Podcast cover is 3000×3000 square JPEGbut at the same time 55×55 thumbnailthe motif and show name must also be readable. To satisfy both constraints at once:
- AI는 large silhouette·simple motifdraw only (no text),
- high-resolution upscaleto secure print·Retina quality,
- show name typesetting uses system fontcomposite with.
Trigger
imagine-podcast "딥러닝을 씹어먹는 밤" --style vintage-radio --primary "#F59E0B"imagine-podcast episode --master ./images/podcast/cover.jpg --ep 12 --title "트랜스포머의 기원"- "make a podcast cover", "Apple Podcasts cover", "audiobook cover", "playlist cover", "episode cover variant"
- "imagine-podcast"
Pipeline (master generation)
- Prompt composition:
config.style_presetsfrom--styleselect key → English phrase.--primary/--secondaryHEX는 name phraseconvert to and inject. Original HEX is_brand.jsonstored only in.positive_suffix:bold silhouette, recognizable at small size, centered motif, solid or simple gradient background.
- Negative enforced:
text, letters, logo text, watermark, cluttered small details, tiny icons, thin outlines, photo of faces, photorealistic skin. Never inject show name·episode title into the prompt. - AI background generation: with Codex image model
ai_generate_size: 1024x10241 PNG. This file iscover_master_bg.png로 text-free original preserved — used when subsequent episode variants recomposite. - Upscale (
scripts/lib/upscale.js):- Default: sharp Lanczos3to 3000×3000 enlarge.
--esrganflag attempts Real-ESRGAN CLI use. On failure fall back to Lanczos3 and warn the user ("upscale: Real-ESRGAN failed (<reason>); falling back to Lanczos3.").- Upscale result is internally processed as PNG for now.
- Show name compositing (
scripts/lib/compose-text.js): default fontPretendard Bold,safe_zone_ratio: 0.80, auto brightness measurement auto-selects white/black. Safe zone leaves top/bottom 20% empty to maintain readability required by Apple Podcasts. - JPEG encoding (RGB, 4:4:4): sharp
.jpeg({ quality: 92, chromaSubsampling: '4:4:4', mozjpeg: true }).flatten({ background: '#ffffff' }). CMYK·grayscale saving prohibited. - Apple Podcasts spec auto-validation (
apple_podcasts_spec.verify_default: true): size 3000×3000 match + JPEG + RGB + byte count 500KB~500MB range. If off, adjust re-encoding quality or warn user. - 55×55 thumbnail preview: save 55×55 JPEG with sharp Lanczos3 (
<master>_thumb55.jpg). entropy/contrast measurement:- entropy <
entropy_warn_below: 3.0→ "motif too monotonous, hard to identify in thumbnail" warning. - contrast score <
contrast_warn_below: 0.25→ "thumbnail contrast insufficient" warning. - Warnings are
_brand.json.warnings[]recorded in and exposed to user, no auto-regeneration.
- entropy <
Pipeline (episode variant)
imagine-podcast episode --master <cover.jpg> --ep <n> --title "<text>"
cover_master_bg.png(text-free original) reused as-is. No AI re-call — quota·consistency protection (episode_variant.never_regenerate_bg: true).compose-text.jstypeset show name + episode number + title at top/bottom.- Save:
./images/podcast/<show-slug>/ep/ep<nn>_<slug>.jpg(filename_pattern: "ep<nn>_<slug>.jpg"). episode_list.jsonlAdd one line to{ ep, title, slug, ts, master_hash }— history only; storing image bytes or raw prompts is prohibited.
Output convention
./images/podcast/<show-slug>/
├── cover_master.jpg ← 3000×3000 JPEG, 쇼 이름 합성 완료 (Apple 규격)
├── cover_master_bg.png ← 글자 없는 AI 원본 (에피소드 재합성용)
├── cover_master_thumb55.jpg ← 55×55 프리뷰 (가독성 자동 평가용)
├── ep/
│ ├── ep12_트랜스포머의-기원.jpg
│ ├── ep13_어텐션-메커니즘.jpg
│ └── ...
├── episode_list.jsonl ← 에피소드 이력 (이미지 데이터 없음)
└── _brand.json ← style/palette/font/warnings/apple_spec_verification
<show-slug>: user--slugor show name slugified.--out-dirIf unspecified,./images/podcast/<show-slug>/fixed. Root./images/writing directly under it is prohibited.
