Four prompting reference sheets, for image, video, LLMs and AI studios: the techniques on them age slowly, the model names and settings quickly.
- Technique outlasts tools. Lighting vocabulary, camera language and shot grammar describe photography and film, so they hold across model releases.
- Model details expire. Syntax, parameters and which-model-to-use advice may last one release cycle, so check them against the models page.
- Supply the specifics. A model fills every gap with an average of what it has seen; the work is giving it the details only you have, not hunting for magic words.
- Context now matters more than wording. Frontier models handle a decent prompt, and the material, history and tools around a request move results more.
- Check what your model supports. Negative prompts and step-by-step instructions help on some models and not others, and Sora is being switched off.
Three things on these sheets moved since they were written.
1 · Sora is being switched off. The consumer app ended 26 April 2026; the API shuts down 24 September 2026. Any sheet below that names Sora as an option should be read as naming a deadline instead.OpenAI Help Center, Sora discontinuation, announced Mar 2026, read at source 10 Sep 2026: “the Sora web and app experiences were discontinued on April 26, 2026. The Sora API will be discontinued on September 24, 2026”
2 · Seedance 2.5 released 31 July 2026 — 30-second single-pass clips, co-generated audio, and up to 50 multimodal references (30 images, 10 video, 10 audio). The bracket format below still applies; the limits are higher.ByteDance’s own Seedance 2.5 page, re-read at source 17 Sep 2026: “Generate 30-second videos using R2V references, multimodal inputs of up to 50 references, and precise editing for ads, social media, ecommerce, and storytelling.” The single-pass length, co-generated audio and the 30/10/10 split of references are not on that page as re-read; they were reported as verified on 22 Aug 2026. An earlier version quoted “create 4K AI videos up to 30 seconds with up to 50 multimodal references”, which the page no longer says.
3 · Prompting matters less than it did. Frontier models handle a decent prompt fine. What moves the needle now is context — the material, history and tools around the request. These sheets remain the fastest way to get a specific result, but they are no longer the whole skill. See context engineering.
What on these sheets still holds
Worth separating, because the two categories age at completely different rates:
Where a sheet below gives a technique, it is likely still correct. Where it names a model or a parameter, check it against the models page, which carries its own check date.
Moved here from the original cheat sheets, word for word. These are the reference pages — the parts you come back to rather than read once.
A model fills every gap you leave with the average of everything it has seen. The work is supplying the specifics only you have — not hunting for magic words.
Technique lasts; model details expire. Lighting, camera language and shot grammar outlive any tool, while syntax, parameters and model picks may last one release cycle.
What each sheet covers
- Subject + Lighting + Camera + Style (this site's shorthand; BFL publish Subject + Action + Style + Context for FLUX.2 — FLUX.2 prompting guide, read at source 23 Sep 2026: “Use this framework for consistent results: Subject + Action + Style + Context”)
- Subject first — models weight earlier words
- Negative prompt only where the model has the field
- Start with [MOTION] always
- Lock subject identity
- Describe light, lens and colour
- Role + Context + Task formula
- Step-by-step only on non-reasoning models
- Specify format and length
- Specific beats vague, always
- Front-load important elements
- Iterate — never one-shot
- Character reference boards
- Identity consistency
- Full production pipeline
📸TEXT-TO-IMAGE — COMPLETE CHEAT SHEET
▸🎯 HOW TO WRITE A GREAT IMAGE PROMPT+
- Subject — who/what with full detail: age, clothing, expression
- Lighting — golden hour, Rembrandt, softbox, neon practical
- Camera — "Canon EOS R5, 85mm f/1.2" = photorealism signal
- Mood/Style — "editorial, serene, Vogue quality"
- Quality tags, only if your model’s guide asks for them — Google’s Imagen guide lists them; other vendors prefer naming the camera and lens (see prompting myths)
