Hand-Drawn Style Director helps you define the paper, drawing materials, line behavior, colors, and spacing for an AI illustration, then reuse those decisions across scenes. Install the Skill in Codex to produce a Style DNA card and a structured prompt. Image generation needs a compatible image tool and a separate generation request. The included two-scene test kept a similar medium and palette, but character proportions and background density changed.
Use it when you need an illustration system you can describe and revise: a set of article headers, a recurring story character, or a campaign with a shared visual treatment. For a single image, you can also use the style-card method below without installing anything. Exact character identity, brand colors, and fixed layouts still need inspection and sometimes manual editing.
Install the Skill and check the first response
Download the v1.0 Skill ZIP. It contains one hand-drawn-style-director folder with seven files: the instructions, interface metadata, an optional Python prompt composer, a sample JSON card, and three references covering style decisions, visual review, and the retained test. Inspect those files before installation. Python is optional for using the instructions; the composer requires Python 3.10 or newer and uses the standard library.
For one macOS or Linux project
Place the downloaded ZIP at the project's top level and run these commands there. This location keeps the Skill available to that repository.
install -d .agents/skills
python3 -m zipfile -e hand-drawn-style-director-skill-v1-0.zip .agents/skillsFor your macOS or Linux account
Run this from the directory containing the ZIP if you want the same Skill available across your projects.
install -d "${HOME}/.agents/skills"
python3 -m zipfile -e hand-drawn-style-director-skill-v1-0.zip "${HOME}/.agents/skills"For one project with PowerShell
Put the ZIP in the project directory, open PowerShell there, and extract it. Check an existing installation before using -Force, which can overwrite files you have customized.
$Target = '.agents\skills'
New-Item -Path $Target -ItemType Directory -Force | Out-Null
Expand-Archive -LiteralPath '.\hand-drawn-style-director-skill-v1-0.zip' -DestinationPath $Target -ForceThe project installation should end at .agents/skills/hand-drawn-style-director/SKILL.md. A second nested hand-drawn-style-director folder means the archive was extracted at the wrong level. OpenAI's skill documentation lists repository and user locations and says Codex detects changes automatically; restart Codex if the Skill does not appear.
Use $hand-drawn-style-director to create a Style DNA card and
a text-only prompt for an original hand-drawn editorial illustration.
Scene: at a neighborhood repair cafe, one volunteer repairs a bicycle
while its owner watches. Include exactly two people and one bicycle.
Use a 3:2 landscape composition with the repair activity on the right
and a quiet area on the left for a headline added separately.
Use broken pencil contours and light watercolor washes on warm paper.
Do not include lettering, a logo, or a border.
Do not generate the image yet.The response should contain the eight style fields, the traits to keep fixed, allowed variations, a short avoid list, and the image prompt. It should also say whether a reference image is needed. For this request, it should return text without generating an image; the Skill labels that result PROMPT_ONLY. If you get only a prompt, ask for the missing card so you can reuse its decisions.
Generate the first illustration and check it
Read the card before making an image. In the repair-cafe example, confirm that it allows two people and keeps the headline area clear. If it adds an unrelated traveler, scarf, or kitchen, ask Codex to remove those details. Once the card matches your brief, use this follow-up:
Generate one image using the repair-cafe prompt and Style DNA above.
Keep exactly two people and one bicycle, with the repair activity
on the right and the quiet headline area on the left.
Keep the agreed paper, pencil lines, watercolor treatment, and colors.
Use no reference image. Do not add text, a logo, or a border.
Return the image and point out any visible mismatch with the brief.Check that you can recognize the repair action, that the number of people and bicycles is correct, and that the headline area is usable. Inspect faces, hands, and bicycle parts at full size, then view the image in its intended page crop. If it fails one requirement, name that requirement in the next edit. If it passes, save the image, prompt, and card before requesting a second scene.
If the Skill is missing, check the final path, confirm SKILL.md exists, and confirm Codex is open in the intended repository. Use the exact $hand-drawn-style-director name. If Python is unavailable, extract the ZIP with your file manager and use the Skill without the composer. If Codex has no image tool, copy the prepared prompt into your chosen image generator.
