direction open, goal clear, the user named things or wants directions at onceeffort, top 3 ranked, your own idea, question block with default answers (M1)
goal uncleareffort, at most three goal questions, in the same message Q16 and Q17 when no things are named yet, because a ranking without a goal would be a guess; the next message is then the mockups, or M1 when things were named

Things before directions. Directions built before the user has named things come from the model's own world and rarely change after the answers, so the first message asks instead of proposing. Every named thing then enters the project world, and at least one mockup of the set shows one (stage 3).

The direction question. Exactly three directions (in Studio optionally a fourth card: the user's own idea or a second seed hit) that cover three distances: near the category (in best form, never as a generic canon), at the edge of the familiar, bold. Per direction four lines without jargon:

  1. Name from the project ("Weser view", not "Swiss Minimal"), catalog code in brackets.
  2. What you see: anchor, type in everyday words, color with its role, emptiness or density.
  3. Why for you: tied to something the user said, the brief or the material, never to taste.
  4. Risk or what you give up, plus the distance.

In German the labels are fixed in markets/german.md 1.

A recommendation is mandatory, with criteria and risk next to it, because a recommendation acts like a default. Recommend rank 1 of the ranking. Exceptions: if rank 1 is the category default without an own asset, recommend the best direction at the edge; a direction whose material is missing is never recommended; in risk categories (law, finance, health, insurance, critical infrastructure, public sector) the recommendation never comes from the loud group V (MAX, TYP, RAW) unless the user explicitly asks for boldness; then recommend the best direction outside group V and show the group V direction as the bold option. The recommendation comes first in the list, never in the middle, because otherwise the middle option wins by position alone. If the recommendation changes after the answers or the mockups, say so in one sentence with the reason. Next to it there is always the path "your own idea": one sentence is enough, it gets turned into a direction. Directions without material are downgraded, and the user learns what brings them back up.

Effort like a quote: level, number of directions and mockups, reviewers and reviewer calls, rounds, decision points, rough duration, plus the relative effort (Quick about a quarter, Studio about twice Standard).

Questions only when the answer changes direction, effort, material, the words on the page or a gate. Each has a default answer for silence, except questions on things and trade-off pairs, where a suggested answer steers the user. Never ask with style terms ("Swiss or brutalism?"); ask for things, not for sites or adjectives (Q16 to Q19 in dialogue.md 3.3; only the reference question Q9 names sites). A trade-off pair is a forced choice ("What would be worse: too calm or too loud?"); after M1c or M1a it comes with the mockups, when the user can see what it means. Without an answer and without a dose in the user's own words the dose is medium (dialogue.md M3).

With a question tool (for example AskUserQuestion in Claude Code): first the top 3 as text, then one call with up to four questions: effort (recommendation first), direction path ("all three as mockups", "only [name rank 1]", "only [name edge]", "only [name bold]"; the free answer covers the user's own idea), one material question (Q5 when it applies, the recommendation first, dialogue.md 3.3), one trade-off pair. M1c, M1a and M1 without named things use no question tool: everything goes out as one numbered text block and the turn ends there, because the tool returns as soon as its options are answered, the open questions would stay unanswered, and options would suggest answers. Without a question tool a numbered text block.

Message patterns for all situations, question catalog, trade-off pairs and asset request: dialogue.md.

Stage 2: Design contract

One screen page following templates/design-contract.md: task, goal, audience and scene, requirements (including output language, market and spelling variant, and all brand colors with hex and role), exclusions (including the category default and its predictable opposite), trade-off profile, material, chosen direction with thesis, mode (for example STI Stage, Quiet or Catalog), base plus accent and signature element, success criteria and gate exceptions, user decisions with date, seed and ranking, and in field 11 the operating mode with missing tools and their substitutes. What is not in the contract no reviewer can check; without a mode the review measures against the wrong corridor. In Studio the contract is shown, because changes cost nothing now.

