| name | editorial-fashion-six |
| description | Six fixed 3:4 editorial fashion storyboard images from the same SKU's multi-angle white-background product photos. Covers brand ID, layered non-flat scenes (reference-scenes), styling (reference-outfit), poses (reference-pose-vocabulary), optional web research, triple-choice, hero-image anchoring, HTTP nano-banana generation, and per-frame QC with targeted regeneration only; bright natural daylight; do not paste full prompts in chat. Use for six editorial, magazine, or mood shots, lookbook storyboards, or nano-banana synced generation for one garment. Triggers include requests for six frames, six editorial images, storyboard six, magazine-style, nano banana image gen. Do not use for eight-slot global PDP or marketplace detail-page matrices—use ecommerce-fashion-workflow instead. |
| metadata | {"version":"1.2.0"} |
| compatibility | Network for optional web search and nano-banana image generation (HTTP or IDE-integrated); env NANO_BANANA_BASE_URL optional when using the HTTP API. No Python scripts in this skill package. |
White-background product → 6 editorial fashion shots
Critical rules (read first)
- Output count and ratio: exactly six images, all 3:4; same
image_size across the set (recommended 2K).
- Phased generation: produce frame 1 first, user confirms URL, then frames 2–6 with frame 1 as hero anchor—never all six at once.
- Scope: this skill is not the eight-image Amazon/Shopify PDP workflow—route that to
ecommerce-fashion-workflow.
- Privacy / UX: do not paste full generation prompts in chat unless the user explicitly asks; share triple-choice, frame labels, URLs, checkpoints.
- SKU fidelity: frame 1 must match the exact front flat; if QC fails, regenerate that frame only before continuing.
- Bundled references: depth and taboos—reference-scenes.md; outfit formulas—reference-outfit.md; pose vocabulary—reference-pose-vocabulary.md; model tiers, body/face density, composition scale—reference-model-presets.md (load when executing Step 4 and QC 1/4/9).
Input: multi-angle product images
Users usually supply several angles of the same SKU (white-background or flat lay: front, side, back, etc.). Before running, map angles: from filenames or user notes, label which image is front / side / back; later shots pick the matching product reference. If client assets cannot reliably distinguish front vs back from filename, notes, or image structure (symmetrical styles, only one image, conflicting labels, etc.), ask the client first which is front and which is back (or ask for relabeled assets)—do not guess front/back without confirmation. Do not ask by default whether a back view exists—do not ping the client only for “do you have a back image.” Shot 4 in this skill is a front variation by default (see step-6 table) and does not require a back image. If an angle is actually missing (e.g. no side), substitute sensibly for that frame and briefly tell the user.
When to use this skill vs the 8-image e-commerce skill
| Need | Use |
|---|
| 6 images, magazine / mood / story feel; unified scene and light | This skill (editorial-fashion-six) |
| Global PDP 8-slot matrix (Amazon / Shopify / DTC–style detail page: front / back / side / top–bottom detail / macro / flat lay / styling) | ecommerce-fashion-workflow |
Shared ideas: white-bg reference, hero anchoring, de-gray and lift. This skill is 6 fixed frames (including shot 4 front variation and fabric close-up); it differs from the 8-slot matrix in count and shot mix.
Examples
Example A — Magazine six-pack
- User: “Same hoodie white-bg front/side—I need six magazine-style street editorial shots, nano banana.”
- Agent: Map angles; brand ID; source QA; triple-choice; frame 1 with front flat → confirm URL; frames 2–6 with anchor + angle refs; QC each frame; save to
outputs/editorial-six/….
Example B — Client sheet + Excel row
- User: Uploads images + “SKU-A—the client sheet says coffee-shop window scene.”
- Agent: Required scene column wins over defaults; still layered foreground/mid/background per reference-scenes; execute same six-frame rhythm and QC.
Prompt visibility (chat)
- Internally still write
prompt and run nano-banana image generation (same parameters as below); do not paste the full prompt in the conversation (see Critical rules).
- Show the user: triple-choice summary, frame + shot type, result URL or image, checkpoints; full prompts only if the user asks.
Direct image generation (nano banana / HTTP)
- Implementation: Use nano-banana image generation available in the current environment (Fancybos-style HTTP API or IDE-integrated tooling). No Python scripts ship with this skill package.
- Parameters (unified across the set):
ratio fixed "3:4" (all 6 the same).
image_size recommended "2K" (match production; keep consistent).
app_model_type default "nano-banana-2" unless user specifies pro.
- Frame 1 may need longer runs—increase
timeout (seconds); poll_interval ~2s is fine.
