cover-maker

Distinct from flyer-maker:

• No event details (date / location / CTA) — covers have title + creator
• Aspect varies by MEDIUM, not platform (album = 1:1; book = 2:3 portrait; podcast = 1:1; magazine = 2:3 magazine cover; report = 1:√2 A4)
• Different composition conventions (titles dominate; minimal supporting metadata)

Distinct from image-prompt:

• Structured input (medium / title / creator) vs free-form prompt
• Multi-variant batch (default 2-3 takes per medium)
• Auto-picks model based on text needs + style

This skill does NOT:

• Generate physical book bindings / album sleeves / cover spreads (back covers, spines) — single front-cover image only
• Generate ISBN barcodes / catalog numbers — overlay manually in your editor
• Source rights-cleared imagery — provide your own photo via --photo
• Cover scaling for specific marketplace dimensions (Amazon KDP / Spotify Canvas / Apple Music) — generate at the medium's standard aspect, resize/upscale in a DTP tool

ROLE

Read metadata + medium + optional photo + style → pick aspect from medium → pick text-friendly + (if photo) multi-ref-capable model → assemble per-variant prompts with composition zones → batch execute → save PNGs.

PIPELINE (v2.14.0+ — shared visual-prompt chain, same as carousel-builder)

1. Resolve metadata:

   • Required: --title
   • Strongly recommended: --creator (album artist / book author / podcast host / report org)
   • Optional: --subtitle, --photo <path-or-url>, --brand-colors "<list>"

2. Resolve medium — picks aspect + composition convention:

   • --medium album → 3000×3000 square (Spotify / Apple Music album art)
   • --medium book → 1600×2400 (2:3 portrait — Amazon KDP standard)
   • --medium podcast → 3000×3000 square (Apple Podcasts spec)
   • --medium magazine → 1600×2400 (2:3 — print magazine cover convention)
   • --medium report → 1240×1754 (A4 portrait at 150 DPI)
   • --medium deck-cover → 1920×1080 (16:9 — slide deck title slide)
   • --medium linkedin-doc → 1080×1080 (1:1 — LinkedIn document)
   • Custom: --aspect WxH

3. Resolve style — see common/visual-prompt-library/styles/_index.md (shared 13-style library):

   • --style auto (default): the LLM picks from the library based on title + creator + medium + tone.
   • --style <name>: explicit from the 13-style library (BIOTECH / CYBER-NOIR / BRUTALIST / VAPORWAVE / MILITARY / SCIENTIFIC / STREETWEAR / ART-DECO / BLUEPRINT / GRUNGE / GLAMOUR / NATURE / ADVENTURE).
   • --style custom "<desc>": free-text override passed verbatim.

4. Pick model — see references/model-picker.md:

   • Heavy embedded text (covers always have text) → ideogram-3-quality (default) or gpt-image-2.
   • Photo reference + identity → nano-banana-pro.
   • Photo reference + brand palette → flux-2-pro.
   • Photoreal magazine-style cover → nano-banana-pro.

5. Compose ONE LLM call — load common/visual-prompt-library/system-prompt.md (the shared SYSTEM_PROMPT) and buildUserMessage(opts) with:

   Mode: cover
   Number of images to generate (N): <variants, default 2>
   Aspect ratio: <medium aspect>
   Topic / theme: <title + creator context>
   Title: "<title verbatim>"
   Creator: "<creator verbatim>"
   Subtitle (optional): "<subtitle verbatim>"
   Medium: <medium>
   Visual style: <library entry full description OR customStyle text>
   [Optional: brand colors, photo reference flag, character description]
   
   Respond with a JSON object: { "slides": [...] }
   

   Spawn ONE Agent (subagent_type=general-purpose) with system=SYSTEM_PROMPT and user=<built message>. The agent returns JSON {"slides":[{"number":1,"prompt":"..."},...]} — N short (1–3 sentence) cover prompts, title + creator quoted, layout language, no carousel chrome (single-image mode).

   Discipline (all enforced in the SYSTEM_PROMPT):

   • ONE LLM call, not per-variant subagents.
   • Each prompt 1–3 sentences.
   • Title + creator + subtitle in double quotes exactly.
   • No meta-labels (no TITLE: / AUTHOR: literals).
   • Title-dominant composition; creator in a consistent secondary zone.

   Retry on bad output: if malformed JSON or wrong N, re-run once with stricter reminder.

6. Assemble plan.json — items [{index, label, prompt, kwargs:{size, image_url}}]. prompt is LLM-returned text. image_url points to --photo when provided.

