Prompting Guides

These guides teach you how to write prompts for each model OpenCauldron supports. Every model has its own quirks — Flux rewards compositional precision, Imagen rewards photographic vocabulary, Ideogram rewards explicit text instructions, Recraft rewards style tags. Using a Flux-style prompt on Recraft (or vice versa) leaves quality on the table.

Read this page first for fundamentals that apply everywhere. Then jump to the per-model guide for the model you are using.

How to use these guides

• Copy any example prompt. Hover the code block, click the copy icon, paste it into OpenCauldron’s prompt field.
• Read the “Why this works” notes. The prompts themselves are less important than the patterns they demonstrate.
• Skim the “Common mistakes” section. Most prompt failures come from a small set of repeated errors.

Anatomy of a prompt

Every strong prompt names five things. The order matters less than making sure each is present.

Lever	What it answers	Example phrase
Subject	What is in the frame?	“a weathered fisherman mending a net”
Action / state	What is the subject doing?	“hands working in the foreground”
Setting	Where?	“on a wooden dock at dawn”
Style anchors	What does it look like?	“35mm film, muted palette, shallow depth of field”
Quality / framing	How is it shot?	“medium close-up, eye-level, soft natural light”

A prompt that names all five rarely produces something unusable. A prompt that names only the subject (“a fisherman”) produces wildly different results every time.

Style anchors that work across most models

Some descriptors carry useful signal almost everywhere:

• Photographic vocabulary: 35mm film, medium format, shallow depth of field, golden hour, softbox lighting, natural light, tilt-shift, prime lens. Models trained on photo metadata key off these.
• Art history references: Dutch Golden Age painting, ukiyo-e woodblock, art nouveau, Bauhaus poster. Strong attractors when you want a specific look.
• Material and texture: matte, glossy, brushed metal, worn leather, paper grain. Cheap and effective.
• Mood through palette: muted earth tones, high-contrast monochrome, pastel washed-out. Works better than abstract mood words like “dreamy” or “ethereal.”

Things that almost never help

• Generic quality boosters. “Masterpiece, 8K, ultra-detailed, trending on artstation” — these were useful on older Stable Diffusion checkpoints and are mostly noise on modern models. Some still respond to “high quality” but the signal is weak.
• Negation by hoping. “Without weird hands” rarely works. Models that support negative prompts have a separate field for that; models that don’t will often include the thing you negated.
• Stacking adjectives. “Beautiful, gorgeous, stunning, breathtaking” doesn’t compound — it dilutes. Pick the most specific descriptor and stop.
• Camera spec dumping unrelated to the scene. Listing five lenses and three cameras in one prompt is just confusing the model.

Negative prompts

Some models support a separate negative-prompt field, others do not.

Supports negative prompts	Does not
Imagen, Ideogram, Recraft, Kling, Veo	Flux, OpenAI (gpt-image-*), Grok

For models without negative prompts, you cannot subtract — only describe what you want clearly enough that the unwanted thing has no room to appear.

Aspect ratio matters more than people think

The aspect ratio you pick changes the content of the image, not just its frame. A “portrait of a fisherman” at 1:1 will be a head-and-shoulders shot. The same prompt at 16:9 will likely include the boat, the dock, and most of the fisherman’s body. If you want a tight close-up at 16:9, you need to say so explicitly: extreme close-up, only the face filling the frame.

Iterating on a prompt

When a result is almost right:

1. Identify the one thing that is wrong (lighting too flat, wrong style, subject pose). Do not rewrite the whole prompt.
2. Add the most specific phrase that fixes it. “Side-lit by a single window” beats “better lighting.”
3. Lock the seed if the model supports it, change one phrase, regenerate. This isolates the effect of your change.
4. If three iterations do not converge, the model may be wrong for the job. Try a different one.

Per-model guides

Each guide follows the same structure: TL;DR → how the model reads a prompt → 8–12 worked examples → common mistakes → parameter cheatsheet.

Image models

• Flux — Black Forest Labs, photorealism and editing
• Imagen — Google, photographic quality
• Ideogram — text rendering
• Recraft — design and illustration
• OpenAI (gpt-image) — instruction-following
• Grok — fast and creative

Video models

• Veo — Google, native audio
• Runway — cinematic motion
• Kling — motion quality
• Hailuo — cost-effective with audio
• Luma — camera controls

Per-model guides are added incrementally. If a guide is missing, the Working with Models reference covers the model’s capabilities and parameters in the meantime.