Copy the reusable analysis prompt

Attach your source image to the analysis message and paste this:

Inspect the attached image and write ONE complete English image-generation prompt to recreate its visible appearance as closely as possible using TEXT ONLY. The generator will not see the image. Do not guess or retrieve its original prompt.

Use these priorities:

1. LAYOUT AND PROPORTIONS. Identify aspect ratio, viewpoint, crop, subject size, negative space and left-to-right ordering. Use percentages across and down the FULL image from the top-left, including blank paper or borders. Give the entire bounds of dominant subjects and important empty regions, not only object centers. Cross-check width-to-height ratios so familiar object names do not stretch the observed shapes. Choose three to five crucial landmarks and give approximate horizontal AND vertical positions. For small attached parts, distinguish whole bounds from centers and describe connection length. For multi-panel images, state each panel's bounds and divider positions; preserve unequal panel sizes. Establish the empty layout first, then place content within each region. For multi-panel scenes you may add clearly labelled panel-local placements for the figures, while retaining full-sheet bounds and checking the conversion; never mix the two coordinate systems without labels. Do not size panels automatically around their contents. If the source dimensions are known, equivalent approximate pixel rows may clarify divider positions; keep them consistent with full-canvas percentages. Keep stray hairs, pigment and spatter out of protected blank areas. Do not fabricate precision.

2. DISTINCTIVE FEATURES AND STRUCTURE. Inventory the visible subjects, props and background elements; do not add plausible but absent decorations. Preserve exact counts, pose, expression, asymmetric markings, accessories and legible text. Use viewer-left and viewer-right to mean the left and right of the picture, not the subject's anatomy. Bind important asymmetric actions to approximate positions of their visible endpoints so hand roles cannot be swapped. Describe irregular silhouettes through bends, forks, cutouts and unequal parts, and explain what touches, overlaps or attaches to what. Keep small details on their actual depth plane. Bind each distinctive component's color and material to that component's shape and picture side; an object's overall palette is insufficient when its parts differ. Do not let surrounding light or a familiar prototype swap those component colors. Describe visible geometry before familiar object labels; do not invent extra components those labels might suggest. Preserve the source's hierarchy instead of inventing equally prominent objects.

3. VISIBLE APPEARANCE AND LETTERING. Describe medium, color distribution, saturation, contrast, lighting, materials and texture. Distinguish genuinely uniform flat fills from shaded or textured surfaces; preserve that difference. Describe a flat designed graphic as a direct 2D vector-layout export with uniform digital fills and glyphs, without physical lighting or a photographed-paper treatment. For painted images describe pigment patterns rather than inventing literal weather. For designed text, transcribe all readable words and specify line breaks, alignment, weight, letter proportions, approximate block width and actual letter height; do not replace those measurements with "large" or "bold" alone. Visible glyph height differs from font size: describe the observed glyph bounds. For document-like graphics, fix every text region separately instead of allowing automatic flow. A compact drawing specification, including simple vector commands if useful, may sit inside the standalone generation prompt; use visible geometry and keep it free of external assets. Preserve subdued and imperfect qualities. Avoid size-amplifying language if it contradicts the source's proportions. Do not invent hidden details, unreadable wording, equipment, an artist or an unverified location.

Return only:
- Suggested aspect ratio.
- ONE copy-ready generation prompt, normally 250-500 words. Start with subject, medium and aspect ratio, then a prominent compact layout plan, distinctive details, appearance and a short avoid list specific to this image. Include the aspect ratio inside the prompt. Resolve contradictions; layout and distinguishing features have priority over decoration.
- A brief note about important uncertain details to check, if needed.

The prompt must stand alone: never use "like the reference," "same as the image" or depend on a previous message. State that coordinates and layout-plan labels are invisible instructions and must not appear in the result; only text actually visible in the source should be drawn.

Copy the complete returned generation prompt, including any layout or drawing specification it contains. Keep the aspect-ratio suggestion and verification note separate.

