@99percentpeople/pi-codex-api 0.1.0

package/skills/gpt-image-prompts/SKILL.md

@@ -0,0 +1,282 @@
+---
+name: gpt-image-prompts
+description: Craft and refine production-ready prompts for GPT Image 2. Use when a user needs an image-generation prompt, a stronger visual brief, prompt variations, or precise edit instructions for photos, illustrations, posters, ads, product visuals, characters, diagrams, or reference-guided images.
+license: MIT
+compatibility: Prompt-only guidance optimized for GPT Image 2. It does not require or prescribe an image tool, API, credential, or file workflow.
+---
+
+# GPT Image Prompts
+
+Write clear, visually specific prompts for GPT Image 2. Preserve the user's idea
+and improve only the details that materially increase control, composition, and
+image quality.
+
+This is a **prompt-writing skill only**. Do not turn its output into a tool
+manual, API request, command, path policy, authentication flow, or error guide.
+Tool selection and execution are outside this skill.
+
+## Output contract
+
+- Return a ready-to-use image prompt, not a tutorial, unless the user asks for
+  an explanation.
+- Keep simple requests to one to three clear sentences. Use short labeled lines
+  for complex scenes, edits, layouts, or exact in-image text.
+- Preserve the user's language and all exact quoted text. Do not translate,
+  paraphrase, or "improve" required copy unless asked.
+- If the user asks for multiple options, make each variation intentionally
+  different in art direction, composition, palette, or mood—not minor synonym
+  changes.
+- Do not include tool names, model parameters, JSON, API fields, output paths,
+  credentials, or implementation steps in the final prompt.
+- Ask one concise question only when a missing detail would make the result
+  materially wrong, such as the exact required text, the edit target, or a
+  contradictory composition. Otherwise make the smallest reasonable inference.
+
+## Core method
+
+### 1. Preserve intent
+
+Identify the user's non-negotiables before adding detail:
+
+- what the image is for;
+- the main subject and action;
+- required style or medium;
+- required composition or orientation;
+- exact text that must appear;
+- elements that must stay unchanged;
+- elements that must not appear.
+
+Do not invent extra characters, props, slogans, logos, brand colors, story
+beats, or side-specific placement that the request does not imply.
+
+### 2. Choose the right amount of detail
+
+- If the request is already specific, normalize it into a clean visual brief
+  without adding creative requirements.
+- If it is vague, add only high-value visual decisions: a useful framing,
+  coherent lighting, a compatible palette, and concrete material or texture
+  cues.
+- Prefer observable descriptions over vague praise. Use "soft window light from
+  the left, low contrast, warm highlights" instead of "beautiful lighting".
+- Prefer a few strong constraints over a long negative-prompt inventory.
+
+### 3. Structure the prompt
+
+Use this order when relevant:
+
+1. **Purpose and format** — editorial illustration, campaign image, book cover,
+   product visual, poster, diagram, character sheet, and so on.
+2. **Scene or backdrop** — place, era, weather, environment, and depth.
+3. **Subject and action** — appearance, pose, gaze, scale, interaction, and the
+   focal moment.
+4. **Style or medium** — photograph, watercolor, gouache, ink, oil, 3D, collage,
+   printmaking, flat graphic design, and its concrete visual traits.
+5. **Composition** — framing, viewpoint, perspective, subject placement,
+   hierarchy, negative space, and crop.
+6. **Lighting and color** — direction, softness, contrast, time of day, palette,
+   atmosphere, and mood.
+7. **Materials and finish** — surfaces, fabric, skin, paper, brushwork, grain,
+   imperfections, or print texture.
+8. **Text** — exact wording, typography, size, color, placement, and frequency.
+9. **Constraints** — what to preserve, what to avoid, and what must not change.
+
+Not every prompt needs every section. Remove empty or low-value lines.
+
+## Visual direction
+
+### Photorealistic scenes
+
+Use `photorealistic` explicitly when realism is the goal. Describe the image as
+a real moment being photographed:
+
+- camera distance and framing: close-up, medium shot, wide shot, full body;
+- viewpoint: eye-level, top-down, low angle, over-the-shoulder;
+- natural light and atmosphere;
+- believable skin, fabric, material wear, reflections, and imperfections;
+- candid pose, gaze, and object interaction;
+- restrained processing when the image should feel documentary or unstaged.
+
+Camera and lens language is useful for overall look, but do not overload the
+prompt with technical specifications that do not affect the requested image.
+
+### Illustration and painting
+
+Describe visual properties rather than relying on an artist's name:
+
+- medium and surface: watercolor on cold-press paper, thick oil on canvas,
+  gouache, charcoal, risograph, woodblock, cel animation, digital matte painting;
+- mark-making: loose washes, dry-brush texture, visible impasto, fine crosshatch,
+  clean geometric shapes, soft pencil underdrawing;
+- edge treatment: crisp silhouette, lost-and-found edges, soft atmospheric
+  contours, bold ink outline;
+- shape language and depth: rounded friendly forms, angular silhouettes,
+  layered foreground/midground/background;
+- palette and value structure: muted earth tones, limited duotone, luminous
+  jewel colors, high-key pastels, deep chiaroscuro;
+- finish: handmade irregularity, subtle paper grain, screen-print misregistration,
+  polished animation background.
+
+For a consistent character, repeat the defining facial features, proportions,
+outfit, palette, and personality in every revision.
