Ten tools, and each vendor publishes its own idea of what a prompt is. They do not agree — and the disagreements are the useful part.
- Four of seven models take no negative prompt at all. FLUX.2, Runway, Seedance and ChatGPT Image have no field. Runway documents that using one may produce the opposite.
- Every vendor formula puts the subject early — except Google, which puts the camera first.
- Word order carries weight. Black Forest Labs states plainly that FLUX.2 attends most to what comes first.
- 30–80 words is the documented sweet spot for FLUX.2. Longer is for genuinely complex scenes, not for more obedience.
- In a negative field, write nouns, never sentences.
blurry, watermark— notno blurry. - OpenAI rewrites your prompt before generating it. What you typed is not what the model received.
- Audio has no prompt box. On ElevenLabs the script is the prompt — punctuation and capitalisation are the controls, and v3 and v2 take incompatible tag syntax.
- Suno is two fields. Style and lyrics are separate; mixing them is why half a prompt gets ignored.
- ByteDance publishes an eight-part Seedance formula that puts format first and action third — not the motion-first bracket format most guides teach.
The table nobody prints
Every negative-prompt guide online hands you one list and implies it works everywhere. Check the vendor documentation and that collapses immediately. This is the single highest-value thing on this page.
--no, comma-separatedYESnegativePrompt on Vertex AIYESBefore you write a negative prompt, find your model in that table. On four of seven, the correct move is not a shorter negative — it is no negative at all, with every exclusion rewritten as a description of what you want present.
The formulas, as published
Image and video first, then text, then audio. Where a vendor publishes no consolidated guide — Midjourney, Suno, Higgsfield — the entry says so and the sourcing is flagged as weaker.
FLUX.2 — Black Forest Labs
Subject + Action + Style + Context
The distinctive claim is about order, not content: the model attends most to what comes first. Their stated priority is main subject → key action → critical style → essential context → secondary details. Bury the subject at the end of a long sentence and you are working against the model.
Length: 10–30 words for quick concepts; 30–80 for most work; 80+ only for complex scenes. They also support a JSON prompt schema for production use, and note that hex colour codes only work reliably when attached to a specific object — a stray code with no object attached will not hold.
Black Forest Labs, FLUX.2 [pro] & [max] prompting guide, docs.bfl.aiVeo 3.1 — Google
[Cinematography] + [Subject] + [Action] + [Context] + [Style & Ambiance]
The only major formula that opens with the camera rather than the subject. Shot size and camera move come first, then who, then what they do.
Audio has its own syntax, which most guides miss entirely — dialogue goes in quotation marks, effects are prefixed SFX:, and background is prefixed Ambient noise:.
On negatives: Veo has the field, but Google's instruction is to describe the exclusion as something present. Their own example — write "a desolate landscape with no buildings or roads", not "no man-made structures".
Google Cloud, Ultimate prompting guide for Veo 3.1Runway Gen-4 / 4.5
Subject → Action → Setting → Camera → Motion over time → Style
Full sentences, natural language. Runway is explicit that simple prompts already produce strong results and that added length buys stylistic control rather than obedience — the opposite of the keyword-stuffing habit carried over from older models.
The hard rule: "Avoid negative prompting, such as no clouds in the sky, for the best prompt adherence." And the reason it matters — their documentation states that including a negative may result in the opposite happening. This is the one model where the folk warning is the vendor's own.
Runway, Gen-4 Image and Gen-4 Video prompting guides, help.runwayml.comSeedance 2.5 — ByteDance
Format + Subject + Action + Environment + Camera + Look + Timing + Audio & Constraints
ByteDance publishes this eight-part structure through Dreamina, its own platform. It is not the bracket format most guides teach — including, until now, this site. Format comes first (duration, aspect ratio, shot structure), the subject second, and the action third, not first.
Their published template: [Duration, aspect ratio, shot structure]. Use [@reference] as the consistent [character/product/location]. In [environment and time], the subject [precise action]. Camera: [shot size + movement + lens/focus]. Visual direction: [lighting + palette + texture + realism]. Timing: [beat 1] → [beat 2] → [beat 3] → [resolution]. Audio: [dialogue/ambience/SFX/music]. Preserve [identity, wardrobe, logo, object geometry]. Avoid [unwanted cuts, camera behavior, artifacts, extra objects, text].
Four rules inside that template that most guides miss:
— Bind references with @Image1, @Image2. Tag assets to specific beats rather than hoping the model infers which reference governs what.
— Shot size before style language, and one camera move per shot. Two moves in one shot is a contradiction the model has to resolve.
— Roughly 6–8 seconds per beat is the stated default for timing.
— Preserve, then avoid. The template gives exclusions their own slot, short, concrete and grouped at the end — after a positive Preserve [...] clause. That ordering matters: state what must survive before what must not happen.
