SKILL.md — photo-monet-postcard-v1.0

1. Name

photo-monet-postcard-v1.0

A compact, self-contained skill for generating a minimal postcard front and back from one source photo, where every drawn element is rendered as a Monet-style impressionist oil study.

Use only this document. Do not retrieve or reference external skills.

Lineage note: this skill descends from photo-zine-postcard-v3.3. The layout discipline and "original photo is primary" rule are inherited. What changed: the rendering language is impressionist oil paint rather than watercolor/ink line work, and the paper, typography, and border language shift to a 19th-century exhibition register to match.

---

2. Goal

Generate two coordinated images:

1. Front — the untouched source photo above, one impressionist oil motif below
2. Back — a matching, writable postcard back

The design must remain:

• minimal and airy
• source-specific
• light-driven rather than object-driven
• painterly but controlled
• print-friendly

The tension the whole card lives on: a modern photograph sitting quietly above a passage of oil paint derived from it. The photograph is the record; the painting is the impression of the same moment. Do not blur that distinction by softening the photo or by tightening the painting.

---

3. Default Format

• card ratio: portrait 2:3 for landscape or square source photos; landscape 3:2 when the source photo is itself portrait (see §5)
• reference print size: 100 × 150 mm
• background: warm linen / raw-canvas tone, slightly cooler and greyer than ivory paper
• texture: visible fine canvas weave, strongest in empty areas where light rakes across it
• original photo: embedded directly, unmodified
• lower passage style: monet_oil_first, full-scene repaint
• palette: enabled by default, sampled as light-and-shadow rather than as flat colors

---

4. Fixed Front Structure

Do not redesign the layout.

1. original photo in the upper area
2. metadata block in the breathing band directly below the photo, left-aligned with the photo's left edge
3. one large impressionist oil repaint of the whole scene occupying the lower area
4. exactly three small color swatches, stacked as the final rows of that same block (§10)

Everything on the left sits on one shared left margin, so the card reads as a vertical spine of small quiet marks running down the left edge with the painting spread out to the right of it. That opposition — a thin left column against a wide right mass — is what keeps a large painting from unbalancing the card.

For a portrait source photo the card turns landscape 3:2 instead: photo on the left, painting on the right, metadata in the band directly below the photo, swatches in a column under the metadata. The relationship stays the same — photograph and painting side by side, small marks on one shared margin — only the axis rotates.

Describe each element exactly once. A rule restated in two places is not emphasis — the model satisfies both statements. Told in one bullet that the swatches are the label's final rows and in another that they are a column below the metadata, it painted two sets of three.

---

5. Main Photo

• Use the actual source photo.
• Preserve its exact original aspect ratio.
• Do not stretch, repaint, replace, filter, or crop it by default. In particular, do not apply the oil treatment to the photo — the contrast between photograph and painting is the point of the card.
• Place it in the upper half, centered horizontally.
• Add a very thin frame: a fine warm-grey line, optionally with a hairline of deeper tone inside it, echoing a mounted plate rather than a gilt frame.
• Leave a small even canvas gap between photo and frame.
• At most one tiny archival tape or corner-mount detail is allowed.

Portrait source photos

A portrait photo embedded at its original ratio would eat almost the whole height of a 2:3 card and leave the painting nowhere to go. When the source is taller than it is wide, rotate the card to landscape 3:2 and place the photo on the left, the painting on the right. Everything else — unmodified photo, thin frame, label under the plate, swatch column on the shared margin — carries over unchanged.

Do not solve this by cropping the photo to landscape. The photograph's framing is the user's, not ours.

The photo fidelity limit — read this before promising fidelity

Every rule above says the photograph must be reproduced untouched. A single image-generation call cannot deliver that. Tested repeatedly on the same sources, with the prohibition stated explicitly and then restated more forcefully:

• a frame shot at a deliberate diagonal gets silently leveled
• a frame shot at an extreme roll gets rotated upright and recomposed
• the plate's proportions get squeezed toward square — a 3:2 source lands near 6:5, a 4:3 source near 1.1:1

The model is not embedding the photograph; it is repainting it convincingly. For ordinary, level, conventionally framed sources the difference is hard to see and the result is fine. For anything with deliberate tilt or unusual proportion, the user's own framing will be overridden.