Turn your brief into visible style decisions
Start with the subject, action, required objects, and destination. An article header needs room for a heading; a storybook page needs room for page text; a small icon needs a recognizable silhouette. Give the intended aspect ratio when you know it. Resolve conflicting requirements before generating: an opaque paper background and a transparent cutout require different surface instructions.
Keep the scene and style in separate parts of the brief. The scene might be a volunteer repairing a bicycle while its owner watches. The style might be broken pencil contours with light watercolor washes and quiet outer margins. Changing the bicycle to a kettle should not require choosing new drawing materials.
| Style field | What to specify | Quiet Paper Journey example |
|---|---|---|
| Surface | Background material and strength of grain | Warm off-white paper with restrained visible grain |
| Medium | The drawing and paint materials that should appear | Graphite, dark ink, and translucent dry-brush watercolor |
| Line logic | Line weight, breaks, pressure, and use of hatching | Thin broken contours, heavier downstrokes, sparse shadow hatching |
| Shape language | Silhouettes, proportions, and simplified features | Round head, dot eyes, compact clothing, small hands |
| Palette and value | Colors and their roles, including where dark marks belong | Graphite leads; orange marks the focal accent; blue and sage support |
| Texture and density | Where paint and detail accumulate and where they stay sparse | Dry paint gaps near the subject, little detail at the page edges |
| Composition | Focal placement, whitespace, depth, and border treatment | One focal group slightly below center, generous open paper |
| Mood and motion | Visible gestures and spacing that express the mood | Small gestures, gentle diagonals, still surroundings |
Style DNA is this package's name for a written description of the illustration's appearance. It gives you specific parts to revise. For example, line_logic becomes a line of prompt text when the composer runs; it does not change a model setting. The same description can still produce different results across images.
After the eight fields, define the traits that must stay fixed in locked_anchors. Put allowed changes in variation_budget. Use avoid for identifiable failures, such as an extra person or a busy background behind a caption. If a story character matters, include its proportions and costume in the anchors. If the subject changes every week, lock the medium and palette without locking a recurring person.
A complete card you can copy
This is the JSON card included in the ZIP. Save a working copy as my-style-card.json and edit its values. Keep every required field, use double quotes, and leave out comments and trailing commas. The nine text fields and three non-empty lists are required by the composer.
{
"name": "Quiet Paper Journey",
"surface": "Warm off-white paper with visible but restrained natural grain",
"medium": "Pressure-sensitive graphite and dark ink contours with translucent dry-brush watercolor fills",
"line_logic": "Mostly thin broken contours, occasional heavier downstrokes, and sparse hatching reserved for shadow",
"shape_language": "Simplified geometric objects and a round-headed character with tiny dot eyes, compact clothing, and small hands",
"palette_value": "Graphite leads; burnt orange is the focal accent; muted blue and sage support; most of the page remains warm paper",
"texture_density": "Watercolor blooms and dry gaps remain visible; detail clusters around the subject while the outer page stays quiet",
"composition": "One calm focal group slightly below center, low horizon, generous negative space, and no decorative border",
"mood_motion": "Quiet anticipation expressed through small gestures, gentle diagonals, and still surroundings",
"locked_anchors": [
"warm paper grain",
"broken graphite-ink contours",
"translucent watercolor fills",
"graphite, burnt orange, muted blue, and sage palette roles",
"one calm focal group with generous negative space"
],
"variation_budget": [
"scene and weather may change",
"camera distance may move from medium-wide to medium",
"lighting may shift between overcast daylight and soft evening warmth"
],
"avoid": [
"no glossy 3D shading",
"no photorealism",
"no logos or watermarks",
"no legible text unless supplied",
"no extra characters"
]
}The sample permits camera distance to change from medium-wide to medium. That is useful for varied storytelling, but too loose for a set of matching cards. For a fixed campaign, replace that permission with a common framing requirement. To retain the traveler's identity, add the accepted head-to-body proportion and clothing details after reviewing a base image. These are suggested refinements; the retained test below used the original prompts.