Stage 3: Directions

From Standard, following exploration.md:

  1. Project world (8 to 15 concrete parts) and, for existing brands, the heritage inventory, deepened from the silent groundwork. Things and words the user named come first, word for word, with the user as origin.
  2. Name the category default R0 and its predictable opposite; both are excluded as the main direction. Priority for R0: competitor scan before the case row before the category table; if the industry is clear, the more specific category table applies; when choosing an industry, the buyer's industry counts.
  3. Ranking following directions/README.md 11, recomputed with the answers: register and material gate, then effect (W), use case (C), project-specific distinctiveness (D), audience (A), material (M), brand (B). Rank 1 only with material (M from 2). Three ranks in the three distances. An external random seed ($PY -c "import secrets; print(secrets.randbelow(10**6))") fills the bold rank in Standard and Studio: pool sorted by score, hit equals seed modulo pool size; after the answers the same seed applies to the recomputed pool. Never invent a random number yourself; without Python use the fallback in dialogue.md 7.1, without code execution 7.5.
  4. Per direction 2 (Standard) or 3 (Studio) project-specific candidates with core idea, source in the subject, anchor, signature element and "designed against"; every candidate has a tangible image idea from an object or a situation of the project world that becomes visible on the key surface (a primitive, a typeface or a color alone is not an idea); from different starting points and at least three material families, at least one from a forced constraint, for campaigns and ads at least one as a functional idea or asset as idea. Rate each candidate obvious, own or surprising; only candidates rated "own" or higher go into a mockup. Swap test: if it works almost unchanged with another name, sharpen or discard it. At least one mockup of the set shows a thing the user named, when there is one; an object, a document or a place from outside the category may supply anchor and form (dialogue.md 4.2).
  5. Mockups in the target medium: the same key surface, same content, same maturity, plus a system strip (fonts with role, color roles, imagery, signature element). Files A.html, B.html, C.html with a randomized assignment; logo and wordmark with class logo or wordmark for the cover-the-logo test. The three differ on at least two of five levels (anchor type, type voice, density, color strategy, signature moment), not only in color or light and dark.
  6. Show the contact sheet (pixels.py sheet), the spectrum in words, then the ranking with fit, distinctiveness, risk and "what you give up". The user chooses, mixes along system levels, says "none, safer" or "none, bolder", or brings their own idea.

Stage 4: Build

Idea and material first, then system and build. Generic results happen when text is set first and then ordered with boxes.