- Negative prompt — only if your model has the field. FLUX.2, Runway, Seedance and ChatGPT Image do not; on those, fold the exclusions into the description instead. See what the vendors say
▸🚫 UNIVERSAL NEGATIVE PROMPT+
- For portraits add: airbrushed, asymmetrical face
- For products add: hands, background clutter
▸📸 LIGHTING REFERENCE GUIDE+
- Golden hour — warm 5600K backlight. Portraits and outdoor editorial
- Rembrandt — dramatic triangle shadow. Character studies
- Studio softbox — clean, even. Product shots and beauty
- Neon / practical — colorful atmospheric. Urban, night scenes
- Overcast natural — diffused, shadow-free. Documentary
- Blue hour — cool ambient sky. Architectural
- Window light — soft directional. Indoor editorial
▸📷 CAMERA & LENS CHEAT SHEET+
- Canon EOS R5, 85mm f/1.2 — portrait, creamy bokeh
- Sony A7R IV, 35mm f/2 — street and candid
- Canon EOS R5, 24mm f/8 — wide landscape, everything sharp
- Hasselblad H6D, 120mm f/16 — product/macro, zero distortion
- Leica M11, 50mm f/1.4 — film aesthetic, timeless
- ARRI ALEXA, anamorphic — cinema grade, oval bokeh
▸🤖 WHICH MODEL TO USE WHEN+
- ChatGPT Images 2.5 — iterative editing, text in images, conversational workflow
- Seedream 5.0 — cinematic scenes, complex compositions, high realism
- Midjourney v8.2 — editorial aesthetics, artistic composition
- Nano Banana 2 Lite — rapid testing
- Ideogram — typography and graphic design
Describe the image you want, not a hope. Give the subject in detail, the lighting and the camera; add a negative prompt only if your model has the field, and otherwise fold exclusions into the description.
🎬TEXT-TO-VIDEO — COMPLETE CHEAT SHEET
▸✅ MUST-KNOW TERMS+
- CFG Scale — how strictly AI follows prompt. Ranges, and where CFG does not exist, are in the FAQ below
- Steps — refinement passes. Draft and final counts are in the FAQ below
- Seed — same seed + same prompt = same image. Lock before iterating
- Denoising strength — img2img: 0.35–0.50 = enhance, preserve
- Prompt weighting — (word:1.4) to boost, (word:0.5) to reduce
- LoRA — add-on file that specializes a model for a style or face
▸🎯 HOW TO WRITE A GREAT VIDEO PROMPT+
- Start with motion — describe what is moving first
- Verb-forward language — "walks slowly" beats "is walking"
- Lock your subject — "same face throughout, no morphing"
- Specify camera — tracking shot, dolly, aerial, handheld
- Describe the look — light source, lens and colour, not a director’s name (why)
- Set duration — "8 seconds, 24fps, no scene cuts"
▸📋 SEEDANCE BRACKET FORMAT (2.0-ERA)+
▸🎥 PLATFORM SELECTION GUIDE+
- Seedance 2.5 — human motion, dance, sports
- Higgsfield Studio — upload face, generate that person in video
- Kling 3.0 — physics-accurate action, water, fire, fabric
- Runway Gen-4.5 — VFX, transformations, commercial quality — View Runway Gen-4.5 guide →
- Pika — stylized, animated, artistic content
- Agent Opus — auto-edits video into TikTok/Reels clips
▸🎬 CAMERA MOVEMENT REFERENCE+
- Tracking dolly — moves alongside subject
- Dolly push-in — slow move toward subject, dramatic
- Handheld — natural shake, documentary feel
- Aerial drone descending — establishing wide shot
- Locked off tripod — static, subject moves within frame
- Pan — rotates on fixed axis
▸✅ TEMPORAL CONSISTENCY TIPS+
- "same consistent face and hair throughout all frames"
- "no morphing or flickering between frames"
- "no scene cuts" — forces single continuous clip
- Keep it tight — in practice, long prompts dilute the lines that matter
- Upload face photo to Higgsfield for identity fidelity
▸🎞️ LOOKS TO DESCRIBE, NOT DIRECTORS TO NAME+
A director’s name carries a look, unpredictably, and some tools refuse living artists’ names — see what a director’s name does to a prompt. Write the look itself:
- Naturalistic — practical light sources, restrained colour
- Epic scale — wide lens, small figures, hazy atmosphere
- Neon melancholy — warm neon practicals, slow motion, muted mood
- Vivid handheld — saturated colour, fluid handheld camera
- Soft pastel — soft window light, pastel palette
- Atmospheric — haze and smoke catching backlight, deep environments
- Clinical — desaturated grade, precise locked-off framing, cool fill
Describe the motion first, then lock the subject's identity, name the camera move and set the duration. Lock the seed before iterating so you can repeat a result.
