Krea 2 / Ideogram 4 prompt skill: route first, then write

TL;DR

• The skill covers Krea 2 (RAW, Turbo, hosted) and Ideogram 4.0 (natural language, structured JSON, Magic Prompt, Canvas and Edit), plus Ideogram 4.5 Precise Edit with mask and multi-turn editing.
• Its core claim is that the two models are conceptual relatives — both LLM-encoded, not CLIP — but their syntaxes do not interchange: Krea 2 wants dense natural prose, Ideogram 4 wants declarative hierarchy or JSON.
• Every prompt output follows one profile: FINAL PROMPT (EN), positive constraints, a negative prompt when the workflow has one, suggested settings, and the single variable to change next.
• A failure table maps symptom to cause to the exact reference section, which is how the skill keeps you from re-rolling blindly.

What it is and who it is for

Krea 2 / Ideogram 4 Prompting is an agent skill for prompt engineering on two image models at once. It is built for the person switching between them all day — generating in Krea 2 and touching up or retyping in Ideogram — and who loses quality because the two get the same prompt.

The premise is worth stating plainly: Krea 2 and Ideogram 4.0 are related models, but the sentence that works in one does not work in the other. Krea 2 reads dense natural prose. Ideogram 4 reads a declarative hierarchy, or structured JSON. A skill that treats them as one model produces prompts that half-work in both.

What it does

The README documents these behaviors:

• Routes before writing. A 12-row intent table picks model × surface × goal first — create from zero, strict layout control, style reference, moodboard, region edit, identity edit, inpaint or outpaint, typography, reverse-prompting, failure repair, genre presets — then loads exactly one reference file.
• Guardrails on every output. No quality-keyword confetti (masterpiece, best quality, 8k…), no SD-style weights like (word:1.3) because Krea 2 ignores them and wants restatement instead, rendered text always in double quotes and early in the prompt, one change per edit turn as the reliable default, describe the result rather than the operation, and positive constraints instead of negative prompts where the surface has no negative field.
• Structured JSON for Ideogram 4, with the observed schema, sampler presets and a bundled validator.
• A consistent output profile, so you always get the prompt plus what to change next rather than a wall of text.
• A failure diagnosis table: symptom, likely cause, and the exact reference section, covering melting edits, a style that swallows the subject, late text rendering wrong and whole-image regeneration on 4.0 Edit.

How to use it

Copy the folder into your agent's skills directory:

git clone https://github.com/artificialguybr/krea2-ideogram4-prompt-skill.git
cp -R krea2-ideogram4-prompt-skill ~/.claude/skills/

The README notes it also works with Kimi and Codex skill folders. SKILL.md loads on demand and the eight reference files load lazily, so keep the relative paths intact. Then ask in plain language and let the skill pick the route: a minimal poster with a title in big serif type goes to the typography reference, keeping a face inside a new outfit goes to Krea 2 Identity Edit, extending a photo to a wider ratio goes to Ideogram Canvas Extend, and changing only a clock in a photo is routed to Ideogram 4.5 Precise Edit with High plus mask.

Validate a JSON prompt before sending it:

python scripts/validate_ideogram4_json.py my_prompt.json

Final prompts are written in English because both models respond best to it; routing questions and explanations come in Portuguese.

What is inside

SKILL.md holds the golden rule — route first — plus the guardrails and the output profile. Eight reference files carry the judgment: model-facts.md for per-variant parameters, krea2-architecture.md for T2I, style reference and moodboard, ideogram4-json.md for the structured schema, Magic Prompt and presets, editing.md for Krea Edit and Annotate, Identity Edit, Canvas, 4.5 Precise Edit and the hosted fallbacks, plus typography, reverse-prompting, failure-repair and genre-presets. INSTALL.md covers install and maintenance notes.

Honest limits

The README is unusually direct about what is a reconstruction: the Ideogram schema is a research reconstruction and you should validate it against the official docs before production use, because parameters change between versions. Seed is accepted on the current v2 generate endpoint while the legacy v1 endpoint needs a docs check. negative_prompt does not exist on Ideogram 4.0 or 4.5, so exclusions go as positive constraints. Pixel-preserving edits belong to Ideogram 4.5 Precise Edit, which is hosted; the open weights of Ideogram are 4.0 text-to-image only, so locally you use generic ComfyUI inpainting.

When to use it, when not to

Use it when you move between the two models and want the prompt to land in the right contract the first time, and when a single wrong step (a weight, an SD-style negative, a quality-word list) is what keeps costing you generations.

Do not expect it to fix a model that you have not set up. The skill writes prompts; it does not host anything, and it does not make a locally unsupported feature work — a pixel-preserving edit on open weights is not what 4.0 text-to-image does. And do not transfer a prompt from one surface to the other after the fact: that is the exact mistake the skill exists to prevent.

The skill follows the skill-creator structure and is MIT-licensed.

Sources

1. Krea 2 / Ideogram 4 Prompting — repository — routing table, guardrails, JSON workflow, output profile, failure table and honest limits quoted above.
2. README.md — install commands, structure, license.

Project page: Krea 2 / Ideogram 4 Prompting

For the prompt structure itself, see the Krea 3 prompt guide; for the photoreal side of Krea 2, Krea 2 Turbo photorealism; and for what to do when a surface has no negative field, negative prompts across models.

FAQ

Do Krea 2 and Ideogram 4 use the same prompt format? No. The README states their syntaxes do not interchange: Krea 2 wants dense natural prose, while Ideogram 4 wants a declarative hierarchy or structured JSON.

Why does the skill forbid weights like (word:1.3)? Because Krea 2 ignores SD-style weights, so the skill restates the idea in prose instead of weighting it.

Can I use a negative prompt with Ideogram 4? Not on 4.0 or 4.5 — the README says negative_prompt does not exist there, so exclusions are written as positive constraints.

Does a JSON prompt guarantee the layout? No. It puts the prompt in the contract the model reads, which is the skill's claim; it does not promise that the layout survives every take.