When fidelity actually matters, split the work: have the model produce the card with an empty plate area, then composite the untouched source file into it with an image library and draw the frame line, label, and swatches in code. The painting stays generative; the photograph and the type stop being negotiable. Be honest with the user about which path is in use rather than claiming a fidelity the pipeline cannot hold.

---

6. What the Painting Depicts

The painting is a repaint of the whole scene, not an extracted fragment. It shows the same view as the photograph above it, in the same framing, rendered in oil. The card then reads as one moment recorded twice — once by the camera, once by the brush.

What still requires judgment is which light condition the painting pushes. Impressionism is about a specific state of illumination, and a photograph usually contains more information than a painting can carry. So decide, before painting, what the light is doing in this image, then let everything else organize around it.

Identify the light event first

Name it explicitly: low sun raking across a slope, backlit foliage, mist dissolving a valley, overcast flatness broken by one bright reflection, warm window light on a cool street. That description drives the whole palette.

Then decide what dissolves

A repaint is not a transcription. Distant and atmospheric passages — sky, mist, far hills, water haze — dissolve into loose strokes and lose their detail. Passages carrying the light event keep the most structure and the strongest color. Foreground clutter simplifies into a few masses.

• keep: the light-bearing forms, the main silhouettes, the spatial layers
• summarize: foliage into masses, crowds of small objects into rhythm, ground texture into stroke direction
• dissolve: sky, mist, distance, reflections, anything already soft in the photograph

Difficult sources

Some photographs resist this. Flat, shadowless light gives the palette nothing to work with; hard-edged manufactured subjects (signage, vehicles, screens, machinery) resist broken color and tend to come back looking like a bad filter. When the source is like this, push the atmosphere rather than the objects — find the warm-cool relationship that does exist, however subtle, and build on that. Say so plainly if the source is a poor fit rather than producing something muddy.

Faces are the real limit. Broken color destroys likeness, and a repainted portrait usually reads as a distorted stranger. When the photograph's subject is a person's face, keep the painting to the setting around them, or tell the user this source suits the skill poorly.

---

7. Monet Oil First Rule

The painting is rendered as an oil in Monet's idiom by default. This is the core of the skill, so it is worth being specific about what that actually means — "oil painting style" alone tends to produce a generic muddy filter.

Required treatment

• Broken color. Build every area from small adjacent strokes of distinct hues rather than blended flat fills. Green foliage is made of yellow-green, blue-green, rose, and ochre touches sitting side by side; the eye mixes them, the brush does not.
• Visible directional strokes. Let stroke direction describe the form: horizontal drags across water, upward flicks for grass and reeds, short comma-strokes for foliage, broad soft sweeps for sky. Strokes are short and loaded, roughly comparable in scale to each other.
• Light impasto. A few raised ridges of thicker paint on the brightest highlights, catching light. Restrained — this is a 100 mm card, not a museum canvas.
• No black, no outlines. Shadows are colored, not darkened: violet, ultramarine, deep teal, madder. Never draw a contour line around the motif; edges come from where one color mass meets another.
• Warm-cool structure. Sunlit passages go warm (cream, coral, chrome yellow, rose); shadows swing to the complement (violet, blue, cool green). This complementary tension is what makes the light read.
• Soft, dissolving edges. Outer contours break up and let a little raw canvas show through at the boundary, so the painting sits in air rather than being pasted on. See §8's "painted boundary" for what this has to look like in practice — a rectangle with feathered edges does not qualify.
• Preserved identity. The scene's silhouettes, spatial layers, and color character stay recognizable. A viewer holding the photo should see the same view.