  1. Claim, anchor and idea per surface as a short list before building: key message in one sentence, anchor (the element seen first that is not body text), source of the anchor. For sequences a through-line from one concrete example of the product's world (a case, a customer, a day, an object) that runs through all surfaces; every surface answers the question the audience has after the previous one. The image proves the claim: the anchor shows what the sentence states (for "capsules end up in the general waste" the general waste with capsules, not a cup of coffee). Proof on the key surface (hero, product post and ad in Persuade and Feed, solution slide in the deck): the offer itself or its use. Software as an interface, built as an HTML mockup from the features in the brief, with the through-line as sample data (one case runs through the main steps of the product); an abstract diagram does not replace it. A physical product as a photo or cutout in the situation it is made for (own or stand-in); a service as a situation; on websites a free photo never carries the hero alone unless the user asks for it (images-and-assets.md 2). Type, icon or flat drawing alone are no proof; directions without photos (TYP, MAX, RUL, ORD) still carry it at this point (craft.md 5.1). Scope follows the brief: Studio deepens the requested surfaces and delivers no additional applications; add-ons and open variants appear as an offer in the wrap-up, never in the delivery. The core function of the brief is visible at a glance: if the brief asks for a status display, an overall value stands visibly in place, not only in small detail marks. When the product produces, fills in or changes something, the anchor shows its contribution as a visible difference (marked with a legend, or before and after), not only the finished result; the before state follows the ledger (craft.md 12.1).
  2. Plan material following images-and-assets.md within the image source chosen in the kickoff (section 2 there): the ideal motif first, then sources in order: the user's material, self-built (product interfaces, data forms, objects, illustration), public-domain archives, free libraries with a license, image generation only after approval and never for people, product UI, logos, data or events. If something is missing that only the user can supply, ask for it bundled and concretely and build a best-form stand-in until then that shows the claim: a free photo of material, ingredient, place, situation or problem (never passed off as real staff, customers or as the real product), the product as an illustration from its description, a typographic or data anchor. Stand-in photos are sharp (at least 1.5 times the output size, never upscaled), contemporary and without filters over everything (sepia, vintage, grain); self-drawn objects only in illustration quality. If no available stand-in shows the claim, the motif gets a final-form slot (an area in the brand color with an outline or crop of the motif, never gray, never the anchor of the whole work); that is better than a topic photo without a claim. Illustration only with its own handwriting from the project world (one technique, one stroke, one texture with an origin); flat stock vector style with soft gradients and circle rings is generic, then rather a photo stand-in with appetite or touch. Never paint or composite the product into a foreign photo (a device pasted onto a stock photo, soft fake 3D); then illustration as a whole or a final-form slot. An image appears at most twice in full form per sequence. Labeling is small (label in the review view, list in the wrap-up).
  3. System before the first surface from the datasheet of the chosen direction (directions/), in the mode from the contract: type roles, color roles with saturation strategy, spacing scale, grid, radii, as tokens with roles. Fonts from the direction's pairings in fonts.md 4, chosen from the project world; check a font outside the catalog first with font-check.mjs --lang <languages> (characters of the output language, tabular figures; fonts.md 3). Starter palettes of the datasheets are placeholders: the accent comes from brand, material or the counter-move to the competition.
  4. Copy pass, when the user supplies no finished text (copy.md 1): benefit and feeling of the audience first, in their words (what they gain or do not want to lose); three headline variants, test questions "could this sentence be on a competitor's site?" and "is the headline understood positively on its own?" (in the feed and at the top of a slide people read only the headline; a pun that says the negative first fails). The audience's trade vocabulary instead of buzzwords, microcopy and mandatory zones in the same voice. Draft missing story copy (why now, comparison, positioning, teaser) from the brief and the category and mark it as a draft in the wrap-up; only facts stay open, and an open number never becomes the headline. Story copy introduces no new facts (laws, deadlines, market figures, competitors): what is not in the brief is researched and sourced or left out. In the feed and in B2C the headline hits the audience's feeling or ritual, not the category or the news value ("New in stores" is no headline). In the feed the caption adds to the image (context, benefit, next step) instead of repeating its headline or explaining what is shown. Outsourcing never replaces the content of a surface: "link in bio", "on request" or "more in the meeting" add to a surface, they do not fill it; if