Generate from the returned text

  1. Check the description beside your source. Verify counts, picture-left/right, markings, shapes and every readable word. Ask the analysis assistant to recheck anything it misread. OpenAI’s vision documentation describes limitations with small text and precise spatial relationships.
  2. Start a fresh generation session. Submit only the complete returned generation prompt and select the suggested aspect ratio if that setting is available. Leave the source picture and analysis conversation out of this session. Save the exact text you submit.
  3. Generate twice with that same text. Compare both with the source: crop and empty space, distinctive features, important object positions, medium and color distribution. Keep both results, including failures.

“36% down” means slightly over one third of the way from the full image’s top edge, including blank paper. A small object also needs its whole bounds and connection: a tag’s center can be correctly placed while a long attachment changes the portrait. For several panels, make sure any panel-local positions are explicitly labelled and agree with the full-page layout.

Five source images, two consecutive outputs each

The earlier template passed our cat and watercolor checks, then failed when we first tried three new types: a product photo, a flat event graphic and a three-panel 3D character sheet. We revised the reusable template and retested all five sources after each revision. The final version met our checks in both consecutive repeats for every source.

Three new source types compared with both final text-only repeats: blue travel mug, event graphic and three-panel 3D character sheet. Objects and broad positions match the checks; details and fills vary.
Left: source. Middle and right: both consecutive final text-only generations. All shown outputs were generated without attaching the source. Click for the complete comparison. Credit: Grow AI Skills original synthetic mug and fictional 3D fixtures; original event graphic; actual built-in text-only outputs. Diagnostic composite from uncropped retained images. Open full-size image.

Pass required the specified recognizable features, broad visual treatment and approximate layout. Every checked landmark had to remain within five percentage points of the source’s full canvas, including allowance for reading uncertainty. The final set passed all 66 coordinate checks. This measures selected positions, not an overall similarity score or exact reconstruction.

SourceWhat both repeats retainedVisible differences
Travel mugBlue body, fitted black lid, right handle, blank cream oval, oak table and wallLid seam lower; handle shape, badge proportions and surface texture changed
Event graphicAll six text regions, shape order, bars, button and approximate positionsGlyph sizes, heading width and fills changed; mild shading remains
3D character sheetThree poses, picture-side hand roles, clothes, scarf, boots and blue benchFace, furniture details and panel proportions changed; one lower face remains about four points higher
Cat portraitEye-color sides, ear notch, facial markings, rose collar and one blank silver tagFur, face proportions, notch outline and top margin changed
Watercolor cityBent unequal bridge legs with their colors, low left sun, attached slats and captionSkyline, sun size, paint width and pigment distribution changed
See the cat and watercolor retests with the final template
Original cat and watercolor sources alongside both new final text-only retests using the final reusable template.
Both original examples were also retested after the template changed; these are the new final outputs, not the earlier v13 pair. Credit: Grow AI Skills synthetic cat; existing AI watercolor derivative of Alexander Psiuk’s Unsplash photo; four actual text-only outputs. Open full-size image.

The mug needed its complete bounds and width-to-height ratio. The character sheet needed fixed panel sizes and separate, clearly labelled placements within each panel. The city scene needed the dark and pale colors bound to the correct bridge legs. For the event graphic, fixed text regions prevented the date and venue from being pushed down.

The graphic’s returned prompt included a self-contained SVG drawing specification based on visible shapes and measured glyph bounds. We submitted that specification as prompt text to the image generator; we did not render it separately or alter the returned image pixels. The generator still changed the type and fills, so the specification did not produce a lossless copy.

Open the exact product photo prompt used for both final outputs
Create a 4:3 landscape natural product photograph of one cobalt-blue lidded handled cup on pale oak in front of an off-white plaster wall. Reproduce a moderately broad cylindrical body, not a slender thermos.