💬LLM PROMPTING — COMPLETE CHEAT SHEET
▸✅ MUST-KNOW LLM TERMS+
- System prompt — instructions set before conversation begins
- Zero-shot — no examples provided
- Few-shot — 2-5 examples before your request
- Temperature — 0 = predictable, 1 = creative/varied
- Context window — how much the AI can "remember"
- Hallucination — AI confidently states incorrect info
▸🎯 THE 3-PART FORMULA+
- Role — "You are a senior brand copywriter for Chanel..."
- Context — "...writing for affluent women 30-50..."
- Task — "...write a 120-word product description."
▸🧠 CHAIN-OF-THOUGHT — NON-REASONING MODELS ONLY+
- Non-reasoning models only, and from practice. The idea is that it brings reasoning to the surface before the answer; OpenAI’s guides do not make that claim
- Unnecessary on reasoning models. OpenAI: “Since these models perform reasoning internally, prompting them to “think step by step” or “explain your reasoning” is unnecessary.”
- OpenAI: reasoning models do better with "only high-level guidance"; GPT-class models need "very precise instructions"
- Give a reasoning model the goal and the done-condition, not the steps
▸📐 OUTPUT FORMATTING+
- Specify format: "respond in JSON", "max 150 words"
- Exclude explicitly: "no corporate jargon", "no bullet points"
- Few-shot: show 2-3 example outputs you want matched
- XML tags: "wrap answer in <answer> tags"
- Word limits: forces concision, removes padding
Give role, context and task, and spell out format and length. Ask for step-by-step thinking only on non-reasoning models; give a reasoning model the goal and the done-condition.
🎭AI STUDIOS & AI INFLUENCERS — COMPLETE CHEAT SHEET
▸🎭 WHAT ARE AI INFLUENCERS?+
AI influencers are fully computer-generated personas — they post content, promote products, and build audiences without ever physically existing. Lil Miquela is one well-known example.
- 24/7 availability — never tired or unavailable
- Zero talent fees — no agents, no contracts
- Full brand control — persona never goes off-message
- Global scalability — same face, every market
▸🎬 AI MOVIES — WHAT'S POSSIBLE IN 2026+
- Script — GPT-6, Claude for screenplay
- Storyboard — ChatGPT Images 2.5 for visual boards
- Characters — APOB AI, Higgsfield Studio
- Scenes — Seedance 2.5, Runway, Kling 3.0
- Voice — ElevenLabs for realistic AI voices
- Music — Suno AI for original scores (Udio has disabled downloads; see the model guide)
- Edit — Agent Opus for automated assembly
▸🚀 FULL AI INFLUENCER CREATION STACK+
- Step 1 — Create persona: APOB AI or Glam AI
- Step 2 — Photos: ChatGPT Images 2.5 with reference
- Step 3 — Video: Higgsfield Studio
- Step 4 — Captions: GPT-6 brand voice prompt
- Step 5 — Short clips: Agent Opus
- Step 6 — Automate with n8n pipeline
▸🎯 AI INFLUENCER CONTENT STRATEGY+
- Define your niche — fashion, fitness, travel, tech, luxury
- 3 posts/week — consistency over volume
- Platform strategy — Instagram for photos, TikTok for video
- Brand voice — use GPT-6 brand voice system prompt for all copy
- Brand deals — pitch brands at 10K followers
- Affiliate income — promote AI tools for passive revenue
An AI influencer is a stack of tools: a persona tool, an image model, a video studio and a caption prompt. When posting, consistency beats volume.