the content is missing, it appears in final form (point 6). Every key figure appears at most twice (stage and closing; in a deck the closing slide uses the second appearance, craft.md 12.2). The product is recognizable as what it is: the core property from the brief (for example waterproof, foldable, tool-free assembly) stands in the headline or anchor, not only in small print, and color and form in the image fit the product. No on-image hedges: a company's own statements need no note in its own channels; sources only for third-party figures, collected in a footnote or appendix; notes like "depiction based on assumptions" belong only in the review view (a made-up line on a sample case is content; craft.md 5.1, point 9). Weighting follows the brief: what the brief calls a side note ("also available decaffeinated") never becomes the loudest element; what belongs together stays together. Decision-critical sections always stay (for example construction software: data location and rights to the photos; freight forwarder: liability and transit times; dental practice: costs and insurance coverage): known facts from brief and answers, open items in final form following point 6, asked about in the kickoff; never leave them out and never evade ("details on request"); their heading promises only what the section answers (copy.md 1, point 10). Quotes only approved, with person and role. English copy: headlines in sentence case unless brand, brief or a fixed genre demand otherwise; one spelling variant throughout (en-US or en-GB, copy.md 2 and 3); no buzzwords, negation parallels or triple cadences (copy.md 5). German copy follows markets/german.md 2 to 4. Check environmental claims against the word list and mandatory buttons against the prescribed wording where a module has them (markets/eu-dach-compliance.md 6 and 3); otherwise the triggers in compliance.md 3.
  5. Build impact in the Persuade and Feed registers (craft.md 5): the anchor carries about 40 to 60 percent of the visual weight, a clear size jump between display and text, color as area rather than line, where the direction allows at least one level of depth (overlap, crop, cutout; on scrolling pages a crop only at the outer edge of the viewport, never at an inner section edge, craft.md 5.1, point 5), composition changes with the content type. Sequences (deck, series, carousel, landing page) have rhythm (5.5): at most three surfaces in a row with the same composition, stage surfaces for title, proof, strongest figure and closing, no unreasoned empty third. Calm directions change too: a deck has at least two full-bleed surfaces (color from the brand world or an image bleeding off the edge) and one color ground, a series changes ground or light and dark; a signature element or a component appears at most three times in full form, otherwise as a crop or detail. Slide titles in a talk deck at least 56 px at 1080 height, in a reading deck 44 px. The hero headline in the Persuade register holds against the image: a strong weight (500 or more, or a display cut) and at most four lines on desktop. The dose from the trade-off profile is binding: if the user says "too calm would be worse", it gets loud.
  6. Two views: the presentation view is free of production notes and decorative gap markers (underscore lines, square brackets, "open", "to follow" on every surface): a missing secondary detail is dropped when the surface works without it (source, period, contact line), or phrased with what is known. Open items of decision-critical sections (point 4) appear in final form: phrased with what is known, as complete content in the typography of the content (a card, row or field such as "Liability" or "Insurance coverage" stands only with its value), never as a "to be added" label, dashed box or footnote. A label, row or field without a value never appears in the presentation view; it counts as a placeholder (V26 in generic-catalog.md 6), and the open value goes to the review view and the wrap-up. Open items in final form stand on no stage surface (title, proof, strongest figure, closing), at most one per surface and three per sequence; everything else only in the review view and the wrap-up. Mandatory notices stand where the real medium has them and only when their trigger is in the medium (unit price only next to a price, where the market requires it; compliance.md 2). The review view (CSS class review-view on body or a file open-items.md) shows all mandatory notices, stand-ins, draft copy and open items. The presentation view is what gets rendered, judged and delivered; the review view goes along in addition.
  7. Build in the target medium. Craft rules in craft.md, media rules in the profile of the use case.
  8. After every render the gates (stage 5). In Studio two build-outs of the chosen direction for the key artifact (another anchor or another idea), a blind pair comparison by R1 in both orders, the stronger one is refined.
  9. Reduce: remove at least one element that has no job; reducing never means weakening the anchor, the idea or the impact.
  10. Brief check before the review: list every statement of the brief (audience, offer, price with its unit, trial or terms, each feature, the goal) and note where it stands visibly in the delivered images (a software feature as a control, state or result). Add what is missing in the existing typography before the reviewers see it, or note it for the wrap-up as left out on purpose. In feed formats only the statements the brief asks to show.