Choose a hand-drawn look beyond watercolor
Use a few compatible drawing materials and describe how their marks should appear. The table offers alternative wording for the card. These are starting recipes; the images tested on this page use the pencil-and-watercolor example.
| Look | Wording to try | What to inspect |
|---|---|---|
| Pencil sketch | Graphite on light paper; thin uneven contours; sparse pencil hatching; grayscale with one small color accent | Whether the silhouette and action remain clear when the image is small; whether hatching hides a required detail |
| Ink with flat color | Variable-weight ink contours; simple shapes; flat muted color fills; little surface grain; no dimensional shading | Whether outlines become too heavy and whether separate objects remain distinguishable |
| Matte gouache | Opaque matte painted shapes; rough brush edges; minimal outlines; limited colors; broad unpainted margins | Whether filled shapes merge together and whether the margins remain open |
| Pencil and watercolor | Broken pencil contours; translucent washes; dry paint gaps; visible warm paper; detail near the subject | Whether paint and paper texture compete with the subject, captions, or small features |
Changing the medium also requires checking the other fields. For ink with flat color, replace watercolor blooms in texture_density and remove translucent washes from locked_anchors. For gouache, avoid retaining a rule that all fills must show paper through them. For the repair-cafe scene, replace the sample's traveler-specific shape description with the two requested people and remove any unwanted character details.
To choose between two looks, keep the repair-cafe scene and framing the same and request one image in each treatment. Compare them in the actual header crop. Choose the version in which the repair action reads clearly and the headline has room. Change the scene after choosing the treatment, so you can distinguish a style difference from a composition difference.
Choose how references enter the image
Select a route before writing the final prompt. Start with text when the written card describes the appearance adequately. Use an accepted generated image when the next output needs visual identity information. Use an authorized external reference when its actual marks, materials, or subject details are necessary.
| Route | Best starting point | What to check |
|---|---|---|
text-only | A new illustration style or unrelated subjects sharing a medium | Whether each output retains the written palette, line, and density rules |
generated-anchor | A recurring character or an established campaign appearance | Whether the anchor meets the brief before reusing it; whether later outputs copy an unwanted pose |
authorized-reference | Your own photo, licensed artwork, or another image you have permission to use | Which details may transfer and which must be replaced |
A photo of your subject and a drawing used for style have different jobs. Label them explicitly: image 1 supplies the subject's geometry; image 2 supplies broad line and paint treatment. State the new action and setting in words. OpenAI's image prompting guide recommends assigning roles to references and naming what each edit must preserve or change.
Style reference only. Preserve broad line quality, medium feel,
palette tendency, texture, and visual density. Ignore and replace
the reference's subjects, people, animals, objects, setting, action,
pose, camera angle, composition, layout, text, logos, and story elements.
The written scene is the sole content source.Use that clause only for a reference whose people, objects, and setting should be discarded. If a supplied photo is the subject you want to preserve, state which details must remain. Otherwise, telling the model to ignore all reference subjects conflicts with your reason for supplying the photo.
Image 1 is my subject reference. Preserve the building's roofline,
doorway, window arrangement, and street position.
Image 2 is a permitted style reference. Use only its broad pencil-line
and watercolor treatment; do not copy its building, people, or layout.
Illustrate the building from image 1 using the written Style DNA.
Do not add buildings, signs, people, or lettering.This request assumes you supply two images in those roles. For several style references, decide which one controls the lines, which controls the palette, and which controls the paper. If one shows dense crayon and another shows delicate ink, choose the treatment you want before generating. You can describe the selected traits in text when the images disagree or the tool accepts fewer inputs.
The available image tool must support the intended reference or edit input. OpenAI's image-generation documentation describes generation and editing workflows; the Skill ZIP does not add those capabilities to a tool. If references are unavailable, use the text-only prompt and record that no image input was used. A written isolation instruction still requires visual review of the result.
Adapt the card to different illustration jobs
The following briefs show how to change the requirements for different destinations. They are suggested requests, not additional generated tests. Keep the style fields that serve the project and replace the ones that conflict with its output.
An editorial header with room for a headline
For the repair-cafe example, specify who repairs what and where the headline will sit. A wide header can place the people and bicycle on the right, with a quiet area on the left. Put that placement in the composition field and keep background detail out of the headline area. Request the illustration without words if the site will add the heading as HTML text.
Inspect the actual mobile crop as well as the wide image. If the bicycle or the repair action disappears, move the people and bicycle into the area shared by both crops or prepare a separate narrow composition.
A storybook series with the same traveler
Choose a base image in which the face, coat, scarf, and body proportions are visible. Review it against the character description before using it as an anchor. Save the accepted image and the card together. Change the action and setting in later requests while repeating the identity details that matter.