What to avoid

These are the common failure modes, and each pulls the result away from Monet toward something else:

• muddy, over-blended grey-brown — usually caused by mixing on the canvas instead of juxtaposing
• a uniform digital "oil paint filter" texture applied evenly across the whole area
• thick, writhing, high-relief strokes — that is Van Gogh, not Monet
• flat poster-like color blocks — that is Gauguin or vector illustration
• black shadows or ink contour lines carried over from pen-and-wash habits
• photorealistic detail rendered in paint texture; forms should stay summarized
• smudged, blurry mush; strokes must remain individually crisp even while forms stay soft

---

8. Painting Size and Placement

Broken color needs room. Each stroke has to stay individually legible at 100 mm print size, and there have to be enough of them to carry a whole scene — so this passage is the largest element on the lower half of the card, not an accent tucked into a corner.

Size:

• width: 55%–75% of card width
• height: 35%–45% of card height

Placement:

• occupying the lower right region, with its left edge clear of the swatch column
• an even margin of bare linen on all four sides — the gap above it, below it, and to its right should read as the same width, and the gap on its left at least that much once the swatch column is accounted for. Uneven margins, a generous left and a sliver at the bottom right, make the painting look shoved into the corner rather than placed.
• no side of the painting may come closer than about 6% of the card width to a card edge

State the margin as a positive requirement — "leave an even margin of bare canvas on all four sides" — rather than as a prohibition on bleeding off the edges. Tested both ways on the same sources: the prohibition was ignored almost every time, the positive form worked on most cards. Models follow a described target better than an avoided one.

Since models overshoot width by roughly 15-20%, ask for the LOW end of the size range in the prompt. Requesting 60% tends to land near 75%, which still leaves the margins intact; requesting 75% lands near 85% and destroys them.

The painted boundary

The edge of the painting is where the brush stopped, not where a crop was applied. This is what separates a study on canvas from a pasted rectangle, and it is the single detail most likely to come back wrong.

• strokes end at different lengths along every side, so the contour is ragged and uneven
• the corners are gone entirely — dissolved, not rounded
• bays and notches of raw canvas bite into the painted area along the boundary, and stray strokes reach out past it
• the four sides do not match each other; one side may dissolve over a wide band, another end more abruptly
• no straight side anywhere, no clean crop, and no uniform frayed border applied evenly around a rectangle — an evenly feathered rectangle is still a rectangle

Say all of this explicitly. Asked only for "soft dissolving edges", models return a straight-sided block with slight fraying at the corners.

It stays subordinate to the photograph in the reading order, but it is a painting, not an ornament. Give it the space it needs; the whitespace budget is what yields.

When size and stroke detail compete for attention in a generation prompt, state the size first — descriptive brushwork language reliably crowds out numeric constraints that follow it. Expect image models to overshoot the width anyway, typically landing near 80%; that overshoot is what eats the bottom and right margins, so if the margins fail, the width is the thing to attack.

---

9. Sets and Batches

When several photographs are processed into one set, the cards need to look like they came from the same press, which matters more than any individual card being optimal.

Hold constant across the set:

• the same linen ground tone and weave
• the same frame line weight and photo margin
• the same type treatment, label position, and swatch column geometry
• the same stroke scale relative to card size, so one card does not look painted with a much larger brush than its neighbors
• sequential catalogue numbers — No. 01, No. 02, and so on, in the order the user listed the photographs

Let vary per card:

• the palette, which follows each photograph's own light
• the card ratio, when a portrait source forces landscape (§5)
• how much dissolves versus holds structure, which follows each scene

A single shared back serves the whole set. Generate it once and reuse it rather than making a near-identical back per card — the differences would be invisible and the cost is real.

---

10. Color Swatches

Exactly three small swatches, sampled from the source photo but read through impressionist logic. Instead of "three colors that appear in the image", pick the three that describe the light:

1. light — the color of the illuminated passages
2. shadow — the complementary cool the shadows fall into (violet, blue, cool green), not a darkened version of the light color
3. accent or ground — the most distinctive local color, or the pale canvas-adjacent neutral

Rules:

• small and secondary
• soft-edged brushy dabs with irregular, hand-made outlines, as if smeared straight onto the canvas with a loaded brush — bristle drag visible, weave showing through the thin edges. Say this explicitly and name what to avoid: asked for "soft-edged dabs" alone, models return geometric squares, rounded rectangles, or flat color chips almost every time.
• no palette strip, no texture sample cards, never more than three
• stacked in a single vertical column as the final rows of the metadata block, not as a separate element placed elsewhere on the card — reading top to bottom as light, shadow, accent

Getting the column to stay put

Swatch position resisted three passes of increasingly explicit percentage-based wording: stated as "on the left margin", then "left-aligned with the metadata", then "left edges on exactly the same vertical line as the title text", the column drifted 10-55 px left of the title on nearly every card, and on one card relocated to the far right edge.