Stage 5: Review

Order: render and look, gates, triage the generic check, measurement profile, reviewers, disposition. Commands, thresholds, reviewer briefing and rubric in review.md; script path $S following scripts/README.md.

Render and look. First the 360 px view (squint: what do you see first?), then full size, for web mobile at 390 px, for sequences as a contact sheet. Widths: Quick 1440 and 390 plus the 320 run, Standard plus 768, Studio plus 1024.

Gates that block (review.md 3, accessibility.md 2): contrast WCAG 2.2 AA (text 4.5:1, large type and controls 3:1), no horizontal overflow at 320 px (responsive surfaces), export sizes, minimum type size per medium, truth (nothing invented; stand-ins, draft copy and gaps marked in the review view; a rebuilt real interface follows its reference capture), language (lang set to the output language, no word broken without a hyphen, hyphens: auto where the language needs it, grammar and spelling of every text in the image, target-market spelling), target sizes and states for product UI, reduced motion, user rules, sources and credits, the accessibility gate following accessibility.md 2 for every screen and every sent document, legal mandatory points when triggered (open [REQUIRED: ...] placeholders mean draft, not ready to publish), generic index after triage at most 2 without a complete cluster; a stricter value from register or datasheet applies (Operate register at most 1).

Generic check: node $S/generic-check.mjs <file> --medium <web|ui|slide|social|document> [--selector .slide] [--deck talk|reading] --allow-fonts "<house typeface>" --brand-colors "<#hex>" [--intentional "<IDs>"] --md generic.md. First answer the visual tells on the image, then read the report; triage every red tell as truly generic, intentional (stated in the contract) or measurement error. Replace, don't delete: a truly generic element is replaced by an element from the product's world with the same job (warmth, emphasis, movement), for example a marking or scale from the project world instead of the italic accent word; deleting without replacement makes the surface emptier, not better (review.md 5). Calibrated at 1440 x 900: none of 53 top references and unseen award-winning sites is generic, 82 percent of real AI sites miss the gate. The index measures closeness to templates, not originality; catalog and visual tells in generic-catalog.md. If the check cannot run or a part of it fails (for example the pixel check: the report then says "pixelCheck": "not run" with the reason), say so; never treat a missing result as clean.

Measurement profile of the direction (measure.mjs, pixels.py palette) in the mode from the contract: type scale, words in the first screen, saturated pixels, calm area and further corridors from the datasheet. It checks fidelity to the direction, not quality. Up to about 20 percent outside: justify or adjust; further: correct or record as an intentional deviation in the contract.

Reviewers in a fresh context:

RoleGuiding questionLevel
R1 Art DirectorDoes it look finished and convincing as if from a studio, does one tangible idea from the product's world carry the whole, does the key surface show the offer, and do the images prove the claims?all (in Quick as a short review, when subagents are available)
R2 AudienceUnderstanding test: who offers what, what should I do, do I believe it?Standard, Studio
R3 GenericWhat would a standard generator have done the same way?Studio (in Standard the lead agent handles the detector report)
R4 Production and AccessDoes it work in real use for everyone?Studio (in Standard through the gates)

Reviewers get only a review folder following templates/review-brief.md: contract, rubric and images with neutral names; no code, no history, no reasoning, no ranking from the generator. R1 states the thesis from the images first and reads the direction in the contract only afterwards. R2 gets only the persona from the contract and answers fixed questions on the reduced image; otherwise it answers from the text instead of the image. All name three flaws with location first, then scores, then a disposition. R3 and R4 see detector findings only after their own judgment. Pair comparisons always in both orders; only consistent judgments count. Reviewers do not discuss with each other; results are merged by rules: veto for gates and blockers, for taste the majority among R1, R3 and R4; R2 counts for understanding and credibility. Every flaw from "important" upward is fixed or named in the wrap-up with a reason. A truth finding is solved without sacrificing impact: stand-in, built mockup with sample data, draft copy or open item in final form (stage 4, point 6) instead of an empty surface or a deleted section. Reviewers see the presentation view; visible placeholders, gray photo areas and empty zones are flaws of impact. Reviewer briefings are in English; copy is quoted verbatim in its language. R2 gets the persona in the contract's language and answers with words in the output language.

Disposition and stop: ship, targeted fix (three to five points by lever; every fix replaces or adds instead of just making the surface quieter, and removes the duplicate it creates; after the last reviewer call only local fixes that align with what exists, review.md 13 and 15), rebuild or restart. At most two review rounds per stage and never more reviewer calls than in the effort table, unless the user books one more round (review.md 16). The same flaws after two rounds mean a change of course, not a third round. The best version ships.

Ship threshold for every level, Quick included: looks finished (no placeholder as anchor, title or main figure, no gray photo area, no production note, no "to be added" label, no label or field without a value, no dashed box and no on-image hedge in the presentation view), one tangible idea from the product's world visible, the offer shown on the key surface (Persuade, Feed, solution slide in the deck), an anchor with a named origin, at least one deliberate decision against the category and AI default (in expression, not in the frame), all gates green. If it is missed, change course or report honestly instead of delivering.

Stage 6: Wrap-up

Brief: what was made and where it is; how it fulfills the brief (from the brief check, stage 4, point 10; never report a statement as fulfilled that is not visible in the delivered images); what was checked, with results (gates, generic index, measurement profile, reviewer dispositions) and what stayed unchecked; not run or substituted: every check that did not run or ran as a substitute, with the reason; intentional deviations with reasons, including every deviation of a rebuilt real interface from its reference capture; image and font sources with licenses (credits following templates/credits.md); labeling of generated images with location; stand-ins and draft copy, open [REQUIRED: ...] points and asset requests (details in the review view); the directions not chosen as one-liners, if they were not shown as mockups. For ads and social also test-plan.md following use-cases/README.md 6 and the note that reviewer judgments do not replace a market test. Then acceptance or change; anything beyond the scope is named as an add-on with its effort.