A close-up naturally makes the head occupy more of the page. Compare head size relative to the body, rather than judging only its size on the canvas. For a rear view, facial traits are not-visible; costume and silhouette can still be reviewed. If several views disagree, settle which image defines the character before producing more pages. The character-consistency workflow covers that identity work in more detail.
A social carousel or brand illustration set
Use the same frame dimensions, caption area, line treatment, and palette roles across cards. Change the scene or object to communicate each card's message. A tea scene and a bus-stop scene can share paper, line, and color treatment while using different object arrangements.
Write whether a brand color is an accent, a background, or a clothing color. A hex value alone does not guarantee an exact rendered color, particularly in translucent paint and shaded areas. If exact brand color is required, check the exported artwork and apply the approved colors in your design editor. Add a supplied logo as a separate asset when its shape and lettering must be exact.
A teaching diagram or illustrated instructions
Specify the objects, relationships, arrows, and sequence before selecting a decorative treatment. For bicycle maintenance, a wheel, a pump, and the correct valve connection must remain recognizable. Simplifying the wheel into an attractive circle is acceptable only if the instructional relationship remains clear.
Keep texture light around connections and labels. Put labels and precise arrows in a layout editor when correctness matters. Compare the completed diagram with the source instructions, including arrow direction and step order. Use hand-drawn marks for illustration; preserve the technical information that the reader must follow.
A small icon, sticker, or transparent cutout
Replace warm paper with a transparent ground if the asset must sit on different backgrounds. Reduce grain and hatching, simplify the silhouette, and describe whether the contour stays thin or becomes heavier for the small destination. Ink with a flat color accent is an alternative to the sample's layered watercolor treatment.
Inspect the icon at its actual display size. If eyes, fingers, or fine pencil marks disappear, simplify those features before producing the rest of the set. Verify that a transparent export has actual transparency and view its edges on light and dark backgrounds. A visible checkerboard drawn into the image is not a transparent background. A hand-drawn raster illustration also needs separate vector work if you require an editable vector logo.
A photo transformed into watercolor
Use the photo as the subject reference and list the landmarks that must remain recognizable: the roofline, doorway, street arrangement, or a person's clothing. Put the watercolor treatment in the style fields. Keep the source composition only if it suits the intended print or page; say explicitly when cropping or simplifying the background is allowed.
Review the recognizable features first, then inspect paint and paper treatment. If the building loses its defining windows, ask to restore their arrangement while preserving the illustration treatment. For a one-off travel image, the Watercolor Travel Photo Prompt and photo-to-watercolor guide may be enough without installing a reusable Skill.
Compare the two original test images
On September 6, 2026, Grow AI Skills generated these two original scenes in separate calls through Codex's built-in image tool. Both used text prompts with the Quiet Paper Journey treatment. Neither received a reference image or an artist name. The tool did not identify its underlying image model. The exact prompts used for those calls are copied below.
Original prompt A: rainy bus stop
Use case: illustration-story
Asset type: first behavior-test image for an English tutorial about repeatable hand-drawn visual style
Primary request: an original round-headed traveler character in a burnt-orange scarf and muted-blue coat waits under a small bus shelter in gentle rain, holding a folded paper map; a bicycle leans against the shelter; wet pavement and three sparse weeds establish the place
Scene/backdrop: quiet neighborhood bus stop, no signage, no readable text
Style/medium: original hand-drawn editorial storybook hybrid; irregular pressure-sensitive graphite and dark ink contours; translucent dry-brush watercolor fills; visible warm off-white paper grain; simplified geometric forms; imperfect handmade edges
Line logic: mostly thin broken graphite-ink contours, occasional heavier downstrokes, minimal hatching used only for shadow
Shape grammar: round head, tiny dot eyes, compact rectangular coat, small hands, simplified objects with slightly uneven perspective
Composition/framing: 3:2 landscape, one calm focal group slightly below center, generous warm-paper negative space, low horizon, no border
Color palette: graphite #2B2A28, burnt orange #D46A3A, sage #7D8F69, muted blue #5F7E9A, warm paper #F3E9D2
Lighting/mood: overcast but comforting, quiet anticipation
Constraints: completely original visual language; exactly one traveler; no artist names; no copyrighted characters; no logos; no watermark; no legible text; no glossy 3D; no photorealism; do not imitate a recognizable living artist