7. Estimate cost + confirm — inherits SKILLS_CAROUSEL_BUDGET=1.50.

8. Batch execute — python3 -m common.runners.cli.cover --plan-file <plan.json> --yes (or via scripts/run.py).

9. Output:

   ./generated/cover/<slug>/
     <medium>-v1.png
     <medium>-v2.png
     <medium>-v3.png    (if --variants 3)
     manifest.json
     style-used.md
     prompts.md
   

MODES

Required

• cover-maker --title "<text>" --medium album|book|podcast|magazine|report|deck-cover|linkedin-doc

Recommended

• --creator "<name>" — artist / author / host / org

Optional content

• --subtitle "<text>" — secondary line
• --photo <path-or-url> — reference image
• --lang en|ru — language hint (default: auto-detect from title)

Visual

• --style auto|<library-id> — visual style
• --style-mod "<override>" — append a tweak
• --variants N — variants (default 2)
• --aspect WxH — custom aspect (overrides medium default)
• --model auto|<slug> — image provider

Two-pass typography (v2.11.0 fallback — opt-in only)

The default v2.14.0 chain is LLM-prompt-then-image (text rendered by the image model, baked into the picture — same chain as carousel-builder). For book covers where text must be pixel-perfect (publisher imprint precision, multilingual layouts the model can't render), opt in with:

• --typeset overlay — runs the legacy two-pass: AI generates a TEXT-FREE background (per the imprint's prompt fragment) + Pillow typography composer overlays title + creator with bundled OFL fonts at the imprint's proper layout fractions.
• --imprint nyrb-classics|penguin-marber-grid|mit-essential-knowledge|picador-modern|faber-modernist — design-system preset for the typography composer (only with --typeset overlay).
• --genre literary-fiction|thriller|non-fiction|academic|memoir|poetry|... — auto-picks imprint (only with --typeset overlay). Mapping in common/runners/cover_imprints.py:GENRE_DEFAULT_IMPRINT.
• Default for all mediums: --typeset ai (single-pass, LLM writes the prompt, image model renders title + creator inside the image).

Execution

• --execute — actually generate
• --output <dir> — custom output
• --parallelism N — concurrent calls (default 2)
• --yes — skip cost confirmation
• --resume — retry failed
• --prompts-only — dry run

REFERENCES (load on demand)

File	When to load
references/cover-types.md	Step 1-2 — per-medium conventions, what fields are needed, typography expectations
references/aspect-presets.md	Step 2 — exact pixel dimensions per medium + platform target
references/composition-zones.md	Step 5 — per-medium composition templates (album / book / podcast / magazine / report / deck)
references/imprints.md	When --imprint is set — full per-imprint design system specs (layout fractions, typography family, palette, prompt fragment)
references/model-picker.md	Step 4 — model auto-pick, when to override
references/troubleshoot.md	When text renders wrong, layout fails, photo doesn't integrate

EXAMPLES

See examples/before-after.md — 3 calibration runs: album cover with reference artwork, business book cover with author photo, podcast cover with bold typographic style.

CONSTRAINTS

• Title is the dominant element. Cover composition prioritizes title legibility. Keep titles ≤6 words for best results.

• Creator name is recommended but optional — for an album with featured artists, list one main creator and put the rest in --subtitle.

• One medium per run. Don't mix album + book in the same call. Run twice if you need both.

• One style for variants. All variants share the same anchor; they differ in stochastic interpretation, not in style.

• Embedded text quality is paramount. Default to ideogram-3-quality for any text-heavy cover. The model picker enforces this.

• Photo-reference covers: pair --photo with nano-banana-pro (identity) or flux-2-pro (brand palette transfer).

• Cost confirm ONCE per batch. Sum across variants.

• No print-bleed marks / crop guides. Output is the cover image only. For physical print: import to InDesign / Affinity Publisher for bleed + crop marks.

• Backside / spine: out of scope. Single front-cover image only.

• Never print API keys. Mask in errors.

INVOCATION HINTS

When the user says any of:

• "album cover for X", "book cover for Y", "podcast cover", "report cover"
• "magazine cover", "deck cover slide", "LinkedIn doc cover"
• "обложка для альбома / книги / подкаста / отчёта"
• "сделай обложку для X"

If the medium isn't clear from context, ask once. Default if unspecified: album (most common request).

If the user mentions a platform (Spotify → 3000×3000 album; Amazon KDP → 2:3 portrait book; Apple Podcasts → 3000×3000 podcast), bias --medium accordingly.

Defaults: --medium album --variants 2 --style auto --model auto. Without --execute, returns prompts.

This skill is distinct from:

• flyer-maker — events with date/location, NOT covers with title/creator
• image-prompt — free-form image generation; this is structured covers
• avatar-maker — single subject portrait, NOT title-driven cover
• thumbnail-maker — 16:9 with bold title for content marketing, NOT artist/author covers