Two changes fixed it to within 1-3 px on the next attempt:

• Bind the swatches into the label. Describe the exhibition label as one block of seven stacked rows sharing a single left edge — title, LOCATION, DATE, number, then the three dabs — and say the dabs are the bottom of that block. An element described as independent will be placed independently; an element described as part of a block travels with it.
• Measure with things in the picture, not percentages. "Each dab's left edge lines up with the left stem of the N in No. 01 above it" and "each dab is about as wide as the word DATE is long" both landed within a few pixels. Percentages are abstractions the model cannot see; a letterform on the same card is something it can.

Anchor the block itself the same way, or it drifts as a unit: state that the block's left edge begins directly below the left edge of the photograph's frame line. Bound to the swatches but not to the photo, the whole block slid 46 px out onto its own margin.

This device generalizes to any placement that matters. It did not, however, fix the painting's size or margins — see §8.

• each swatch no wider than about 6% of the card width — they are a footnote to the painting, and image models tend to inflate them if given no number
• never compete with the painting

For a lily-pond source, this would typically be: warm pale green-gold (light), deep violet-blue (shadow), rose-pink (accent).

---

11. Metadata and Typography

Because the visual language moved to 19th-century painting, the type moves with it: set metadata like a small exhibition label under a canvas.

Allowed text only:

• title
• optional short subtitle
• LOCATION
• DATE
• small index or catalogue number

Type register:

• a light serif or old-style roman for the title, small and generously letterspaced
• small caps or petite caps for LOCATION and DATE
• ink-brown or warm charcoal rather than pure black
• optional very thin rule above the block, the width of the text only

Arrangement: the label is one block of seven stacked rows sharing a single left edge — title, LOCATION, DATE, number, then the three swatches as its final rows (§10). Left-aligned, no colons after the labels, never strung onto one horizontal line. Spelling this out matters because an unspecified detail is where output drifts — left to itself the same spec will produce a stacked block on one card and a single strung-out line on the next, and a set then looks unrelated.

Placement: in the breathing band directly below the photo, with the block's left edge beginning directly below the left edge of the photograph's frame line, restrained and compact. Sitting it right under the plate is how an exhibition label actually behaves — it belongs to the photograph, not to the painting. It also keeps the band below the photo from reading as dead space once the painting takes over the lower right.

Deriving a title

When the user has not supplied a title, derive one from what the photograph shows: at most five words, descriptive, naming the subject and where useful its light condition — Rolling Hills at Sunrise, Willow and Lily Pond at Dusk, Obedient Plant, Shade Light. A recognizable landmark may be named directly (Hall of Prayer for Good Harvests). This is a description of the image, verifiable by looking at it, which is why it is safe to generate.

LOCATION and DATE are different in kind. They are factual claims about where and when the photograph was taken, and nothing in the image establishes them — recognizing a building is not the same as knowing the shot's location and date. Leave both blank unless the user provides them, and never soften the gap with Unknown or Undated. When a landmark is recognized, put the name in the title rather than asserting a location.

Do not add keyword lists, descriptive paragraphs, or multiple labels. One title, not a string of nouns.

---

12. Forbidden Front Elements

Do not add:

• multiple paintings or rows of studies
• gilt or ornate frames, easel or canvas-corner mockups
• palette-and-brush props, paint tube illustrations, artist-signature flourishes
• "Monet" as printed text, or any real artist's name presented as authorship
• texture sample boxes, material cards, image grids
• large flat color blocks, decorative dots, wave doodles
• badges, seals, logos
• keyword lists, long captions
• full-width lower compositions
• rounded card corners unless explicitly requested

If uncertain, simplify. The painting should be the only expressive element on the lower half.

---

13. Back

Generate a unified postcard back on the same linen ground.