On negatives: there is no negative_prompt field, so exclusions share the prompt and compete with it. Long generic lists contradict the positive direction — banning blur while asking for shallow depth of field, banning shake while asking for handheld. Keep the Avoid clause to a handful of unusable-clip failures.
Kling 3.0 — Kuaishou
A real negative field — around 2,500 characters, separate from the prompt
The most conventional of the video models, and the one where habits from Stable Diffusion transfer cleanly. The field recognises entries as exclusions automatically, so do not write "no" — just the terms.
Capacity is not permission: the documented advice is a handful of focused terms aimed at glitches you have actually seen, not a dump of every fear.
Kling API documentation and published prompting guidesMidjourney v8.2
--no at the end, comma-separated
One --no per prompt — a second one does not stack. It is equivalent to a negative weight, so excluding a term behaves like a strong push away from it, not an absolute ban.
The trap: each word is treated separately. Writing two words after --no without a comma reads as two separate exclusions, not one phrase. Use commas deliberately.
ChatGPT Image — OpenAI
No published formula — and your prompt is rewritten before use
OpenAI documents implementation, not prompt craft. The consequential detail is buried in the API reference: the model automatically revises your prompt before generating, and the rewritten version is returned in a revised_prompt field.
This changes what iteration means. When output does not match your prompt, the cause may be the revision rather than the wording — so read revised_prompt before rewriting. OpenAI also states the model struggles with precise text placement and layout-sensitive composition, which no amount of prompt tuning fixes.
ElevenLabs — voice
The text is the prompt. There is no separate prompt box.
Punctuation and capitalisation are the primary controls: ellipses create pauses and add weight, CAPITALS add emphasis, ordinary punctuation sets rhythm. These work on every version and are the safest lever.
The version split matters more than the tags. v3 takes inline audio tags — [whispers], [excited], [sigh] — and does not support SSML break tags at all. v2 takes <break time="1.5s" /> up to a 3-second maximum, and ElevenLabs warns that too many in one generation "can cause instability." Syntax from the wrong version fails silently rather than erroring.
The limit that catches people: a tag cannot override the voice. ElevenLabs states a voice will not contradict its training — [shout] on a voice trained on whispering will not shout. Tag effectiveness varies by voice, and Professional Voice Clones are not fully optimised for v3.
Suno — music
Two fields, not one — and this is where most prompts leak.
The style prompt carries the sound: genre, mood, instrumentation, vocals, BPM, as comma-separated descriptors rather than sentences. The lyrics field carries the words plus structural metatags — [Verse], [Chorus], [Bridge]. Genre written into the lyrics box, or lyrics into the style box, is the usual reason half a prompt appears ignored.
Reported limits: roughly 1,000 characters of style and 5,000 of lyrics on v4.5 and later; about 200 and 3,000 on older v3.5/v4. The practical ceiling sits below the cap — past roughly 3,000 characters of lyrics, songs start rushing or truncating. Reported sweet spot is 5–8 style descriptors; beyond about ten they conflict.
Weaker sourcing than the entries above. Suno publishes no consolidated prompting reference, so these figures come from third-party guides and API documentation (Aug 2026) which disagree at the margins. Verify in the app before relying on a number.Higgsfield — motion
Name the move in the prompt and select the matching preset.
Higgsfield's distinguishing feature is a library of 50+ named cinematic motion presets — crash zoom, bullet time, dolly, crane, FPV drone, Snorricam. The workflow is not prompt-only: you pick the preset in the app, then describe subject and scene in the prompt. Naming the same move in the text as well as selecting it is what makes results repeatable rather than approximate.
Because the camera move comes from the preset, the prompt's job is narrower than on other video tools: subject, identity anchoring, environment and action. Describing an elaborate camera move in text while a different preset is selected puts the two in conflict — the same self-contradiction problem that breaks long negative prompts elsewhere.
Higgsfield camera-controls preset library, higgsfield.ai/camera-controls, checked 25 Aug 2026. Higgsfield does not publish a single consolidated prompting guide; the workflow above is drawn from the preset documentation and platform guides.Where all six agree
Strip the branding and three rules survive every vendor document:
The checklist
Before you send a prompt:
SFX: and Ambient noise: prefixed?□Every claim here traces to a vendor's own documentation, named in the source line under each section. Model versions and documentation both move — this page was checked on 25 August 2026 and the negative-prompt table is the part most likely to change first. Re-read the vendor page before relying on any row of it. Errors get logged at corrections.
Where vendors contradict each other, this page says so rather than picking the tidier answer — see the Seedance note above.
The through-line: prompting is not one skill applied to seven tools. The tools disagree about what a prompt even is, and the vendors have written that down.