1{2 "skill_name": "dreambase-micrographic-design",3 "evals": [4 {5 "id": 1,6 "name": "apparel-legal-gate",7 "prompt": "make me a graphic for the back of a t-shirt for our dev tool launch — I want it to look like the care label / spec tag inside a garment, with the washing symbols and a CE mark and a barcode. black on white, we're printing 200 of these next week. save it as tee-back.svg",8 "expected_output": "Raises the certification-mark and care-symbol legal problem before building (a t-shirt is a shippable product), proposes invented marks with fictional identifiers and an original symbol family rather than reproducing CE and GINETEX symbols, generates a real scannable barcode or omits it, and scales the type system up for screen printing rather than using the 4-9pt band"
11 "Raises the legal issue proactively — a t-shirt is a product placed on the market, so a CE mark or CE-lookalike is not safe decorative use",
12 "Does not reproduce the real CE mark on the garment artwork, and does not propose a slightly-distorted near-CE as the solution (the lookalike prohibition covers 'meaning or form')",
13 "Flags that the ISO/GINETEX care symbols are registered trademarks and builds an original symbol family using the modular construction logic instead of copying the standard set",
14 "Any invented conformity mark is given a fictional identifier (file/body/licence number), because a bare glyph reads as fake",
15 "Addresses screen-printing production limits — scales type and stroke up rather than using the 4-9pt compliance band, which will not survive garment printing",
16 "Either generates a genuinely valid/scannable barcode or omits it, rather than drawing decorative stripes",
17 "Does not require the user to have asked about legality before raising it"
18 ]
19 },
20 {
21 "id": 2,
22 "name": "web-component-legibility",
23 "prompt": "restyle our pricing page's plan comparison table in the micrographics / spec sheet look — hairline rules, tick marks, technical micro type, that whole aesthetic. it's a React component, here's roughly what it has: 3 plans, 12 feature rows, monthly/annual toggle, and a CTA button per plan. make it feel engineered.",
24 "expected_output": "Applies the visual grammar (hairline rules, compartments, LABEL: value pairs, tabular figures, monochrome plus one accent, no shadows or radius) while keeping prices, feature labels, the toggle, and CTAs at readable sizes and accessible contrast — explicitly reserving the micro band for chrome only",
25 "files": [],
26 "assertions": [
27 "Keeps load-bearing content (prices, plan names, feature labels, CTA buttons) at readable sizes; the 5-12px micro band is used only for chrome like source lines, revision codes, or corner marks",
28 "Does not put interactive controls (toggle, CTA) in micro type",
29 "Maintains WCAG contrast (4.5:1 body, 3:1 large) and does not treat the technical aesthetic as an accessibility exemption",
30 "Applies the grammar concretely: hairline rules and boxed compartments instead of card shadows, no border radius, no gradients",
31 "Specifies tabular/lining figures with slashed zero for the pricing numbers",
32 "Marks decorative elements (tick strips, crosshairs, corner brackets) as aria-hidden or otherwise non-semantic",
33 "Applies the two-type-size discipline in the data area rather than inventing an intermediate size ladder"
34 ]
35 },
36 {
37 "id": 3,
38 "name": "rating-plate-authenticity",
39 "prompt": "design a fictional rating plate graphic for a made-up hardware product for our landing page hero — an edge compute box called the NODE-7. I want it to look genuinely real, like something photographed off the back of actual hardware. SVG please.",
40 "expected_output": "Produces a rating plate that reproduces the authentic incoherence — mark band with mixed typography, unequal gaps, optical rather than baseline alignment, satellite identifier text off the grid, a separate overlapping variable-data layer — with real internally-consistent specs, invented marks with fictional identifiers, and correct IP-code and symbol conventions",
41 "files": [],
42 "assertions": [
43 "Reproduces the incoherence deliberately: marks are not baseline-aligned, gaps between marks are irregular, and marks carry differing typography rather than a unified type system",
44 "Includes satellite identifier text (file/licence numbers) at sizes unrelated to the main text grid",
45 "Uses invented conformity marks with fictional identifiers rather than real CE/UL/FCC marks, or restricts itself to free informational symbols",
46 "Specs are internally consistent and plausible (voltage/current/wattage actually relate; IP code well-formed with no spaces or hyphens and both positions filled)",
47 "Applies rating-plate structure: identity block, rule, LABEL: value spec block, mark band collected in one region rather than interleaved with data",
48 "Includes authenticity details such as a separate overlapping serial/barcode layer, an artwork part number in the margin, or a revision code",
49 "Uses two type sizes in the data area rather than a smooth size ladder"
50 ]
51 },
52 {
53 "id": 5,
54 "name": "applied-mode-not-simulation",
55 "prompt": "make me a poster for our AI infrastructure conference in that micrographics look — big and loud, it's going on a wall and also as an Instagram post. conference is called SUBSTRATE, it's in Brooklyn in October. I want it to feel like a piece of industrial propaganda, not like a boring spec sheet.",
56 "expected_output": "Applied-mode work: a saturated or dark ground, huge display type carrying SUBSTRATE, mono micro-layer as counterpoint, extreme scale contrast, deliberate art-directed composition, invented institutional telemetry, and one or two texture moves — not a monochrome compliance-plate simulation",
57 "files": [],
58 "assertions": [
59 "Works in applied mode rather than producing a monochrome rating-plate simulation — the user explicitly asked for loud, not spec-sheet",
60 "Uses a saturated or dark ground rather than defaulting to black-on-white monochrome",
61 "Establishes extreme scale contrast — the conference name at display size against mono micro-type, with no intermediate sizes",
62 "Composes deliberately (centred/balanced/art-directed) rather than applying the deliberate-misalignment rules, which belong to simulation mode",
63 "Invents an institutional frame (division, programme, sector, issue number) so the telemetry is internally consistent rather than random",
64 "Includes at least one texture move (line-screen fill, halftone, grain, grid, or hazard striping) rather than flat rules alone",
65 "Handles both output sizes — wall poster and a square/vertical social crop — rather than one artboard",
66 "Real information (name, city, month) is accurate; invented telemetry is clearly world-building rather than false claims"
67 ]
68 },
69 {
70 "id": 4,
71 "name": "brand-fit-pushback",
72 "prompt": "we're a meditation and sleep app, very soft and calming brand — pastel gradients, rounded everything, friendly illustration. can you redo our onboarding screens in that technical spec-sheet micrographics style? I saw it on a Nike thing and loved it.",
73 "expected_output": "Names the brand-fit problem directly — the style signals rigor, engineering and measurement, which is at odds with a warmth/calm brand — while still delivering something useful: either a scoped application where it genuinely fits, or the requested work with the tension stated",
74 "files": [],
75 "assertions": [
76 "Explicitly raises the brand-fit failure mode rather than silently executing the request",
77 "Explains the specific mismatch: the idiom signals rigor/engineering/measurement, which conflicts with a warmth-and-calm value proposition",
78 "Still delivers something concrete rather than only refusing — e.g. a scoped surface where the idiom genuinely fits, a hybrid, or the requested design with the tension flagged",
79 "Does not moralize or lecture; states the concern briefly and moves to useful work",
80 "If proceeding, notes the specific conflicts with the existing brand (gradients and rounded corners are incompatible with the idiom's flat single-colour logic)",
81 "Treats the user's decision as final if they reaffirm the request"