Include:

• thin outer border in warm grey-brown
• one vertical divider slightly right of center
• stamp box in the upper-right
• 3 or 4 address lines on the right
• large blank message area on the left

Optional:

• small POST CARD in the same serif near the upper-left
• tiny catalogue index near the lower-left

Keep the back plain. An earlier version allowed a faint painted echo of the front scene in the message area; in practice it lands as a stain rather than as atmosphere, and it prevents one back from serving a whole set. Omit it unless the user asks.

Do not add a large collage, palette, sample blocks, or any painted passage dense enough to make the card hard to write on. Ink must sit legibly on the surface.

---

14. Quality Requirements

Target:

• highest detail quality, sharp overall rendering
• individual brushstrokes crisp and distinguishable, even where forms are soft
• visible pigment granularity and light impasto relief
• clean, even canvas weave in empty areas
• low noise, no blur, no smudging, no muddy mixing
• suitable for later 4× super-resolution upscaling

These requirements improve rendering only. They must not add elements, and they must not tighten the painting into photorealism — "sharp" refers to stroke definition, not to detail count.

---

15. Priority Order

1. original photo embedded correctly and unmodified
2. fixed layout preserved
3. clear breathing space kept between the photo and the painting, and around the metadata block
4. the light event identified and driving the palette
5. Monet oil treatment: broken color, visible strokes, colored shadows, no outlines
6. painting large enough for strokes to read
7. exactly three light/shadow/accent swatches, small and stacked
8. minimal exhibition-label metadata
9. consistency across a set (§9) ahead of any single card's optimum
10. omit everything else

If a later rule conflicts with an earlier one, keep the earlier rule.

---

16. Front Prompt Template

Create a minimal portrait postcard front, 2:3, on warm linen canvas ground with a fine visible weave.

LAYOUT FIRST — these positions and sizes are requirements, not suggestions:
- The source photograph sits in the upper area, reproduced faithfully AS A PHOTOGRAPH. Do not repaint, filter, stretch, or crop it. Preserve its exact original aspect ratio of <RATIO>. Center it horizontally, thin warm-grey frame line, small even canvas gap between photo and frame.
- IMMEDIATELY BELOW the photograph, in the narrow band between the photo's bottom edge and the painting, place the label block. Its left edge begins directly below the left edge of the photograph's frame line. This band must not be left empty.
- The label is ONE BLOCK OF SEVEN STACKED ROWS sharing a single left edge, reading downward: the title <TITLE>, then LOCATION with its value left empty, then DATE with its value left empty, then <NO>, then three small paint dabs one below another as the block's final rows. No colons after any label, never strung onto one horizontal line. The dabs belong to this block and continue its left edge — they are not a separate column placed elsewhere on the card, and they must never sit in the outer margin or on the right side of the card.
- Each dab's left edge lines up with the left stem of the N in <NO> directly above it, and each dab is about as wide as the word DATE is long. They are soft-edged brushy smears with irregular hand-made outlines and bristle drag, weave showing through the thin edges — NOT geometric squares, rounded rectangles, or flat color chips. Top to bottom: <LIGHT COLOR>, <SHADOW COLOR>, <ACCENT COLOR>. There are exactly three dabs on the card in total, and they exist only inside this label block.
- ONE large impressionist oil painting in the lower area, about 60% of the card width and 38% of its height, clearly narrower than the photograph above it. It is the dominant element of the lower half. Do not shrink it because of the brushwork detail described below.
- The painting must leave an EVEN margin of bare linen canvas on all four of its sides — the gaps above, below and to the right reading as the same width, none closer than 6% of the card width to a card edge.
- The painting's boundary must be an irregular hand-made edge, not a crop: strokes ending at different lengths so the contour is ragged, corners dissolved away entirely, notches of raw canvas biting into the painted area and a few stray strokes reaching out past it, the four sides not matching each other. No straight side anywhere. An evenly feathered rectangle is still a rectangle and is wrong.
Nothing else on the card.

The painting repaints THE WHOLE SCENE of the photograph — the same view, the same framing — in oil. The light event to build it around is <LIGHT EVENT>. Passages carrying that light keep the most structure and strongest color; sky, mist, distance and reflections dissolve into loose strokes; foreground clutter simplifies into a few masses.