FIRST preserve this full-canvas size plan. The COMPLETE mug, including its lid and handle, fits from about 32% to 73% across and 13% to 84% down; keep oak visible throughout the lower 16%. The blue cylindrical BODY alone fits 32-61% across and 22-84% down. Its physical height is about 1.6 times its width. The body has almost parallel sides with a slight taper and a softly rounded base. Its base center is(45%,83%); do not let it drop toward the bottom edge. The black lid occupies 32-61% across and 14-22% down, with a small raised drinking flap reaching 13% down at upper-right. A thick blue D-shaped handle occupies 60-73% across and 28-64% down, showing the wall through its open center. One blank horizontal cream oval spans 36-55% across and 36-53% down, centered at(46%,44.5%). These are approximate invisible instructions, never printed annotations.

The blue surface is matte and finely pebbled, with restrained cobalt shading. The oval is warm ivory with tiny natural speckles and a thin brown edge. Keep its horizontal ellipse shape and leave it entirely blank. The handle has rounded edges and a darker interior. Show the black lid slightly from above: a low stepped outer rim, dark oval recess and one irregular upturned drinking flap on its viewer-right, attached to the fitted cap.

The tabletop/wall boundary slopes gently from 54% down at the left edge to 64% down at the right. Pale honey oak fills the foreground with visible grain and a few knots. The wall is simply uneven off-white plaster. Only broad blurred diagonal gray LIGHT SHADOWS cross its left side; no visible plant, branch, window frame or other object is present there. Soft upper-left daylight brightens the mug's left face and casts its shadow rightward across the oak. Preserve ordinary tactile photography, natural subdued color and a mildly soft background.

Keep the inventory to this one mug, table, wall and their light/shadows. Avoid foliage, flowers, additional props, hands, steam, a saucer, writing or logos, a circular badge, a left handle, shiny metal lids, glossy plastic-looking blue, an elongated body, a base below 84%, a tighter crop, dramatic lighting or coordinate labels.
Open the exact flat event graphic prompt used for both final outputs
Generate a 4:5 portrait PNG of a plain flat event flyer by following the self-contained vector drawing specification below. This is a direct digital layout, not a redesigned poster or a photograph of paper. Preserve the specified text positions, normal Arial-like glyph proportions and uniform solid fills. There are no gradients, lighting, textures, shadows, bevels or perspective.

Use a 1200 by 1500 coordinate space for the whole image. The code specifies invisible drawing instructions; do not show code, XML, coordinate labels or color labels in the result. Render only its visible shapes and exactly six text regions. All lettering uses plain non-condensed sans-serif forms. If font availability differs, keep the visible glyph bounds and widths: the headline is approximately 652 pixels wide and 71 pixels high, rather than an oversized condensed title. The two subtitle lines together occupy about 434 by 77 pixels; the bold day/time line about 503 by 35; the venue about 263 by 33; the white button lettering about 297 by 35; the registration line about 385 by 29.

Keep the positions independent: do not shift the date or venue down to make a new text flow. The orange circle is clipped only slightly at the right canvas edge. The teal square, lavender rounded square and navy X remain separate, exactly ordered on the right. The bottom rule stays above the lower navy band.