+
+### Composition and people
+
+When composition matters, specify:
+
+- orientation or canvas intent: portrait, landscape, square, panoramic;
+- focal point and visual hierarchy;
+- foreground, midground, and background relationships;
+- usable negative space only when copy or UI genuinely needs it;
+- body framing such as "full body visible, feet included";
+- gaze and action such as "looking at the open book, not at the camera";
+- hand/object interaction such as "both hands naturally gripping the handle".
+
+Use left/right placement only when the user or intended layout requires it.
+
+## Text inside images
+
+Text is a visual constraint, not an afterthought:
+
+- Put exact copy in quotes or ALL CAPS.
+- Say `verbatim`, specify that it appears exactly once, and prohibit extra text.
+- Specify typography category, weight, color, alignment, placement, and relative
+  size.
+- Keep copy short where possible.
+- For an uncommon name or difficult word, add a spelling cue such as
+  `S-T-R-I-P-E`, while retaining the normal verbatim text that must be rendered.
+- For posters, diagrams, and infographics, define hierarchy and reading order;
+  require sharp, legible text and adequate whitespace.
+
+Example constraint:
+
+```text
+Text (verbatim, exactly once): "WEEKLY PLAN"
+Typography: bold white geometric sans-serif, centered at the top, large headline
+No other words, letters, logos, signatures, or watermarks.
+```
+
+## Reference images
+
+When images are provided, assign each a role instead of treating every image as
+an edit target:
+
+```text
+Image 1: base scene and composition reference.
+Image 2: subject identity reference.
+Image 3: color and brushwork reference only.
+```
+
+Then state the relationship explicitly: what to borrow, what to transplant,
+what to change, and what must remain untouched. For compositing, specify
+placement, scale, perspective, lighting, contact shadows, and occlusion.
+
+## Precise edits
+
+Lead with the single requested change, followed by a preservation lock:
+
+```text
+Change only <target change>.
+Preserve <identity, geometry, pose, framing, lighting, background, typography>.
+Keep everything else exactly unchanged.
+Do not add <unwanted elements>.
+```
+
+Repeat the preservation lock on every edit iteration. For identity-sensitive
+work, explicitly preserve face, facial features, skin tone, body shape,
+proportions, pose, expression, hair, and camera angle as applicable.
+
+## Iteration
+
+Improve images through small, diagnosable revisions:
+
+1. Establish the core scene and composition.
+2. Review the mismatch in concrete visual terms.
+3. Change one major variable at a time.
+4. Restate critical invariants so the image does not drift.
+
+Good follow-ups:
+
+- `Keep the same composition; make the window light warmer and softer.`
+- `Change only the jacket to dark navy wool; preserve identity and pose.`
+- `Remove the extra tree; keep the landscape, camera angle, and palette unchanged.`
+
+Avoid rewriting the entire prompt when a targeted correction is enough.
+
+## Prompt templates
+
+### New image
+
+```text
+Purpose: <intended asset and audience>
+Scene: <setting, era, weather, backdrop>
+Subject: <main subject, appearance, action, gaze, interaction>
+Style/medium: <visual medium and concrete stylistic traits>
+Composition: <orientation, framing, viewpoint, placement, negative space>
+Lighting/color: <direction, softness, contrast, palette, mood>
+Materials/texture: <important surfaces and finish>
+Text (verbatim): "<exact copy>"
+Constraints: <must include; must preserve; must avoid>
+```
+
+### Edit
+
+```text
+Primary change: Change only <specific target>.
+Preserve: <identity, geometry, pose, composition, lighting, background, text>.
+Integration: Match <perspective, scale, material behavior, shadows, color temperature>.
+Constraints: Keep everything else exactly unchanged; <additional exclusions>.
+```
+
+### Reference-guided image
+
+```text
+Image 1: <role>.
+Image 2: <role>.
+Primary request: <new scene or transformation>.
+Borrow from references: <identity/style/palette/composition cues>.
+Preserve: <locked attributes>.
+Do not copy or add: <logos, unrelated text, extra elements>.
+```
+
+## Examples
+
+### Natural photograph
+
+```text
+Create a photorealistic candid photograph of an elderly sailor adjusting a net
+on a small fishing boat while his dog rests nearby. Medium close-up at eye
+level, soft coastal daylight, shallow depth of field, subtle film grain, natural
+color, weathered skin, worn fabric, and salt-aged wood. The moment should feel
+honest and unstaged—no glamorization, heavy retouching, text, or watermark.
+```
+
+### Painterly illustration
+
+```text
+Create a quiet fantasy editorial illustration of a lone herbalist crossing a
+misty alpine meadow at dawn. Hand-painted gouache on textured paper, simplified
+rounded shapes, soft dry-brush edges, layered atmospheric depth, muted sage and
+umber palette with a small amber accent on the lantern. Wide landscape
+composition with the figure small against the mountains; no text, logo, or
+watermark.
+```
+
+### Poster with exact text
+
+```text
+Create a minimal vertical event poster built from bold cobalt and warm ivory
+geometric shapes with generous negative space and crisp screen-print texture.
+Text (verbatim, exactly once): "NIGHT MARKET". Set it in a heavy condensed
+sans-serif, ivory, centered in the upper third as the dominant element. No
+other words, letters, logos, signatures, or watermarks.
+```
+
+### Surgical edit
+
+```text
+Change only the white dining chairs to natural oak chairs with realistic wood
+grain. Preserve the exact room geometry, camera angle, table, walls, floor,
+lighting direction, contact shadows, and all surrounding objects. Match the
+original perspective and color temperature. Keep everything else exactly
+unchanged; add no new decor or text.
+```