Curated delivery: the requested views as key images plus one overview board that comes first, with a title per view (pixels.py sheet ... --labels labels each view by its file name, so name the files by title, for example 01-defect-list-desktop.png) and one sentence per view in the wrap-up; check renders (320, 768, extra states) and working views go into a subfolder checks/, not into the delivery. A view the brief did not ask for is delivered only when it is as strong as the requested ones. HTML decks are built on templates/deck-shell.html and also ship as a PDF from pdf.mjs, its page count equal to the slide count (craft.md 12.2).

Stage 7: Log and learn

Direction log per brand following templates/direction-log.json, in the project folder as design-log.json and, if writable, also in ~/.unaverage/direction-log.json: date, brand, case, level, register, R0, set with classes, seed, chosen and discarded directions, fonts, palette, signature element, generic index. Within a brand the log keeps things consistent (continue instead of reinventing). Across brands it rotates: on a tie in score the direction that did not win for the last three other brands takes precedence; the same applies to fonts, signature element and prop.

Collect improvements to the skill. User corrections that go beyond the single case, that is, cases where a rule of this skill led to a different result than the user wants, or where the user had to ask for something by hand. Collect them without changing the skill; project-specific wishes do not count. Propose them at a fitting moment (result is done, end of work, user brings it up), per proposal the observation, the cause in the skill, the concrete change with file and location, the effect next time; at most five, sorted by effect. Change only after explicit consent.

Effort levels

QuickStandardStudio
Meant forsingle piece with a short life, brand and content are setwebsite, campaign, UI, document; direction open or partialidentity, high-stakes pitch, new product UI, brand without guidelines, "the best"
Content questions0 to 23 to 55 to 8 plus discovery
Candidates internal3 one-liners for the built direction2 per direction3 per direction, 4 to 6 competitors rendered
Directions to the user1, in the wrap-up with its derivation3 mockups with the ranking (top 3 first when the user already named things); none when the direction is given3 to 4 mockups with the ranking (top 3 first when the user already named things), ranking judgment by R1 in both orders; none when the direction is given
ReviewersR1 short review (when subagents are available), else a cold quick checkR1, R2R1 to R4
Reviewer calls at most1310 including blind comparisons
Rounds1 fixat most 2at most 2 per stage
Contract and versions5 lines internal, at most 3 render versionscontract, versions by roundcontract shown, versions by round
Measurementrender 1440, 390, 320; generic check; contrast; 3 key valuesall gates, full measurement profile, plus 768plus 1024, all states, print
Duration, rough5 to 15 minutes30 to 60 minutes1.5 to 3 hours
Effort relativeabout a quarter1about twice

Time and effort figures are estimates; relative effort counts tokens, from one measured Studio run, and duration is an estimate of its own. In tests, 24 to 36 reviewer calls per Studio run brought no gain in impact over a light process; hence the caps. Build no further measurement or review scripts beyond scripts/, except a throwaway step-through for builds in a live deck (review.md 3).

Triage: lifetime and reach, openness of the direction (missing CI counts only for identity or brand building), reinvention or correction, stakes (funding, launch, trade fair, law). No "high": Quick. One "high": Standard. Propose Studio from two "high" when one of them is the stakes or identity. Examples: webinar visual Quick, landing page Standard, Series A pitch deck Studio. If material that the best directions need is missing, recommend Studio only with an asset request.

Signal words take precedence: "quick", "short", "just quickly" mean Quick; "directions to choose from", "variants" at least Standard; "the best", "time doesn't matter", "review independently" mean propose Studio without asking again; "just do it", "you decide" mean the Autonomous mode. If the user says nothing, Standard applies; Studio needs an explicit yes. German signal words: markets/german.md 1.