Prompt conventions
- Injecting show name or episode title into the prompt is prohibited. All text enters only via
compose-text.jspost-processing. - Direct rendering of Apple Podcasts/Spotify logos is prohibited. Do not place platform brand marks on the cover (include
logo textin the negative + composite your own brand logo manually only when the user wants it as a separate layer). - Keep motifs large and simple:
bold silhouette, recognizable at small size가positive_suffixfixed at . Blocks complex decoration. - Color: Directly writing HEX is prohibited; inject name phrases; the original HEX goes only in the manifest.
Upscale mode
- Default (
--esrgannone):upscale.jsresizes to 3000×3000 with sharp Lanczos3. Deterministic, zero external dependencies. - opt-in (
--esrgan): Real-ESRGAN CLI(예:realesrgan-ncnn-vulkan) call attempt. If successful, use the SR result; if it fails, immediately fall back to Lanczos3 + reason warning. Do not fake success. - The upscale module treats the "no external binary" state not as an error but as a graceful fallback.
Automatic Apple Podcasts spec validation
config.apple_podcasts_spec:
| Item | 값 |
|---|---|
| Size | 3000×3000 |
| Format | JPEG |
| Color space | RGB (CMYK/Grayscale prohibited) |
| bytes | 500KB ~ 500MB |
On validation failure:
- Below the byte lower bound →
jpeg_qualityraise one step and re-encode. - Above the byte upper bound →
jpeg_qualitylower and re-encode. - Color space mismatch → force conversion with sharp
.toColorspace('srgb')and re-save. - Once spec compliance is confirmed, record
_brand.json.apple_spec_verification: "pass".
55×55 thumbnail readability check
- After automatic resize, compute Shannon entropy와 and mean contrast.
- If below the threshold, record a warning in
_brand.json.warnings[]and advise the user to "simplify the motif". - Do not automatically regenerate — the simplicity of a podcast cover is also a matter of user intent.
Failure mode handling
| Symptom | Action |
|---|---|
| Motif smeared at 55×55 | Entropy/contrast warning + --style recommend a simpler preset (modern-minimal / paper-grain). No automatic regeneration. |
| Real-ESRGAN CLI missing/failed | Lanczos3 fallback + cause (esrgan_spawn_error / esrgan_exit_N / esrgan_missing_output) warning. |
| 1024 → 3000 Lanczos smearing | Keep the motif large and simple (enforce positive suffix). If high quality is needed, --esrgan opt-in. |
| Saved as CMYK·Grayscale JPEG | sharp toColorspace('srgb')force to and re-save. |
| Below/above capacity spec | jpeg quality auto ±5 adjustment loop (max 3 times). If it still fails, expose to the user. |
| compose-text.js absent | explicit error (npm i sharp). Preserve the textless master PNG. |
| master PNG absent in episode variant | Reject. imagine-podcast <show> Guide to run it first. Do not cover it up by re-calling AI. |
Taboo
- Do not inject show name or episode title into the AI prompt.
- Do not re-call AI in episode variant (
never_regenerate_bg: true). Master PNG reuse required. - Do not save formats other than JPEG (Apple standard). Even if PNG is used in the internal process, the final output is JPEG RGB.
- Do not save in CMYK or Grayscale.
- Do not report upscale failure as success — Real-ESRGAN failure must disclose the reason.
- Do not hide the 55×55 warning. Always expose to the user.
./images/podcast/No external writes (--out-dirunless specified).- Do not AI-render showname → logo.
masterpiece/8k UHD/fulfill all requestsDo not use booster or bypass phrases such as .
relationship with the imagine family of skills
compose-text.js: dedicated to compositing show name / episode title — shares the same module withimagine-thumb·imagine-og·imagine-logo.upscale.js: the shared upscale module introduced by this skill. Reusable in other skills (print-grade posters, etc.).- Style Guardian:
per_showscope (cross_show_leak: false). Palettes of multiple shows do not mix.show_slug/palette/motif_family/font를_brand.jsonpreserved in . imagine-thumbdistinction from (YouTube thumbnail): podcast is 1:1 3000, YouTube is 16:9 1280×720. This skill's episode variant mode is the core.imagine-ogdistinction from (social card): OG is a story-unit image; this skill is show identity + episode history management.imagine-icondistinction from: app icons are a platform-specific matrix; this skill is a single square + 55px legibility constraint.