Asset Enhancer

asset-enhancer skill

Comprehensive behavior spec for producing production-grade software-development assets. Drives every asset_* MCP tool.

Cardinal rules

1. Producing production-grade software assets is a routing + post-processing problem, not a prompt-engineering problem.
2. The user may not have an image-model API key. The plugin must work anyway.
3. There are real zero-key and free-token paths — always surface them first before asking the user to pay for anything. The free_api.routes block in asset_capabilities() enumerates the ranked free programmatic routes: Cloudflare Workers AI (Flux-1-Schnell + SDXL, 10k neurons/day), NVIDIA NIM (Flux / SDXL / SANA, 1k requests/month with NVIDIA_API_KEY, no card), HF Inference (free HF_TOKEN), Stable Horde (anonymous queue), Pollinations (zero-signup HTTP, last resort), and local ComfyUI. Google's Gemini / Imagen image API is paid for programmatic image output — unbilled GEMINI_API_KEY → HTTP 429, limit: 0. The AI Studio web UI at https://aistudio.google.com is still free for interactive generation — treat it as a paste-only flow (external_prompt_only → asset_ingest_external).

See rules/asset-enhancer-activate.md for the condensed always-on version. This file is the long-form spec.

Three execution modes

Before generating anything, decide (with the user) which mode to use. Call asset_capabilities() to learn what's available in the current environment.

1. inline_svg — zero key, Claude writes the SVG, server persists

Two-step round trip:

1. The server returns an svg_brief (viewBox, palette, path budget, style hints, skeleton) and instructions_to_host_llm. You emit the <svg>…</svg> inline in your reply as a ```svg code block so the user can see it.
2. Immediately after emitting, call asset_save_inline_svg({ svg, asset_type }). The server validates the SVG against the brief (viewBox, <path> count, palette), writes master.svg to disk, and — for favicon / app_icon — runs the full platform export (favicon.ico + apple-touch-icon + PWA 192/512/512-maskable; or AppIconSet + Android adaptive + PWA). Returns an AssetBundle with variants[].path. Show those paths to the user.

Skip step 2 and the user gets a code block with no file. Do not skip step 2.

Good for: logo, favicon, icon_pack, sticker, transparent_mark, simple app_icon masters. Flat, geometric, ≤40 paths.

Not good for: illustrations, hero art, photoreal, OG cards with web-font typography.

2. external_prompt_only — zero key, user pastes into a web UI

The server returns the dialect-correct prompt + a list of paste targets with URLs in free-first order: AI Studio web UI (https://aistudio.google.com — free interactive Gemini / Imagen), Pollinations (zero-signup HTTP), HF Inference (free HF_TOKEN), Stable Horde (anonymous queue), then paid UIs (Ideogram web, Recraft web, Midjourney, fal.ai Flux, BFL Playground, OpenAI Platform Playground, etc.). The user generates elsewhere, saves the image locally, and calls asset_ingest_external({ image_path, asset_type }) to run the matte / vectorize / validate pipeline.

Note: GEMINI_API_KEY does NOT give free image-gen — Google removed that tier in December 2025. The free Gemini/Imagen route is the AI Studio web UI, not the API.

Good for: any asset type. Best for illustration, hero, text-heavy logos, when the user has a Midjourney / Ideogram subscription but no API key.

3. api — requires a provider key

Server calls the routed provider directly, mattes, vectorizes, exports, validates. Writes a content-addressed AssetBundle with variants[].path for every artifact.

Requires at least one of: OPENAI_API_KEY, IDEOGRAM_API_KEY, RECRAFT_API_KEY, BFL_API_KEY, GEMINI_API_KEY (or GOOGLE_API_KEY). The router falls back among configured providers automatically.

Recommended flow

user brief
  ↓  asset_capabilities()                   → what's available RIGHT NOW
  ↓  asset_enhance_prompt({ brief })        → AssetSpec + modes_available
  ↓  ask the user which mode they want
  ↓  asset_generate_<type>({ brief, mode }) → InlineSvgPlan | ExternalPromptPlan | AssetBundle
  ↓  if inline_svg:
       1. emit <svg> in reply (user sees the shape)
       2. call asset_save_inline_svg({ svg, asset_type })
       3. show the returned variants[].path list (user sees files on disk)
  ↓  if external: show prompt + paste_targets; wait for asset_ingest_external
  ↓  if api: show variants[].path and validation warnings