Quick in eight steps:

  1. Read the case row and profile, minimum type size from use-cases/README.md 4.
  2. Short calculation only over the starter set of the case row (directions/README.md 11.5); the recommendation is built following stage 1 (11.5, point 7).
  3. Kickoff with at most two questions, keep building with a best-form stand-in instead of waiting.
  4. Three one-liners for the built direction, each with a tangible image idea from the product's world, never "obvious"; build the strongest.
  5. Parameters from the datasheet (sections 4 and 5), font from fonts.md 4.
  6. Build, render (web 1440, 390, 320; fixed formats at target size and as a 360 px preview), gates, generic check, three key values.
  7. R1 short review or cold quick check (review.md 17), one fix.
  8. Check the ship threshold, short wrap-up (what was made, sources and licenses of image and type, stand-ins and open items, not run or substituted, ranks 2 and 3 as one-liners; for ads and social one sentence on testing in the feed instead of test-plan.md), log.

Modes next to the levels (details in dialogue.md 6):

  • Correction: improve an existing design, the direction stays. Render and measure the current state, 1 to 3 questions (what bothers you, what stays, what for), heritage inventory, before and after, one reviewer in a before-after comparison.
  • Review only: for "does this look generic?", "design check", "make it premium" when nothing is to be built. Quick check with measurement, verdict in one sentence, three changes with location and measure, what should stay. Build nothing new unless asked.
  • Autonomous: level from the triage, at most Standard, no waiting points, direction by ranking without a mockup choice. All assumptions and the two other directions in the wrap-up.
  • No user available (subagent, workflow, non-interactive run): like Autonomous; Studio only when the prompt explicitly asks for it.

Directions at a glance

13 directions in five effect groups, plus five registers and four modules. Each direction is a parameter card plus a principle, derived from design history, studios and measured references. This table is the only summary; datasheets in directions/.

CodeDirectionCoreLead effectMaterial needed
ORDObjective Orderconsistent grid, one sans in few sizes, objective photos, one signal colorReassuranceno
PROProduct as Proofthe real product interface with a running case story as the image, color means stateCompetencea presentable product
DATData with a Pointevery surface one sentence that claims something, and exactly one piece of evidence in gray plus an accentInsightreal data
STIStillness and Materialdirected emptiness around a real object in consistent lightValueobject in photo quality
EDIEditorial Storylarge photography of its own and a characterful display cut in alternationRelevancephotography
HERHeritage and Quotationa form from the brand's own past, set in the presentProvenanceheritage or archive
WRMWarm Closenessreal people in everyday life, daylight, soft formsClosenessphotos of real people
DRWDrawn Worlda character or illustration world of its own with a fixed techniqueLikabilityillustration or a simple technique
ONEThe One Ideaone visual punchline, readable in three seconds, the brand inside the ideaPunchlineone idea
RULOne Rule, Many Formsone primitive from name, logo or tool plus a nameable variation ruleRecognitionno
MAXContained Maximalismlead color, pattern, bold type, contained, one focus per surfaceJoyno
TYPType as Imagetype becomes the main image, a small function layer keeps it readableTalkabilitya typeface with character
RAWRaw and Honestcontent first, visible construction, loud only in word and materialAuthenticityattitude and copy

Modules, combinable with every direction: Motion, Space and Scene, Human Trace, Trend Quote (modules.md). Combine as base plus at most one accent per surface and two per project; the accent changes exactly one axis.

Reading plan

Read only what the stage needs. The references are large; most tasks need few files.

WhenFile
Stage 0case matrix and section 4 in use-cases/README.md, plus the profile in the group file; more than one language or market: section 7 there
Quickonly the use-case profile, the direction's profile (directions/README.md 5) and sections 4, 5, 7 of the datasheet, directions/README.md 11.5, review.md 3 and 17, generic check
silent groundwork and kickoff (from Standard)dialogue.md for message patterns and questions; exploration.md 2, 4 and directions/README.md 11.4 for world and R0; for M1 also that README (registers, profiles, material gate, ranking) and exploration 6.4; images-and-assets.md 2 when Q5 applies
Stage 3directions/README.md (registers, profiles, material gate, ranking); exploration.md in full; per direction in the set datasheet sections 4, 5, 7
after the choicedatasheet of the chosen direction in full; with an accent the second one only in sections 4 and 7; with a module accent modules.md
Stage 4craft.md (section 13 when another tool continues the work or a design system is asked for, cases C9, D2, J4); copy.md when writing copy; fonts.md 1 to 3 and the direction's rows in 4; images-and-assets.md when images are needed (section 8 for image generation)
Stage 2 and 5templates/design-contract.md, templates/review-brief.md, review.md, generic-catalog.md, accessibility.md 2
Stage 6 and 7templates/credits.md, templates/direction-log.json
German output copy, or the user writes Germanmarkets/german.md (whole file for German copy, section 1 for German conversation only)
law triggered: shop, prices and discounts, purchase and cancellation, environmental, health or finance claims, AI features, image generation, photos of people and customer logos, images from the user's website, a live website with tracking or web fonts, email and business mailcompliance.md 1 to 5; for the EU, Germany, Austria or Switzerland also markets/eu-dach-compliance.md; for AI features also case C10 in product-design.md
modes Correction, Review only, Autonomous, Direction givendialogue.md 6