- Optional env:
NANO_BANANA_BASE_URL overrides default http://ai-api.fancybos.com (test/prod).
- Frame 1:
img_urls = user’s front product image (URL or local path).
- Frames 2–6: frame 1 output URL as hero anchor; per shot combine matching-angle product refs with anchor in
img_urls (e.g. [frame1_URL, side product]); count per API limit. Frame 4 is front framing (see table): img_urls = front product + frame 1 anchor; not default back view; no back asset required.
Step 1: Brand ID and scene recommendation
- Identify brand: from user copy, product logo/hang tag, filenames, or client sheet—brand (or line) when possible.
- If brand is clear: recommend scene and light in that brand’s usual visual tone (e.g. American street, Korean fresh, outdoor tech, quiet luxury minimal) and why it fits the silhouette; mirror public campaign location types when appropriate—do not invent off-brand mood. Summarize typical model hair/makeup mood (influencer-polished, outdoor natural, sporty clean, etc.) for step 4.
- Background reference (avoid flatness): if word lists alone yield empty, flat, single-plane backgrounds, web-search same-brand, same category/season campaigns, lookbooks, official social—abstract richer layered elements into internal prompt; do not copy whole copyrighted frames. Without brand, search category benchmarks. When locking location, follow reference-scenes.md “Scene diagram: layering and light vocabulary” · §1 How to use: (1) Lock scene family from triple-choice and MLB quick reference; (2) Foreground / midground / background each at least one clear element; (3) add light (§3 there); (4) ground and footing (§4); (5) taboos per MLB and main Skill. Background defaults to avoid: no default luxury hotel lobby, office elevator hall (symmetrical elevators, glass shafts, polished stone, cold top light, “HQ” corridor)—prefer outdoor natural or lifestyle interior (coffee/dessert, window-lit casual). Optionally
avoid hotel lobby, office elevator hall, symmetrical elevator doors, marble corporate corridor (don’t stack); if client Excel explicitly wants such interiors, follow that job.
- No brand or unclear: recommend from category and experience; layer reference-scenes.md MLB quick scenes and taboos; if client Excel has required scene, that row wins.
- For the user: short takeaway only—no long essay.
Step 2: Source image QA
Check each new garment batch: light too dark, creases/shadows crushed, glare, heavy crop, messy background.
- If risky: note possible inherited creases/shadows/crop; ask for cleaner white-bg / flat or confirm proceed.
- If OK, go to triple-choice.
Step 3: Style and narrative (triple-choice)
From silhouette, fabric, audience, and step 1 (including web layering), give 3 differentiated directions; each has scene, light mood, emotional tone (2–4 sentences); scene must show depth (avoid one blank wall / solid cyclorama). Default bright natural light (clear white daylight)—airy, clean, slightly higher saturation; avoid gray murk and dirty shadows; if brand/user wants warm / golden hour / film cast, state in triple-choice or that option; if brand wants low-key, state separately. After pick, internal prompts stay on that rail.
- No lazy templates (hard rule): same as root Skill—no rotating place nouns with identical light/story/depth; don’t default to terrace/seaside/café window every time; checklist for differentiation (hook, light/time, layers/footing, brand/category, two dimensions differ); pull from MLB high-frequency scenes, grade/time, and reference-outfit.md social UGC & trend tags.
Scene library (MLB): reference-scenes.md MLB; client required scene column wins per row.
Styling vocabulary and formulas (required)
When user/client has not specified bottoms/shoes/socks/inners: run reference-outfit.md §1 Decision order → §2–§5 picks; then “Social UGC & trend styling” in the same file for one compatible trend tag—abstract to English (no web required). Rich ≠ more items. Check §8 Pre-prompt checklist.
- Outfit information density (hard rule): state bottom silhouette, shoe family, sock logic, at least one visible layer; don’t use only
stylish outfit.
Trend search (global social & editorial, optional)
After required styling block: optional web search for global social / editorial and trend keywords (Agent cannot browse gated logged-in social apps as the end user). Abstract to phrases—no stealing finals, no hero swap. When/How/What/Forbidden same as root Skill; signage taboos reference-scenes.md MLB.
Step 4: Model and mood (editorial)
Full detail lives in reference-model-presets.md (tiers, sweet-clear addendum, body/face density, limbs, accessories, composition scale). Open that file when drafting internal prompts or fixing QC for frames 1–6.
In SKILL, keep these hard rules:
- Pick one model tier (default female-leaning) or user/custom tier; phrases in English, pick don’t stack—see reference Sections 1–3.