<svg xmlns="http://www.w 3.org/2000/svg" width="1200" height="1500" viewBox="0 0 1200 1500">
<rect width="1200" height="1500" fill="#F4EFE5"/>
<rect x="0" y="0" width="1200" height="83" fill="#10263B"/>
<rect x="0" y="1414" width="1200" height="86" fill="#10263B"/>
<circle cx="1048" cy="288" r="168" fill="#EF6A47"/>
<rect x="915" y="492" width="196" height="200" fill="#55B8B1"/>
<rect x="840" y="730" width="271" height="282" rx="35" fill="#D7C9FF"/>
<path d="M 882 772 L 1061 951 M 1061 772 L 882 951" fill="none" stroke="#10263B" stroke-width="14"/>
<rect x="100" y="1030" width="561" height="130" rx="65" fill="#10263B"/>
<path d="M 100 1352 L 1101 1352" fill="none" stroke="#BFAC8E" stroke-width="5"/>
<g font-family="Arial, Helvetica, sans-serif" fill="#10263B">
<text x="100" y="258" font-size="96" font-weight="700">OPEN STUDIO</text>
<text x="100" y="510" font-size="42">Watch a live print demo</text>
<text x="100" y="556" font-size="42">and meet the makers.</text>
<text x="100" y="733" font-size="48" font-weight="700">SATURDAY 11:00-16:00</text>
<text x="100" y="832" font-size="48">RIVER HALL</text>
<text x="100" y="1270" font-size="32">Register: northline.example</text>
</g>
<text x="380.5" y="1120" text-anchor="middle" font-family="Arial, Helvetica, sans-serif" font-size="48" font-weight="700" fill="#FFFFFF">FREE ENTRY</text>
</svg>

Preserve exact capitalization, punctuation, the plain hyphen in 11:00-16:00 and the two subtitle lines. Avoid enlarged lettering, condensed type, reflowed blocks, any added words, photographic effects, color gradients, texture, a QR code, logo or visible drawing code.
Open the exact three-panel 3D character prompt used for both final outputs
Create a 2:3 portrait stylized 3D character sheet with three separate lavender studio scenes. FIRST fix the three horizontal scene slots on the full page: TOP0-30.1%, MIDDLE30.1-69%, BOTTOM69-100%. The top and bottom slots are almost equal in height; the middle is only about 1.3 times either one. Do not compress the top slot or let the standing pose determine the divider positions. Two thin white lines span the whole width at 30.1% and 69% down, around rows 462 and 1060 of a 1024x 1536 image.

Use the same sculpted adult man in every slot: tan skin, angular jaw, thick dark eyebrows, a predominantly pale silver-gray swept quiff with short darker sides, charcoal denim jacket with rolled sleeves and chest pockets, loose charcoal trousers, burnt-orange knitted fringed scarf and separated chunky black lace-up boots. Use sculpted animated-film-like features, cloth creases and woven scarf texture. All scenes have soft lavender wall and floor, gentle warm upper-left light and contact shadows.

Every scene includes a matte cobalt-blue horizontal beam resting on two chunky block legs with diagonal inner cutouts. No higher backrest or extra furniture parts. Preserve the furniture's deep blue against the lavender surroundings.

The following positions are LOCAL to each slot, measured from that slot's top; do not mistake them for whole-page positions. Nothing crosses a divider. In every slot include the entire figure and both boots, leaving a little floor beneath.

TOP LOCAL: face centered 44% across/18% down this slot; hands near 45% across/57% down; boot soles near 95% down. The man sits leaning forward, knees spread, both elbows on his knees and hands meeting between them with one wrist over the other. The stern face looks to the picture's LEFT. Scarf hangs between the legs. Beam extends 17-82% across, top around 53% of this slot, leg bases 90%. This seated figure should occupy nearly the complete slot height, without squeezing the slot shorter.

MIDDLE LOCAL: face 49% across/11% down this slot; boot soles 97% down. He stands wide-legged. The hand on the LEFT side of the picture goes into the trouser pocket near 42% across/38% down this slot; the hand on the RIGHT grips the scarf at chest near 52% across/26% down. Head turns firmly toward picture-RIGHT. The blue beam is behind the legs, top 67% down this slot, bases 90%.

BOTTOM LOCAL: face 56% across/15% down this slot; boot soles 93% down. He sits relaxed, legs apart, picture-left hand on the knee near 43% across/48% down; picture-right forearm along the beam top, hand hanging near 66% across/52% down. His head tilts up toward picture-LEFT with a small amused smile. Beam top 42% of this slot, bases 81%; scarf hangs down his front and over the seat.