⚡ Prompting hacks, FAQs and advanced techniques
🎯Front-load your most important element+
Some models read order this way. Black Forest Labs says it of FLUX.2, and tells you to put the most important elements first; Google’s Veo formula, by contrast, opens with the camera. On a model that weights what comes first, put your subject or most critical instruction at the start — "1woman, golden hour, Canon 85mm" rather than "Canon 85mm golden hour 1woman".Black Forest Labs, FLUX.2 prompting guide, read at source 17 Sep 2026: “Word order matters - FLUX.2 pays more attention to what comes first.” An earlier version of this answer said AI models in general weight earlier tokens more heavily; no source read here says that for every model.
🔢Specify quantity explicitly+
Write "1woman" not "a woman." Writing "1" prevents the model from generating groups, doubles, or ambiguous compositions. Works for any subject: 1man, 1cat, 1product. Explicit numbers = explicit intent.
📷Name the camera and lens every time+
"Shot on Canon EOS R5, 85mm f/1.2" is a quality signal. The model has learned that images tagged with professional cameras are high quality — naming the camera pulls the output toward that quality level automatically. Use it every time.
🎨Use publication references for style+
"Vogue Japan editorial" or "Kinfolk magazine aesthetic" communicates an entire visual world in two words. Publication references invoke lighting style, color palette, composition rules, and mood all at once. More precise than describing each element individually.
⚫Add negative prompts — where the model reads them+
First check your model takes one. Midjourney has a negative field; Veo takes a negativePrompt only on Gemini Enterprise Agent Platform (formerly Vertex AI), not in the Gemini API — see the Veo sheet; Kling has one only on its legacy API and recommends negatives inside the prompt. FLUX.2, Runway Gen-4.5, Seedance 2.5 and ChatGPT Image do not — and Runway documents that including a negative "may result in the opposite happening."
If it has the field: add CGI, plastic skin, airbrushed, watermark, blurry, overexposed. Short and specific. 30 seconds, real quality improvement.
If it does not: the same six items become positive description — natural skin texture with visible pores, even exposure, clean frame. Same intent, the only phrasing the model can act on.
Vendor documentation, checked 25 Aug 2026 — full comparison🎬Video mistake that kills temporal consistency+
Not locking subject identity. Without explicit instructions, the model drifts — face, hair, and clothing change between frames. Always add: "same consistent face and hair throughout every frame, no morphing, no identity change." For Higgsfield: use a reference photo as the identity anchor.
💬How long should a prompt be?+
As long as it needs to be — no longer. A 20-word prompt with specific, well-chosen elements outperforms a 150-word prompt full of filler. For image prompts: 40–80 words is the sweet spot. For video: 60–120 words. For LLMs: as long as the context requires, but structured clearly.
🔄Why does the same prompt give different results?+
AI models use a random seed by default. The seed controls the starting noise — same prompt, different seed, different output. To reproduce a result you like, save the seed number from the output metadata and lock it for subsequent runs.
🖼️Does prompt order matter?+
On some models, yes. Black Forest Labs says FLUX.2 pays more attention to what comes first, so on FLUX.2 lead with the subject, then lighting, camera, style and quality tags. Vendors do not agree on one order: Google’s Veo formula opens with the camera, and for LLMs Anthropic advises putting long documents above the question. Follow the vendor you are prompting.Black Forest Labs, FLUX.2 prompting guide, read at source 17 Sep 2026: “Word order matters - FLUX.2 pays more attention to what comes first.” An earlier version of this FAQ answer said AI models in general weight earlier tokens more heavily; no source read here says that for every model.