- Priority: user/client model preference > step 1 brand hair/makeup > step 3 emotion; user wins on conflict. If ethnicity/tier unstated: infer + one-sentence confirmation. Male / plus-size / outside table: follow user + brand; still use reference for proportions and QC vocabulary.
- Eyes default to camera; pose variety (stand/sit/lean/walk) and outfit visibility per reference-pose-vocabulary.md and reference-outfit.md (reference-model-presets Section 8 points to the same).
- Every prompt must satisfy body proportion lines and facial detail density in the reference—else Image QC will fail; sweet-clear tier also uses Section 3 addendum and QC item 9.
Step 5: Consistency, light, expectations
-
Photo and color: natural-light led; default bright natural, clear white daylight—clean bright daylight, airy natural light, clear white-balanced daylight; also natural sunlight, bright daylight. Do not confuse with step 1 taboo “cold office elevator top light”—here means outdoor/window clear daylight, not that scene type. soft golden hour only if step 3/emotion asks warm—not every frame yellow. Layered background; reuse reference-scenes.md Scene diagram · §2–§7. Slightly higher saturation, clear tones—avoid muddy gray. Even face light. Frame 5 not chest-up only. Full body: head room, feet mostly in frame; background rich, secondary.
- Delivery look reference (clear street lifestyle): under default bright natural (clear white), prefer even highlight, soft shadow—bright overcast or open shade diffused sun (don’t default whole frame golden hour / strong warm skin unless step 3 says so). Neutral, slightly cool-clean WB; may add
soft even lighting, diffused daylight, neutral white balance, minimal warm color cast. For full body or line emphasis, slightly low angle (slightly low angle) aligns with body rules; if still big-head/short-leg, reduce tight face crop, reinforce balanced head-to-body ratio + elongated legs—avoid wide-angle face + bad crop chibi proportions. Street lifestyle: façade (warm gray plaster/brick) + glass doors + distant trees/street depth; foreground light terrace table corner, cup, chair back; mid/back shallow DoF; subject dominant. Clear garment colors vs warm env, white, green—airy, not gray or mushy.
-
Garment + styling consistency: hero locked; styled pieces one family all 6.
-
Hero anchoring: frame 1 only first; then 2–6 use frame 1 URL + angle refs.
-
Frame 4 (front variation): not back; front; different pose/scale vs frame 1; front product + anchor; if user later wants back, restore back logic.
-
Logo / small type: may warp—note PS path.
-
Background + anatomy: same defaults as step 1; limbs/shoes rules satisfied.
Step 6: Six frames (ratio 3:4 only)
| # | Framing | Notes | Internal focus (don’t show full text to user) |
|---|
| 1 | Front | Upper large; to camera; same SKU as white-bg ref | front view, chest-up or waist-up, large scale, looking at camera; exact garment from reference product image, faithful details, no substitute |
| 2 | Side | Upper dominant, large; eyes toward camera | side profile, upper body, tight framing, eyes toward camera |
| 3 | Full | Large in frame; long legs, normal head-to-body | full body, large subject, looking at camera; elongated legs, balanced head-to-body ratio |
| 4 | Front variation | Not back; alternate pose/scale vs 1 | front view, alternate pose vs shot 1, upper or medium framing, looking at camera, not back view |
| 5 | Three-quarter outfit | Mid-thigh, waist/pockets readable | three-quarter, mid-thigh crop, waist and pockets visible, looking at camera |
| 6 | Fabric macro | Large detail; face in frame → to camera | fabric macro, large detail fill; face → looking at camera |
Execution rhythm (two phases)
- Phase A: frame 1 only, white-bg in
img_urls, ratio="3:4", user confirms URL.
- Phase B: frames 2–6 with frame 1 URL anchor + angle refs;
ratio="3:4".
Don’t generate all 6 at once.
Frame 1 must match white-background product (hard rule)
Common failure: model’s top is not the same SKU as client flat (wrong or “similar” piece). Frame 1 must pass SKU fidelity before 2–6.
- Internal prompt must stress (pick): exact same garment as reference flat—
exact garment from reference product image, faithful reproduction, match reference SKU, same neckline/sleeves/hem/pockets/print placement, no alternate design, no generic similar item. Styled bottoms/shoes still per reference-outfit.md—never replace hero.
- img_urls: user-confirmed front white-background product; if multiple refs, state which is hero internally.
- Before delivery: Image QC · frame 1 vs flat; fail → rerun frame 1 only—do not proceed to 2–6 if frame 1 is wrong garment.
Save finals to disk (after approval)
When user/client finalizes all 6 (or each frame), Agent must download each http(s) result URL to a local folder—not only chat links.