Environment and tools

Scripts live in scripts/ of this skill; in the references the path is $S. Before first use run npm install once in the skill folder (it installs puppeteer-core and its dependencies, no browser; needs Node 22.12 or newer). $PY is the Python command found by the tool check: python3, else python, else py -3, each accepted only if --version reports Python 3 (on Windows python3 can open a store placeholder), or the venv ~/.unaverage/venv from the fallback in dialogue.md 7.1. The scripts find Python the same way through scripts/python.mjs and honor the environment variable PY (for example PY="py -3"). font-check.mjs needs network access for Google Fonts and Fontshare, offline only with a local file. Overview in scripts/README.md; which tool feeds which function, with install commands and substitutes: dialogue.md 7.1.

Operating modeConditionConsequence
Fullshell, Node, Chrome, Python, subagentsprocess as described; reviewers ideally on a different model strength
Partialshell available, single tools missing or too oldthe functions of the missing tools drop out or run as substitutes (dialogue.md 7.1); the kickoff says which, and the wrap-up lists them
Without subagentsshell and toolsreviewers by substitute: a separate fresh run, else the user with guided questions ("What do you see first? What bothers you? What is missing?"), else a cold self-review, explicitly named as the weakest signal
Without shell (for example Claude.ai chat)no executiondo not measure, do not invent a seed (substitute rule in dialogue.md 7.5); mockups as one artifact side by side; mark results as "not checked on the image"; generic check as a visual check following generic-catalog.md 9

Missing tools never fail silently. The tool check stays silent only when everything is present. If a tool is missing or too old, the kickoff carries a short block: what is missing, which functions drop out or run weaker (table in dialogue.md 7.1), and the options: install (exact command, run only after the user's OK), a named substitute (for example a visual generic check instead of the measured one), or continue without it and accept the named loss. The user decides; in the modes Autonomous and No user available nothing is installed, the named substitute is used and listed. The wrap-up lists every check that did not run or ran as a substitute, with the reason. Nothing is installed without the user's OK.

If the environment limits the level on offer, the kickoff says so in one sentence.

Limits

  • No invented key figures, testimonials, customer logos, ratings or people to make a design look more credible. Sample data is consistent and, on persuasion surfaces, introduced as made up (in a deck by its "Example" label, craft.md 12.2; no Acme, no Lorem, no round fantasy numbers); outcomes and conditions the brief does not give (a saving, a customer count, a rating, a tax or fee) are never invented, while a made-up sample case keeps its steps and the product's outputs within the brief's ranges (craft.md 5.1, point 9).
  • No generated faces for real people and no generated images that look like photos of real events. A stock photo never poses as a real employee or a real customer. Photos of real people only with approval and edited locally where possible.
  • No copies of specific third-party designs and no imitation of living artists or studios. References provide principles, not templates; signatures such as a studio's logos, patterns or characters are never rebuilt. Third-party images never appear in the result.
  • Image generation only as a reasoned proposal after explicit approval, with labeling where the law requires it.
  • No legal advice: the skill sets mandatory points as placeholders and asks, instead of inventing legal texts; when expert review is needed is in compliance.md 5.
  • No publishing, purchases or uploads to external services without approval.
  • A draft may be experimental; a final result stays understandable, accessible and usable for its purpose.

Related Skills