Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Create original relaxed black-pen graphics from concepts or visual references, including sparse illustrations, narrative scenes, animals, objects, abstract relationships, icons, and logo marks. Use when the user wants quick hand-drawn line energy, causal story staging, non-equilibrium proportion, shorthand forms, open contours, structural line drift, misregistered junctions, selective retracing, physical pen texture, or a transferable illustration system that must not collapse into clean vector
.claude/skills/haorantang97-antibes-holiday/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-19 | ✗→✓ | ▲ Improved | 35% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 14% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 60% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 41% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 121% | 0% |
Create original graphics that feel drawn in one alert but lightly unsteady sitting with a real black pen. Prioritize stroke life, tempo, shorthand, structural instability, and the courage to stop. Semantic compression remains useful, but a clean concept never compensates for lines that look professionally controlled.
When references are supplied:
Default evidence weighting:
60%;30%;10%;0%.Read style-system.md before drawing. Read prompt-recipes.md before using an image-generation model.
illustration, series, icon/badge, or logo.principal actor, counterforce, narrative hinge,and the causal path between them.
distance, pause, anticipation, or release.
overshoot, displaced restart, incompatible pair of fragments, or abandoned path.
Do not default to a person, face, unrequested character traits, accessories, or anthropomorphism.
Draw three fast structural variants.
expected destination.
illustration. They must record discovery, not decorate the page.
trained, or gracefully sketched rather than found while drawing.
Add only the shorthand needed to make the subject legible.
competently explain the whole form.
delete so aggressively that the result collapses into a stick symbol.
Restore the physical behavior that image models and vector tools remove.
and recovering the form.
Outside it, prefer one imperfect attempt over several sketchy alternatives.
Living form: use posture, mass distribution, locomotion, and distinctive anatomy.Object: exaggerate the operational edge, opening, handle, joint, contact, or material path.Place: use one boundary, route, threshold, and scale cue.Abstract relationship: first choose an observable physical metaphor, then express tension, exchange, interruption, growth, or separation through its gesture. Do not draw free-floating lines when the meaning depends on explanation.Logo: begin with a relaxed pen master; derive a separate production vector only after the gesture works.Do not transfer human facial shorthand to non-human subjects. For human figures, preserve posture and character while withholding clean anatomy. Build the figure from unequal, partially incompatible fragments, with enough local specificity to remain observed rather than diagrammatic.
Treat complexity as relationships, not line count.
violate realistic perspective when that makes the event clearer.
narrative hinge where intention becomes consequence:contact, exchange, reveal, interruption, impact, or release.
path, or directional gap. Do not add arrows or motion icons.
supporting actors, and leave the setting largely unstated.
operational feature, displaced small actor, or edge-reaching contour.
When a scene is humorous, put the humor in disproportion, posture, timing, and physical consequence. Do not explain it with sweat drops, surprise ticks, exclamation marks, speed marks, or other cartoon punctuation unless requested.
Choose a composition family from style-system.md according to the event. Do not repeat one family across a series by default.
Inspect the tools available in the current environment before rendering:
imagegen workflow and built-in image_gen tool.while preserving the reference roles, prompt rules, and validation gates here.
stroke specification and state that the calibration image was not applied.
image. Deliver a renderer-ready prompt, the intended reference roles, and the complete validation checklist instead.
service without the user's explicit approval.
Use the selected raster capability for exploration and illustration.
When recent reference crops are available:
When no user references are available, supply the bundled stroke-calibration.png. It controls physical pen instability: pressure, drift, missed junctions, overshoots, displaced restarts, broad lurches, retracing, ink deposition, dry breaks, and abrupt terminals. Ignore every shape and layout in the asset. Never copy a swatch as an object, symbol, border, motif, or composition. Keep structural assembly in the text prompt; do not add a second visual reference whose cleaner lines could dilute the material.
Use a bounded decontrol pass when the first readable render remains too competent:
remove conventional detail, and let one structural drift alter the silhouette.
trained editorial, fashion, or period sketching. Otherwise keep the readable draft and edit one region at a time.
Ask for a photographed or scanned black-pen drawing only when physical stroke texture matters. Keep paper texture subtle and never use it to fake quality.
expressive master: preserves pen irregularity for display use;production master: optically tuned SVG for small sizes.Do not force an illustration line to survive at favicon size by making every stroke thick and geometric.
Validate against at least three recent, same-cluster references. The candidate must pass every primary gate:
Physicality: reads as real pen movement, not a Bézier outline or generic sketch filter.Tempo: contains visible contrast between fast sweeps and slower recognition marks.Instability: at least one structural line visibly loses and recovers control;intended meetings are not all perfectly registered.
Shorthand: forms are inferred from decisive fragments rather than fully described.Relaxation: proportions, spacing, anatomy, and terminals feel unforced withoutturning into uniform noise.
