Prompt craft

A prompt is a list of decisions, not a paragraph of atmosphere

The vocabulary that actually moves a generation, organised by the decision it controls. Written for someone who already knows what they want the shot to look like and needs the words a model responds to.

Three kinds of prompt

Image

Subject, framing, optics and the light that decides whether it reads as a photograph.

26 aspects

Video

Duration, shot size, camera movement and the audio a clip describes for itself.

18 aspects

Audio

Tempo, instrumentation and the structure a generated cue needs to be cuttable.

11 aspects

How to use this

Work down the aspects in order. The order is the recommended order in the prompt too.
One clause per aspect. A prompt is a list of decisions, not a paragraph of atmosphere.
Prefer a number over an adjective wherever a number exists: 116 BPM over upbeat, 85mm over flattering, 15 seconds over short.
Put the negatives last, in one sentence beginning with No.
Reuse the identical phrasing across a batch. Paraphrasing is what makes a set of images look unrelated.

The vocabulary behind the vocabulary

Cinematic language

The vocabulary a cinematographer already uses, and what each term does to a generated image. Lenses, film stocks, lighting styles and grades, plus a mood-to-parameters table that already agrees with itself, and two shot structures worth copying.

Open the cinematic reference

Say what you want, not what you don't

Most image and video models do not expose a separate negative prompt field, so a prohibition inside the prompt can act as a mention and pull the thing you banned into the frame. Where that is a risk, state the positive instead.

Most models don't expose a `negative_prompt`. Phrase positively:
Instead of "no blur" to "tack sharp"
Instead of "no people" to "uninhabited landscape"
NEGATIVE: color drift, photorealism, 3D render, lip-sync, captions, on-screen text, logos, watermark{, plus style-specific bans}.
NEGATIVE: <style drift and realism bans; no lip-sync, captions, text, logos, or watermark>
Regenerate with "no text, no letters, no logos, no captions" appended.
Never: <forbidden treatments>
Change ONLY the person's facial expression to: <phrase>. Keep identity, face structure, hair, pose, body, clothing, logo, background, lighting and composition EXACTLY unchanged, pixel-faithful.
Honor every `forbidden_elements` entry literally. Never replace one forbidden cliché with another generic symbol.
Known failure mode: false-positive `nsfw` flags on innocuous product prompts, the fix is removing mood/atmosphere words ("seamless loop feel", "ambient", "intimate") and re-describing the shot as a plain product film/photo.
Models reject prompts with `nsfw` or `ip_detected` terminal status. Avoid: Real public figures; Sexual content; Trademarks / branded characters

Guidance for specific models

Shown only for models and vendors this directory documents from public sources. Guidance naming products that could not be identified publicly is not published.

GPT Image 2 in the directory
Default high-fidelity image generation; graphic design, UI, banners, typography, and any brief with on-image text; generated product concept / packaging / can / bottle with brand name or lab
Nano Banana 2 in the directory
Fast everyday default for character, cartoon, stylized and reference-driven image work; edits; game sprites, tiles, backgrounds and textures; the style-key image in the explainer pipeline; m
Nano Banana 2 Lite in the directory
Fast reference-driven generation and edits when the brief is simple or speed/cost matters more than Pro fidelity.
Nano Banana Pro in the directory
Harder briefs where Nano Banana 2 falls short; the main thumbnail render; photoreal and reference-driven website imagery; deriving multiple angles of one approved product so every view is re
Higgsfield Soul 2.0 vendor
Aesthetic UGC, fashion editorial, lifestyle character stills; Soul Character reference work.
Recraft V4.1 in the directory
Clean graphic and vector-style design assets, logos, icons, flat illustrations, brand marks, controlled-palette visuals; the only model used for new logo marks.
Flux 2.0 in the directory
Precise prompt adherence; a creative alternative to the Banana family; key-pose generation in the sprite pipeline; cheap photorealistic drafts.
Flux Kontext Max in the directory
Context-aware editing and style transfer; anime, stylized looks, typography remix when defaults feel too generic.
Seedance 2.0 in the directory
Default all-purpose serious video, multi-shot, consistent identity, motion-heavy, cinematic, image-to-video, 4-15s up to 4K; seamless loops; the end-frame cover reveal; dramatic motion from
Seedance 2.5 in the directory
Reference-driven video, editing an existing video, or extending one.
Seedance 1.5 Pro in the directory
Cheap clean single-take shots without cuts; the exact model for the 2D sprite animation pipeline.
Kling 3.0 and Kling 3.0 Turbo in the directory
Cheaper Seedance 2.0 substitute for single-plane scenes without heavy motion; cleanest image-to-video with a start frame; Turbo when the user wants speed or lower cost.
Gemini Omni Flash in the directory
Fast multimodal reference-to-video from image references and optionally one video reference; the clip generator in the explainer pipeline.
Google Veo 3.1 vendor
Ultra-realistic top-tier cinematic quality; Lite for fast batch and volume work.
Minimax Hailuo in the directory
Cheap with strong physics when natural-physics motion matters and audio is not needed.
Wan 2.7 vendor
Synchronized audio with character-consistent video (2.7); stylized, experimental, intentionally artistic cheap work (2.6).

Where this comes from

Vocabulary and ordering advice adapted from production prompt documentation. Frequency statistics from the originating project's private prompt corpus have been removed rather than published, and no end-user prompt is quoted anywhere on this site.

Guidance is shown only for models or vendors this directory documents from public sources. 17 entries naming internal or unidentifiable products were withheld.

Worked example

Eleven films, one arc, one variable

A contributed set of 15-second hand-drawn VFX shorts, published in full with the method behind it: the five-beat arc, the texture language that keeps output from turning into vector CGI, and the trick that makes a generated camera move read as filmed rather than planned.

Open the VFX fusion set

Series concepts

Structures for a run of films rather than one: a fixed arc, a texture discipline, and exactly one thing that changes. Proposed rather than run, and labelled as such.

Ten inspirational concepts transformation-led shorts, one variable each
Educational and developmental toys eight concepts, none depicting a child's face

Audio tags

Bracketed instructions inside an audio prompt: what ElevenLabs documents for delivery and non-speech sound, and why Suno's [Verse] and [Chorus] tags are recorded here as convention rather than specification.

Open the audio tags reference

What the model publishers say

Anthropic and Google both publish prompting guidance for their own systems. Google's Veo and image guidance is first-party instruction for writing a shot; Anthropic's is about language models, and the parts that transfer are the ones about being explicit and showing examples. Linked to the originals, with the PDFs.

Read what they say, and where we differ

The lists

The same vocabulary with the explanation taken out: everything on one screen, in the order it goes in the prompt, with a PDF of the whole list at the bottom of each page. Use these when you know what you are looking for.

187 phrases for video prompts, grouped by the decision each one controls 187 phrases across 17 groups
294 phrases for image prompts, grouped by the decision each one controls 294 phrases across 24 groups
102 copy-paste prompts, each labelled with what it demonstrates 102 prompts across 4 groups
98 words to cut from your prompts, and what to write instead 98 words across 20 groups
14 mood recipes where the lens, the light and the grade already agree 14 recipes across 14 groups

Downloadable decks

The same vocabulary on one page, for pinning next to a monitor.

Open the downloads