Render it in Monet's idiom: broken color built from small adjacent strokes of distinct hues rather than blended fills; visible directional brushwork following the form (horizontal drags on water, upward flicks in grass, comma-strokes in foliage, soft sweeps in sky); light impasto ridges on the brightest highlights only; colored shadows in violet, ultramarine and deep teal with absolutely no black anywhere; warm sunlit passages against cool complementary shadow; soft dissolving edges letting a little raw canvas show through; no contour lines of any kind. Keep the scene's silhouettes and spatial layers recognizably the same view as the photograph. No frame or border around the painting.

The metadata block: light old-style serif in ink-brown, a hairline rule the width of the text on top, then the seven stacked rows described above.

The three swatches: as described above, brushy and irregular, never chips.

Avoid: muddy over-blended grey-brown, uniform digital oil-filter texture, thick writhing Van Gogh impasto, flat poster color blocks, black shadows, ink outlines, photorealistic detail, blurry smudging. Avoid gilt frames, easel mockups, paint tubes, palettes, brushes, signatures, artist names, keyword lists, sample boxes, multiple paintings, badges, seals, logos, long text, rounded card corners.

Quality: highest detail, crisp individually readable brushstrokes, visible pigment granularity, clean canvas weave in empty areas, low noise, no blur, suitable for later 4x super-resolution upscaling.

Overall mood: quiet, luminous, atmospheric, painterly, collectible — a photograph and its impression sharing one card.

For a portrait source photo, change the first line to landscape 3:2 and place the photo on the left with the painting on the right, keeping every other instruction as written (§5).

---

17. Back Prompt Template

Create a matching postcard back, same portrait 2:3 ratio, same warm linen canvas ground, same thin-line style and serif typography as the front.

Keep it functional and mostly blank.

Include:
- thin warm grey-brown outer border
- one vertical divider slightly right of center
- stamp box in the upper-right
- 3 or 4 address lines on the right
- large blank message area on the left

Add small POST CARD text near the upper-left in the same serif. Leave the lower-left index area clear so one back can serve a whole set.

Address lines should start below the stamp box and be evenly distributed down the right column rather than bunched at the bottom.

Do not add a painted passage, collage, palette, sample blocks, or any texture dense enough to interfere with handwriting.

Quality: clean sharp lines, refined canvas texture, low noise, no blur, suitable for later 4x super-resolution upscaling.

---

18. Quick Checklist

Front

• original photo embedded, unmodified, original aspect ratio
• card ratio matches source orientation (portrait source → landscape card)
• thin frame and even canvas gap
• fixed layout preserved, clear breathing band below the photo
• light event identified and driving the palette
• whole scene repainted, same framing as the photo
• broken color from adjacent distinct hues, not blended fills
• stroke direction follows the form
• shadows colored violet/blue, no black, no outlines
• light impasto on highlights only
• soft dissolving edges, some raw canvas showing
• painting 55-75% of card width, strokes legible at print size
• even bare canvas margin on all four sides of the painting
• painted boundary ragged and irregular — no straight sides, no feathered rectangle
• exactly three light/shadow/accent swatches, small, brushy-edged, stacked vertically
• swatches sit as the label block's final rows, left edges flush with the number above them
• label block's left edge begins below the photo's frame line
• metadata block sits directly below the photo, left-aligned with it, stacked in lines with no colons
• title present — supplied by the user, or derived from the subject in five words or fewer
• LOCATION and DATE blank unless the user supplied them, never Unknown
• sequential catalogue number within a set
• no gilt frame, palette props, or artist name
• strokes crisp, not muddy or blurred

Back

• unified functional layout on same canvas ground
• writing space preserved and legible
• plain — no painted echo unless requested
• one back reused across a set
• output clean and sharp

---

19. One-Line Definition

Keep the photograph untouched above, then repaint the same scene below as a large Monet-style oil — broken color, directional strokes, violet shadows, no outlines — on linen canvas, with an exhibition label under the plate and three small light/shadow/accent swatches stacked on the left margin.