Asset taxonomy — closed enum

Pipeline (api mode)

  brief
    ↓  classify()                           → asset_type ∈ enum
    ↓  parse_brand_bundle()?                → BrandBundle | null
    ↓  compute_params(asset_type, brand)    → dimensions, transparency, safe_zone
    ↓  route(asset_type, ...)               → primary_model, fallback_model, postprocess[]
    ↓  rewrite(brief, model, brand)         → dialect-appropriate final prompt
  ── call provider with seed pinned ──
    ↓  validate_0(image)                    → reject checkerboard / bad alpha / moderation
    ↓  matte?        (birefnet | rmbg via PROMPT_TO_BUNDLE_RMBG_URL | difference fallback | native)
    ↓  vectorize?    (recraft | vtracer-on-PATH | potrace-on-PATH | llm-svg | posterize fallback)
    ↓  upscale?      (dat2 | supir | img2img | never)
    ↓  export        (sharp platform fan-out + @resvg/resvg-js + satori + png-to-ico + svgo)
    ↓  validate_final (tile-luma alpha check + bbox + palette ΔE2000 + OCR Levenshtein + WCAG contrast)
    ↓  content-address cache (prompt_hash, model, seed, params_hash)
    →  AssetBundle

Model routing (data-driven)

The routing table lives in data/routing-table.json. Summary:

Dialect rules per provider

Prompts are never forwarded verbatim. Rewrite per target family. See the per-angle research under docs/research/ for derivations; implementation is in packages/mcp-server/src/rewriter.ts.

gpt-image-2 / gpt-image-1.5 / gpt-image-1 (OpenAI)

• Prose sentences. Subject → Context → Style → Constraints.
• Transparency: gpt-image-1.5 and gpt-image-1 accept background: "transparent". gpt-image-2 does NOT — the param 400s (regression). Route transparent jobs to gpt-image-1.5.
• For text: "Acme" in double quotes. Ceilings: gpt-image-2 80 chars / paragraph (99% per third-party); gpt-image-1.5 dense text (~60 chars, LM Arena #1); gpt-image-1 ~50 chars.
• Never rely on negative_prompt field — silently ignored across all three.
• gpt-image-2 has an active noise-artifact bug on structured surfaces (clouds, skin, panels). Pricing is not yet on OpenAI's pricing page.

Imagen 3 / 4 (Google) & Gemini 2.5 / 3.1 / 3 Pro Flash Image ("Nano Banana")

• Narrative prose. Imagen has a default rewriter for prompts <30 words, so write ≥30 words of concrete description to suppress it.
• Do not ask for transparency in the prompt — RGB-only VAE renders a checkerboard. Ask for "solid pure white background" and matte externally.
• Per-model text ceilings:
  • Imagen 4 family: ≤25 chars / ~3-4 short words per Google's own guidance. Photoreal leaders, not text.
  • gemini-2.5-flash-image (original Nano Banana 1): ~80% accuracy, degrades past 1-3 words.
  • gemini-3.1-flash-image-preview (Nano Banana 2): ~90% accuracy, ranked #1 on Artificial Analysis Image Arena at launch — strong-text, not weak.
  • gemini-3-pro-image-preview (Nano Banana Pro): paragraph-length reliable, ~94-96% accuracy.
• negative_prompt: Imagen 4 accepts negativePrompt on the Vertex AI endpoint; the Gemini API surface ignores it. All Nano Banana variants ignore it.
• Gemini supports multimodal edit (in-context image refs, up to 14 on 3.1); Imagen does not.

Stable Diffusion 1.5 / SDXL

• Tag-soup: comma-separated descriptors. "masterpiece, 8k, studio lighting".
• 77-token CLIP limit — front-load important tags.
• negative_prompt is a real CFG sampler feature; use it.
• For transparency: LayerDiffuse adapter, or matte after.

Flux 1 / Flux 2 / Flux Pro / Kontext

• Prose narrative. T5 + (Mistral-3 24B VLM on Flux 2) — long-form sentences work; Flux 2 also accepts JSON-structured prompts and @-image refs.
• Never send negative_prompt. Per BFL's Flux 2 prompting guide: "FLUX.2 does not support negative prompts." On 1.x the fal schema rejects it; on 2.x fal silently no-ops. Use positive anchors instead.
• Text ceilings: Flux 1.1 Pro / Pro: 1–3 words. Flux Schnell / Klein: 1–2 words. Flux 2 (pro/dev): ~5–10 words / one tagline reliable. Flux 2 Flex: BFL claims strongest text in the Flux 2 line.
• Flux 2 accepts up to 10 brand reference images (was 8 on Flux 1).
• Flux 2 family has 5 SKUs: pro / flex (steps+guidance control) / max / dev (32B open weights, non-commercial) / klein (Apache 2.0 distilled).

Midjourney v6 / v7

• Prose + -- flags. --sref, --cref, --mref for consistency.
• No official API — only reachable via external_prompt_only mode.
• --ar 1:1, --style raw, --q 2.