Transfer: the line behavior survives the new subject without importing human accessories or source motifs.Originality: no source composition, character, or signature device is reproduced.For story scenes, add two primary gates:
Causality: the spatial arrangement makes the action and consequence inferable.Hierarchy: scale, crop, density, and empty space clearly prioritize the storywithout defaulting to balanced staging or realistic perspective.
Secondary gates:
Fail the result immediately if it is vector-clean, diagrammatic, evenly polished, uniformly wobbly, decoratively distressed, or only recognizable because of labels. Also fail it when most contours arrive exactly where expected, figures read as competent fashion or period illustration, faces and hands are cleanly resolved, or the image could pass as a professional quick sketch with texture added afterward. Visible shoelaces, fingers, pockets, complete facial features, and fully joined limbs are warning signs unless one carries the story. Fail generic loose sketching when duplicate contours, search marks, and darkened corrections spread evenly across the whole subject instead of gathering at one knot. Fail unrequested cartoon shorthand when sweat drops, emphasis ticks, exclamation marks, or speed marks carry an emotion or action that posture and causality should have made visible.
Do not convert these gates into a single score. A strong semantic result with weak stroke physicality still fails.
Include:
Keep internal reference analysis out of the final response unless requested.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-19 | fail→pass | 16,465 | 5,092 | -69% | 1 | 1 | 0% | 2,550 | 3,436 | +35% | 0 | 0 | — |
case-01 | fail→fail | 47,674 | 36,630 | -23% | 1 | 1 | 0% | 8,298 | 4,324 | -48% | 0 | 0 | — |
case-02 | fail→fail | 39,590 | 8,043 | -80% | 1 | 1 | 0% | 7,423 | 3,881 | -48% | 0 | 0 | — |
case-03 | fail→fail | 47,595 | 29,731 | -38% | 1 | 1 | 0% | 8,275 | 3,800 | -54% | 0 | 0 | — |
case-04 | fail→pass | 48,388 | 29,322 | -39% | 1 | 1 | 0% | 8,259 | 9,388 | +14% | 0 | 0 | — |
case-05 | pass→fail | 53,461 | 17,108 | -68% | 1 | 1 | 0% | 8,233 | 5,349 | -35% | 0 | 0 | — |
case-06 | fail→fail | 54,251 | 31,358 | -42% | 1 | 1 | 0% | 8,240 | 6,867 | -17% | 0 | 0 | — |
case-13 | fail→pass | 20,500 | 10,662 | -48% | 1 | 1 | 0% | 2,620 | 4,184 | +60% | 0 | 0 | — |
case-07 | fail→fail | 19,380 | 22,181 | +14% | 1 | 1 | 0% | 2,805 | 3,645 | +30% | 0 | 0 | — |
case-08 | pass→pass | 13,917 | 23,566 | +69% | 1 | 1 | 0% | 2,158 | 5,074 | +135% | 0 | 0 | — |
case-09 | fail→pass | 35,134 | 8,428 | -76% | 1 | 1 | 0% | 2,805 | 3,963 | +41% | 0 | 0 | — |
case-10 | fail→pass | 15,476 | 22,880 | +48% | 1 | 1 | 0% | 1,924 | 4,252 | +121% | 0 | 0 | — |
case-11 | pass→pass | 17,514 | 15,259 | -13% | 1 | 1 | 0% | 2,437 | 4,549 | +87% | 0 | 0 | — |
case-12 | fail→pass | 14,742 | 11,543 | -22% | 1 | 1 | 0% | 2,184 | 4,319 | +98% | 0 | 0 | — |
case-14 | fail→pass | 16,891 | 13,562 | -20% | 1 | 1 | 0% | 2,393 | 4,158 | +74% | 0 | 0 | — |
case-15 | fail→pass | 51,066 | 14,406 | -72% | 1 | 1 | 0% | 8,218 | 4,499 | -45% | 0 | 0 | — |
case-16 | fail→pass | 16,801 | 16,783 | -0% | 1 | 1 | 0% | 2,610 | 5,099 | +95% | 0 | 0 | — |
case-17 | fail→fail | 16,309 | 11,484 | -30% | 1 | 1 | 0% | 2,485 | 4,479 | +80% | 0 | 0 | — |
case-18 | fail→fail | 27,094 | 9,928 | -63% | 1 | 1 | 0% | 1,873 | 3,876 | +107% | 0 | 0 | — |
case-20 | fail→fail | 12,009 | 19,439 | +62% | 1 | 1 | 0% | 1,791 | 3,713 | +107% | 0 | 0 | — |
case-21 | pass→pass | 19,583 | 36,718 | +87% | 1 | 1 | 0% | 2,622 | 5,643 | +115% | 0 | 0 | — |
case-22 | pass→pass | 14,560 | 10,713 | -26% | 1 | 1 | 0% | 2,048 | 4,193 | +105% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 19 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +36 percentage points is the difference between those two pass rates over the 19 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.