Original prompt B: evening tea
Use case: illustration-story
Asset type: second behavior-test image for an English tutorial about repeatable hand-drawn visual style
Primary request: the same original round-headed traveler character in the same burnt-orange scarf and muted-blue coat sits at a small wooden kitchen table in the evening, pouring tea from a simple ceramic pot into one cup; a folded paper map rests beside the cup; one potted herb and a curtained window establish the place
Scene/backdrop: quiet modest kitchen, no signage, no readable text
Style/medium: original hand-drawn editorial storybook hybrid; irregular pressure-sensitive graphite and dark ink contours; translucent dry-brush watercolor fills; visible warm off-white paper grain; simplified geometric forms; imperfect handmade edges
Line logic: mostly thin broken graphite-ink contours, occasional heavier downstrokes, minimal hatching used only for shadow
Shape grammar: same round head, tiny dot eyes, compact rectangular coat, small hands, simplified objects with slightly uneven perspective
Composition/framing: 3:2 landscape, one calm focal group slightly below center, generous warm-paper negative space, low visual density, no border
Color palette: graphite #2B2A28, burnt orange #D46A3A, sage #7D8F69, muted blue #5F7E9A, warm paper #F3E9D2
Lighting/mood: soft amber evening light, quiet relief after travel
Constraints: completely original visual language; exactly one traveler; preserve the character's scarf, coat, proportions, face language, and palette; no artist names; no copyrighted characters; no logos; no watermark; no legible text; no glossy 3D; no photorealism; do not imitate a recognizable living artist
| Review area | Visible observation | Recorded result |
|---|---|---|
| Surface, medium, and lines | Warm paper, graphite-ink contours, and watercolor with dry gaps appear in both scenes | pass |
| Palette roles | The traveler retains an orange scarf and blue coat; muted green supports the scenes | pass |
| Shape language | The tea scene has a larger head relative to the character's body | partial |
| Texture and density | The kitchen contains more background marks | partial |
| Composition | The tea scene is framed closer and leaves less open paper | partial |
| Reference isolation | No reference image was supplied | Not tested by these calls |
The original review marked composition as partial. Interpret that alongside the sample card, which permits a closer camera distance: closer framing alone is an allowed change, while the reduced open space matters to the intended spacious treatment. If equal character scale is essential to your series, tighten the framing requirement before generating its next scene.
The two images have a similar visual treatment with visible differences. There was no comparison without the card, so this example cannot establish that the card reduced drift. The same editorial workflow wrote the prompts and reviewed the images; there was no blinded review or comparison across models. The ZIP includes both prompt texts, image-file hashes, and the review notes. The other illustration briefs, reference requests, and proposed repair in this guide have not been tested with generated images.
Repair a failed image with a specific change
Compare the output with the written card and the accepted image, then name the failed requirement. A repeated wrong coat color still fails even when every image repeats it. Use pass when a visible requirement is met, partial for a visible mismatch within otherwise usable work, fail for a rejected requirement, and not-visible when the image does not expose the detail.
| Visible problem | Inspect first | Next change to try |
|---|---|---|
| Character looks different | Head-to-body proportion, face placement, costume, and silhouette | Restate the failed identity trait and provide the accepted subject anchor if supported |
| Background becomes busy | Whether the scene requests too many objects or the card allows too much detail | Remove optional objects and define where background marks must remain sparse |
| Artwork looks glossy or three dimensional | Whether lighting or material instructions conflict with the drawing medium | Specify matte paint, flat washes, and the intended contour behavior |
| Colors change their roles | Whether the same accent has spread into large background areas | Assign the accent to specific objects and reduce its permitted background use |
| Reference pose or setting is copied | Whether the reference was labeled as subject, style, or edit target | Clarify its role and replace the pose and setting in the scene instructions |
| Words are wrong or unwanted | Whether exact wording was supplied and all text is readable | Correct the text separately or remove it and add typography in the layout |
| Small illustration loses detail | The actual destination size and background contrast | Simplify the silhouette and remove texture that hides required features |
| Two instructions conflict | Transparent ground versus paper, fixed crop versus changing camera distance | Resolve the conflict in the card before another generation |
For the tea scene, start by reducing background detail while keeping the traveler. The following proposed request assumes you supply the tea image to edit.
