Prompt engineering image

prompt-engineering-image

Translate an image brief into a provider-specific prompt string. Each model has distinct prompt grammar — this skill applies the right structure per adapter.

When to use

• Writing or refining a prompt for any pack-ai-image provider.
• After image-provider-routing has selected the target provider.
• When a prompt is underperforming and needs provider-specific tuning.

Per-provider prompt grammar

Ideogram (text-in-image: logos, banners, typographic art)

• Lead with the text literal in quotes: "BAKERY NAME" in bold serif on a cream background.
• Follow with visual context: background, color palette, style (flat, vintage, art-deco).
• Avoid long scene descriptions — Ideogram renders text best with a focused layout brief.
• Key params: model: V_2, aspect_ratio: ASPECT_1_1 (square logo) or ASPECT_16_9 (banner), magic_prompt_option: AUTO (let Ideogram enrich).

Flux (photoreal: product shots, portraits, scenes)

• Descriptive noun phrase first: subject → lighting → environment → camera.
• Example: "Close-up product shot of a ceramic coffee mug, soft studio lighting, clean white background, 50 mm lens bokeh, ultra sharp".
• Style descriptors: cinematic, hyperrealistic, 8K, golden hour, DSLR.
• Negative prompts accepted: --no cartoon, illustration, text.
• Routes through fal/Replicate — model slug: fal-ai/flux-pro or black-forest-labs/flux-pro.

Recraft (vector / SVG logos and icons)

• Style param is mandatory: style: vector_illustration (SVG output), style: icon (for simplified marks), style: realistic_image (raster fallback).
• Keep prompt minimal — recraft interprets shape semantics: "minimalist leaf icon, single color".
• Avoid photographic language (lighting, bokeh, grain) — it has no effect on vector output.
• Key params: model: recraftv3, response_format: url.

Gemini-image / GPT Image 2 (general art, edits, multimodal)

• Natural language works well — no special syntax required.
• Be explicit about style: "watercolor illustration", "flat design", "oil painting".
• For GPT Image 2 image editing: include the edit instruction after describing the target: "Remove the background and replace with a solid pastel blue".
• Gemini: submit via generateContent (Nano Banana family) or imagen-4.0:predict (Imagen 4).

Procedure

1. Receive the brief — extract: subject, style, output format (raster/vector/banner), target provider (from image-provider-routing or explicitly stated).
2. Structure the prompt blocks — subject · style · composition · technical params.
3. Apply provider grammar from the section above for the target model.
4. Tune for the job shape — text-literal first for Ideogram; noun-phrase first for Flux; minimal + style: param for Recraft; natural language for Gemini/GPT.
5. Inspect the adapter header — open src/scripts/ai-image/adapters/<provider>.sh and confirm the param enums (aspect/style/model) the prompt relies on still match.
6. Emit the prompt in the Output format below.

Output format

1. Target provider — name + adapter file reference.
2. Prompt string — the exact string to pass to the adapter, ready to copy.
3. Key params — any model-specific fields (aspect ratio, style, negative prompts).
4. Variant (optional) — one alternative phrasing when the brief is ambiguous.

Gotcha

• Per-provider param enums drift — aspect_ratio, style, and model enum values are ASSUMED from the adapter header comments. Verify against live API docs before promotion; never hardcode these in production without a smoke trace.
• Adapters are scaffold-tier — prompts authored here are not live-validated. Actual rendering requires adapter promotion to stable per provider-lifecycle-discipline.
• Recraft: photographic descriptors (bokeh, lighting, grain) silently have no effect on vector output — strip them to avoid prompt budget waste.

Good example: Ideogram brief for a bakery logo — lead with the text literal: '"Le Four" in warm serif, vintage French patisserie style, cream and terracotta'.

Bad example: Sending a photorealism-heavy Flux prompt to Recraft — the style descriptors will be ignored and the vector output will be wrong.

Do NOT

• Do NOT send photographic descriptors (bokeh, lighting, grain) to Recraft — they have no effect on vector output and waste the prompt budget.
• Do NOT hardcode ASSUMED param enums into a live run without a smoke trace — verify against the adapter header / provider docs first.
• Do NOT embed a real person's likeness, a trademarked brand mark, or a named living artist's style in a prompt without the rights check (image-likeness-and-rights).
• Do NOT write a prompt before the provider is chosen — route via image-provider-routing first.

See also

• image-provider-routing — select the provider before writing the prompt.
• provider-lifecycle-discipline — lifecycle tier gates live runs.
• src/scripts/ai-image/adapters/ — adapter header comments for param enums.
• image-likeness-and-rights — rights check before generating real-person likenesses or brand marks.