Cross-check the full-page face centers near(44%,5.5%),(49%,34.5%),(56%,73.5%). All percentages and labels are invisible instructions. Avoid unequal top/bottom slot heights, shifted dividers, cropped feet, extra people, swapped hands, fused boots, dark hair replacing silver, a high-backed bench or printed numbers/text.
Open the exact cat portrait prompt used for both final outputs
Create a 4:5 vertical natural head-and-upper-body photograph of one adult long-haired black-and-white tuxedo cat, upright, front-facing at eye level. Include both entire pointed ears and the long white chest bib continuing out through the bottom.

Full-canvas layout has priority. Protect an empty warm gray-beige wall band from 0% to 9% down: no ear or hair enters it. Ear tips begin near 9-10% down. The cat begins near 9% down and the chest continues through the cropped bottom at 100%. The centered cheek fur spans about three quarters of the width. The visible background contains only the wall, without props. Viewer-left eye center is near 40% across/36% down, viewer-right eye near 62% across/36% down, nose near 50% across/46% down. The dusty-rose COLLAR STRIP ITSELF lies within 61-64% down; this range describes its entire visible band, not its center. Keep the collar tight and high around the neck, partly buried in fur. The ENTIRE single small oval tag occupies 65-73% down, with its center near 55% across/69-70% down. The tag hangs directly below the collar by one very short round loop, no long gap or chain. Its top begins immediately below the collar. Leave the long white bib below the tag through the lower quarter of the image. Positions refer to the full frame, are approximate, and must not appear as annotations.

Viewer-left eye is muted amber-gold; viewer-right eye pale gray-green, naturally sized with narrow vertical dark pupils and small reflections. Viewer-left pointed ear has one visible small irregular jagged notch at its upper tip; viewer-right ear is intact. Fine gray inner-ear hairs remain below the protected wall band. Black forehead and cheeks frame a narrow uneven white blaze between the eyes, widening into the white muzzle. The nose is pale dusty pink around a dark charcoal lower-central patch, leaving pink above and at both sides. A separate small vertical black chin patch sits below the mouth. Long fine white whiskers curve unevenly outward. The broad white chest has long uneven downward tufts between dark shoulders.

Use one blank matte-silver oval tag with no engraving. Gentle diffuse upper-left light creates soft shadows, textured black fur, subdued ivory whites and realistic flyaway hairs. The background is a softly blurred uneven warm gray-beige plaster wall. Preserve low saturation, restrained contrast and an ordinary tactile photographic appearance. Avoid clipped ears, a tighter face crop, giant or matching-color eyes, a collar low on the chest, long attachment links, low dangling jewelry, plastic fur, visible paws, extra pets, bows, dramatic lighting, lettering or coordinate labels.
Open the exact watercolor city prompt used for both final outputs
Create a 2:3 portrait watercolor-and-fine-ink waterfront-city illustration on warm ivory textured paper.

PRIORITY 1 — full-sheet layout. Keep the upper sixth (0-16% down) completely blank ivory; no wash or spatter enters it. The irregular painted patch fits approximately 8-94% across and 16-85% down, leaving paper below. The low waterfront bank, bridge deck and feet sit at 65% down. River fills the lower fifth, from 65% to 85%, fading into a ragged edge. Caption "ROTTERDAM" is small, thin, widely spaced uppercase, centered at 88% down.

Preserve the left-to-right order: low blocks near 12% across; tiny cream-gold sun centered (20%,62%) and half-hidden among roofs just above the bank; a narrow rectangular building centered 31% across, occupying roughly 49-65% down with a short antenna tip at 48%; two smaller rectangular buildings near 40% and 45%; a small rounded upright building near 60%; the bridge pylon near 73%, with its tip at 32% down and feet at 65% down. The left rectangular building is only one sixth of the page high, HALF the pylon's height above the bank. It stays one third across, not at the left border. These approximate positions refer to the FULL paper, not the inset painting, and must not be drawn as labels.