Edit the supplied evening-tea image. Reduce background drawing
and texture on the wall around the traveler so more warm paper shows.
Keep the traveler, face, head-to-body proportion, scarf, coat, hands,
teapot, cup, map, foreground table, and framing unchanged.
Preserve the broken graphite-ink contours and translucent watercolor.
Do not add objects or text.Inspect the edited result for changes to the hands, map, character, and table before accepting the quieter wall. Correcting density does not correct the head proportion; handle that as a separate decision. If only one local area is wrong, editing an accepted image may preserve more useful work than generating the whole scene again. If the scene or reference role is wrong throughout, revise that requirement before starting a new image.
Each edit can still alter details you asked to preserve. Restate the important constraints and compare against the original. For artwork that requires unchanged pixels outside the repaired area, composite the accepted local correction into the original in an editor. Prompting alone cannot establish pixel identity.
Continue only while each attempt addresses the named problem without breaking a requirement you need to keep. If the same detail keeps changing, stop producing the remaining scenes and simplify the request, change the reference, or finish that detail in an editor. If a tool change alters the style, check one representative scene before generating the whole series again.
Prepare the illustration for its destination
For a website or slide, view the export at the size where readers will see it and check the actual crop. For a transparent asset, inspect the edges over both light and dark backgrounds. For print, ask the printer for the required file dimensions and color workflow, then check the exported file against those requirements. An aspect-ratio instruction in a prompt does not establish that the file is suitable for the final print size.
If the available export is too small or fine marks become unreadable, use a larger supported output or simplify the artwork and check it again. Upscaling is a separate step that needs another visual check. Keep exact logos, labels, and technical arrows as separate editable elements when those details must survive resizing.
Use the optional prompt composer
The script assembles text from your card and scene. It does not call an image model, load a reference image, or review a generated result. Run it from the installed Skill directory so the relative script and card paths resolve.
cd .agents/skills/hand-drawn-style-director
python3 scripts/compose_prompt.py \
--style-card assets/sample-style-card.json \
--subject "a traveler reading a folded map at a rainy bus stop" \
--use "editorial story illustration" \
--ratio "3:2 landscape"For a personal installation, first change into "${HOME}/.agents/skills/hand-drawn-style-director". In PowerShell, use the installed directory and the Python command available on your machine, such as py -3, and put the arguments on one line. The backslashes above are shell line continuations for macOS and Linux.
The output starts with Intended use: and Primary request:, then prints framing, the style fields, anchors, variations, and avoid rules. Copy that text into your image tool or save it as a prompt file. The --ratio value is a written composition instruction; it does not set the tool's export dimensions.
Point --style-card at your own JSON file to use a different style. Add --text "Repair Cafe" when those exact words belong inside the image, then inspect their spelling. Add --json if you need structured output: it returns style_name, reference_plan, and prompt. The script always reports text-only; supply and label reference images separately through a compatible image tool.
For a missing-file error, check your working directory and file paths. For invalid JSON, check quotes, commas, and required fields. The script rejects designated artist-name keys and explicit style of wording in card values or command-line inputs. It does not recognize every person's name. If it rejects a style description, rewrite that description using the intended materials and marks.
Keep the accepted style usable for the next scene
Save the working card, exact prompt, reference images with their roles, and chosen output together. Record the image tool and settings when available, plus the requirement that passed or failed and the next intended change.
For a multi-view character sheet, use the Character Reference Sheet Prompt to plan view order and visible identity details. For a motion deliverable, continue with the AI product video workflow once the source artwork is accepted. This Skill's mood-and-motion field describes visible gesture in a still image; the ZIP does not animate it.
The package adapts numbered-selection and reference-routing ideas from yang0/handraw-style at commit e51f4f8. That checked tree had no LICENSE file. The download contains original instructions and a sample card, without its 216-image gallery, artist records, source code, or model-capability table. Gallery numbers from that project do not select styles in this Skill.
Before using a finished illustration, check its crop, required text, character details, and the permissions for its source images. The prompt composer cannot clear rights or approve a brand asset. Save any unresolved correction beside the card and chosen image so you can address it before making the next scene.