Ideogram 2 / 3 / 3 Turbo

• Prose + quoted text strings. Best-in-class typography.
• For RGBA transparency: use the dedicated /ideogram-v3/generate-transparent endpoint; set rendering_speed: "TURBO" for the Turbo tier (there is no style: "transparent" API parameter).
• "Magic Prompt" expands literal prompts — toggle OFF for logo/icon/brand work.

Recraft V2 / V3 / V4

• Prose. style_id = brand lock. controls.colors = hex palette enforcement.
• Only production model with native SVG output.
• Path count correlates with quality; reject >200 paths for a clean mark.

Brand bundle

If a BrandBundle is provided, every generation injects it. Shape:

palette: [#hex, #hex, ...]     # DTCG color tokens
style_refs: [path.png, ...]    # IP-Adapter / --sref / Recraft style_id
lora: path.safetensors?        # SDXL/Flux subject/style LoRA
sref_code: "--sref 123456"?    # Midjourney
style_id: "uuid"?              # Recraft
do_not: ["drop shadows", ...]  # Rewritten as positive anchors
logo_mark: path.svg?           # Canonical mark for composition
typography: { primary, secondary }  # Fonts for composited type

Use asset_brand_bundle_parse to canonicalize a brand.md / brand.json / DTCG tokens.json / AdCP spec into this shape.

Validation

Three tiers. Tier-0 always. Tier-1 on first asset in a set. Tier-2 on any failure or on user request.

Tier 0 — deterministic: dimensions, alpha, checkerboard FFT, safe-zone bbox, file-size budget, DCT entropy.

Tier 1 — metric: VQAScore, palette ΔE2000, OCR Levenshtein, WCAG AA at 16×16.

Tier 2 — VLM-as-judge: Claude Sonnet / GPT vision against asset-type rubrics (not wired by default; set PROMPT_ENHANCER_VLM_URL to enable).

Regenerate vs. repair

Caching

Key: (model, version, seed, prompt_hash, params_hash). Storage: content-addressed path assets/<hash[0:2]>/<hash>/<variant>.<ext>. The MCP server is synchronous — prompt_hash is emitted on every AssetBundle so a future hosted tier (BullMQ / SQS / Cloudflare Queues) can use it as a deduplication key.

Output contracts

InlineSvgPlan (mode: "inline_svg")

{
  "mode": "inline_svg",
  "asset_type": "logo",
  "brief": "…",
  "svg_brief": { "viewBox": "…", "palette": {...}, "path_budget": 40, "require": [...], "do_not": [...], "skeleton": "…" },
  "instructions_to_host_llm": "…"
}

Read svg_brief. Emit <svg>…</svg> as a ```svg code block in your reply **and then call asset_save_inline_svg({ svg, asset_type })** to persist the file. That tool returns an AssetBundle with variants[].path — show those paths to the user so they know the files exist on disk.

ExternalPromptPlan (mode: "external_prompt_only")

{
  "mode": "external_prompt_only",
  "asset_type": "logo",
  "enhanced_prompt": "…",
  "target_model": "ideogram-3-turbo",
  "paste_targets": [{ "name": "Ideogram web", "url": "https://ideogram.ai", "notes": "…" }],
  "ingest_hint": {
    "tool": "asset_ingest_external",
    "args": { "image_path": "<path>", "asset_type": "logo" }
  }
}

Show the prompt + the top paste target(s). After the user saves the result, call asset_ingest_external.

AssetBundle (mode: "api")

{
  "mode": "api",
  "asset_type": "logo",
  "variants": [
    { "path": "…/master.png", "format": "png", "width": 1024, "height": 1024, "rgba": true },
    { "path": "…/mark.svg", "format": "svg", "paths": 18 }
  ],
  "provenance": { "model": "recraft-v3", "seed": 1234, "prompt_hash": "…", "params_hash": "…" },
  "validations": { "tier0": { "...all pass": true } },
  "warnings": []
}

Research

Every decision in this skill is traceable to a research angle. See docs/RESEARCH_MAP.md for the full file-level mapping. The load-bearing angles most relevant to day-to-day behavior:

• docs/research/04-gemini-imagen-prompting/4c-transparent-background-checker-problem.md — why Imagen/Gemini never gets transparency
• docs/research/07-midjourney-ideogram-recraft/7b-ideogram-text-rendering-for-logos.md — text ≤3 words rule
• docs/research/08-logo-generation/8e-svg-vector-logo-pipeline.md — three SVG paths (Recraft / LLM-author / raster-vectorize)
• docs/research/09-app-icon-generation/9a-ios-app-icon-hig-specs.md — 824² safe zone
• docs/research/19-agentic-mcp-skills-architectures/ — why 13 small tools and three modes