PRIORITY 2 — irregular shapes and connections. The right pylon is an angular thin blade with two unequal cap points, a shallow notch, and a short dark-blue segment that folds right then angles back around 36-39% down, forming an elbow. A broad dark blue-gray left leg slopes down-left to (66%,65%), and a narrower pale warm-gray right leg slopes down-right to (81%,65%), leaving a long thin triangular gap. Keep the BROAD LEFT leg itself DARK BLUE-GRAY from its upper bend to its left foot; keep the NARROW RIGHT leg itself PALE WARM-GRAY. These are different colored structural parts, not lighting effects. The cream/gold wash behind them must not turn the broad left leg white or beige. Do not exchange the leg colors. Fine cables fan down-left to a low gently curving deck, with two heavier stays to the right. A small dark-blue diagonally slatted structure attaches beside and overlaps the right foot at 74-93% across and 63-69% down, at the bank's depth; clear river remains below it. Loosely suggest the buildings with delicate gray ink and vertical facade marks. Keep the tiny partly visible sun, short rays and broken pale-gold river reflection.

PRIORITY 3 — painted appearance. Broad translucent granular blue-gray and muted lavender pigment blooms above and at right diffuse into pale cream and diluted gold at lower-left. Use irregular abstract blotches, pale gaps and paper grain, not repeated literal cloud bands. Small gray-blue ripples and creamy gold patches dissolve into the river edge near 85% down. Keep contrast subdued, ink thin, borders feathered and all pigment within the painting's bounds. Avoid orange-red sunsets, dense architectural detail, oversized buildings, an elevated large sun, a straight generic pylon, detached foreground fencing, hard rectangular frames, full-bleed paint, extra words or coordinate labels.

One Codex assistant performed analysis and assessment in a continuing conversation with earlier attempts available. Some final descriptions reuse earlier wording; selected source pixels and labelled enlargements supported the checks. The final template was developed on these five sources. We did not run a final test on new unseen images, an isolated captioning request or a blind panel. The generator did not expose its served model, seed or internal prompt processing.

Download the template and final five worked examples (about 35 MB). The complete two-phase test archive (about 247 MB) retains all 98 outputs, exact prompts and failure records. Those 98 outputs include earlier template versions; ten passing final outputs do not establish a future success rate.

The cat, mug and fictional 3D character sources are existing Grow AI Skills synthetic editorial assets; the event poster is an original editorial graphic. The watercolor is an existing AI illustration derived from Alexander Psiuk’s Rotterdam photograph, used under the Unsplash License.

If either result fails, revise the analysis prompt

Use the source and failed result in the analysis conversation. This correction asks for a better reusable template. Apply it to the source again, save the new returned prompt and generate twice without attaching an image:

Image A is the source. Image B is a result generated from text only.

CURRENT UNIVERSAL ANALYSIS TEMPLATE:
[Paste the reusable analysis template.]

RETURNED GENERATION PROMPT:
[Paste the complete text used to generate B.]

Identify the three visible differences that most affect the match. Explain whether the returned description omitted a feature, described it incorrectly, or gave conflicting instructions.

Revise the UNIVERSAL ANALYSIS TEMPLATE so that it asks for the missing information or resolves that ambiguity for other images too. Keep it general: do not insert this image's objects, colors, coordinates or wording into the reusable template.

Return the three differences and one complete revised universal template. I will apply it to the source again, save the new generation prompt, and test that unchanged text twice without attaching an image.

Name the missing information: complete ears with empty space above them, a tag’s bounds and short attachment, the exact hand-to-pocket relationship, or each text block’s independent placement. “Make it more accurate” adds no shape or position information.

If you are testing several sources, apply the revised template to all of them again. If only one repeat matches, retain both outputs and continue revising; showing only the better image would hide the variation you need to check.