- Suggested path:
outputs/editorial-six/<SKU_or_date_summary>/ (or user absolute path); one subfolder per SKU.
- Filenames (match step 6):
01_front.jpg, 02_side.jpg, 03_full_body.jpg, 04_front_variation.jpg, 05_three_quarter_outfit.jpg, 06_fabric_macro.jpg (or .png, consistent).
- How: Download each result with whatever the environment allows (browser save-as, terminal
curl, Agent/IDE file tools); save under the filenames above. No Python batch script required.
- Optional:
urls.txt one URL per line; hero anchor archive: 00_hero_anchor_url.txt with final frame 1 URL.
Image QC and regeneration (before each delivery)
Before handoff or next frame, view output; if any item fails, regenerate that frame only with targeted strengthening (frame 1: exact garment; sweet-clear tier: expression; Step 4 body + facial density + Avoid per reference-model-presets.md; complete visible limbs, looking at camera, shoes on feet, layering)—don’t rerun all 6.
- Body proportion: oversized head, giraffe neck, short legs, stubby/chibi proportions → rerun; reinforce required + Avoid phrases; if big head + short legs, don’t only tweak face.
- Anatomy and limbs: missing/extra limbs, bad hands, bad joints; bad elbow/knee crops.
- Shoes and hands: extra/floating shoes, hand–shoe fusion (unless hold-shoe requested).
- Face and features: output face shape, hair, eyes/nose/lips clearly off prompt (identity drift, hair vs ban)—rerun; common: moon face, too round-flat, chubby cheeks, wide jaw vs oval/V-line/small face, large eyes, high bridge; unfounded light brown hair; bare lifestyle / blunt features vs sweet-influencer target. Retry: facial density + Avoid (
round face, wide jaw, light hair, etc.).
- Eyes and framing: side still toward lens; frame 4 front + to camera.
- Garment vs hero: drift; unreadable logo (note post). Frame 1 strict: top vs front flat same SKU—neckline, placket, sleeves, pockets, hem length, main color, print placement clearly different (“another piece”) → rerun frame 1; don’t enter 2–6 until aligned.
- Background and grade: empty wall, flat gray; default taboo locations unless job asked.
- Outfit: if bottoms/shoes should show, visible and consistent.
- Sweet-clear (strict, only if tier = Asian female sweet-clear or equivalent written):(a) Expression cold, distant, “over-it,” zero warmth, or pink blush but still cold editorial / mature face; (b) face too large, wide jaw, moon/round, wide square, overly mature bone vs oval, petite face, large eyes, high bridge; (c) hair middle part + big curl / heavy wave; (d) hair unfounded light gold/brown/bleach (no client write)—any clear miss → rerun; reinforce reference-model-presets.md Sections 1 and 3: expression, Asian makeup, small face, young hair, dark brown-black hair (
dark brown-black hair, influencer-style petite face, Asian beauty makeup, youthful hairstyle, no middle-part big curls, etc.)—don’t fight garment/body hard rules.
Retry policy: ~2–3 tries per frame; then explain and ask flat swap / relax rule / accept.
Frame 1: if rerun (including mismatch flat), tell user “hero frame regenerated” + reason; 2–6 anchor final frame 1 URL.
User: may auto-retry; if user approved a frame, don’t auto-rerun it.
Troubleshooting
| Symptom | Likely cause | What to do |
|---|
| User wanted 8 PDP slots (front/back/side/detail/flat lay…) | Wrong skill | Switch to ecommerce-fashion-workflow; do not force six editorial frames. |
| Skill should trigger but does not | description mismatch | User query should mention six frames, editorial/magazine/storyboard, same-SKU multi-angle, or nano-banana; re-read frontmatter triggers. |
| Skill triggers for unrelated tasks | Over-broad wording in chat | Confirm user actually wants six 3:4 editorial shots—not a generic “help me generate some images” request. |
| Frame 1 wrong garment vs flat | Prompt or ref order | Rerun frame 1 only; strengthen SKU-fidelity phrases; verify img_urls uses correct front flat. |
| HTTP / timeout errors | API or network | Check NANO_BANANA_BASE_URL, increase timeout on frame 1, verify network; retry; surface error once with next step. |
| Big head / short legs / moon face (sweet tier) | Model block too thin | Apply Image QC items 1, 4, 9; rerun that frame with density + Avoid lines from reference-model-presets.md—not the whole set. |
| Instructions feel too long to follow | Context load | Use Critical rules + step headers first; load reference-scenes.md / reference-outfit.md for scene/styling; load reference-model-presets.md for Step 4 and QC face/body fixes. |
Self-check list (each delivery)