⚙️What CFG scale should I use? (and where CFG does not exist)+
On SDXL-based pipelines, where CFG is a real classifier-free-guidance scale: photorealism 5–7, artistic or stylised 7–9. Above 9 risks over-saturation and artifacts.
FLUX is different, and this answer has been corrected twice. Black Forest Labs' model cards say FLUX.1 [dev] was “Trained using guidance distillation” and FLUX.1 [schnell] was “Trained using latent adversarial diffusion distillation”. Do not assume an SDXL CFG number means the same thing on FLUX. FLUX.2 [pro] and [max] take no negative prompt at all.FLUX.1 [dev] and FLUX.1 [schnell] model cards, read at source 16 Sep 2026. An earlier version of this answer said both open models are guidance-distilled and run at CFG 1, and that a FLUX guidance of 3–4 approximates CFG 7; the cards support none of that for [schnell], and neither card mentions CFG 1. See the FLUX cheat sheet.
ChatGPT Image handles guidance internally — there is no manual setting, and no negative prompt field either.
📊How many steps should I use?+
20 steps for fast drafts. 25–30 for final production output.
🔍Why does the model ignore part of my prompt?+
Token limit and weighting. Very long prompts cause earlier elements to lose influence. Conflicting instructions cancel each other out. The model has trained on millions of images — some concepts are so rare in training data they're effectively invisible to the model. Simplify and front-load the critical elements.
🧬Use attention weighting for emphasis+
In Stable Diffusion interfaces that support weighting syntax: use (word:1.3) to increase emphasis on a term, (word:0.8) to reduce it. Not every pipeline reads this syntax, so check yours first. Example: "(golden hour:1.4), (Canon 85mm:1.2), (natural skin:1.3)" tells the model which elements matter most. Don't overweight — in practice, very high weights tend to cause artifacts.
🔗Chain prompts for iterative refinement+
Generate a base image → img2img at 0.40 denoising to refine → img2img at 0.35 for final polish. Each pass preserves the composition while improving detail. The Universal 8K Upscale prompt is specifically built for the final pass. Three-pass chain = commercial-grade output from any starting image.
🎭Reference images beat text descriptions+
A reference image communicates specificity that 200 words can't. For face consistency: use IP-Adapter Face ID or Higgsfield's reference upload. For style: use IP-Adapter at 0.6–0.7 weight. For video: a reference photo in Higgsfield Studio locks identity across every frame more reliably than any text instruction.
⚡Meta-prompting — use AI to improve your prompts+
Run your rough idea through the Meta-Prompt Upgrader (Prompt 22 in the vault) before sending it to an image or video model. The LLM adds camera specs, lighting, mood direction, and negative prompts automatically. Takes 30 seconds.
🎯Seed-lock for consistent series output+
Find a seed that produces a subject with the exact features you want. Lock the seed. Vary only the environment, lighting, and outfit across generations. This technique produces a consistent "character" across an entire image series without ControlNet or reference images — using only the seed as the identity anchor.
🧠LLM chain-of-thought — only on non-reasoning models+
Check which kind of model you are prompting first. On a fast non-reasoning model, "Think step by step before answering" is a technique from practice, not vendor guidance: OpenAI’s guides say it is unnecessary on reasoning models and that GPT-class models want precise instructions.
On a reasoning model it is redundant and can hurt. Those models already plan internally. OpenAI's guidance is that reasoning models do better "on tasks with only high-level guidance", while GPT-class models need "very precise instructions" — the two want opposite prompts. Give a reasoning model the goal and the done-condition, not the steps.OpenAI prompt engineering guide, developers.openai.com, checked 25 Aug 2026
Check what your model supports before applying a tip. Negative prompts, CFG and step counts exist on some models and not others, and step-by-step prompting is meant for non-reasoning models.