Prompt craft · other people's guides

What Anthropic, OpenAI and Google say about prompting

Three model publishers wrote down how to prompt their own systems. Two of them wrote guides specifically for their video models, and those are the closest thing to a first-party specification for how to write a shot.

Every link below was fetched on 2026-08-17 and is served by the publisher's own domain.

How to read this page

These are other people's guides. Everything here is attributed, linked and summarised rather than reproduced, and where a guide says something this site disagrees with, the disagreement is recorded rather than smoothed over.
Anthropic's guidance is about a language model, not a media model. It earns its place because a media prompt is still a written brief, and the failure modes of a vague brief are the same in both.
OpenAI's Sora 2 guide and Google's Veo guide are both first-party instruction for writing a shot, and they agree with each other more than either agrees with the general prompting advice.

Anthropic

Publishes prompting guidance for Claude as living documentation rather than as a single downloadable book. The docs are the reference; the PDF below is a one-page teaching handout, not the full guide.

The originals

Prompt engineering overview Docs
The entry point. Its own advice is to establish success criteria and a way to test against them before touching the prompt at all.
Prompting best practices Docs
The living reference the overview points at. Covers clarity, context, examples, XML structuring, roles, long context, output format and extended thinking.
6 Techniques for Effective Prompt Engineering PDF · CC BY-NC-SA 4.0
A one-page before-and-after handout on Anthropic's own CDN. Copyright 2025 Rick Dakan, Joseph Feller and Anthropic, released under CC BY-NC-SA 4.0, so it may be shared and adapted with attribution for non-commercial use under the same licence.
Interactive prompt engineering tutorial Tutorial
Nine chapters with exercises, linked from the official docs. There is a lighter spreadsheet version linked alongside it.

Their techniques, and what each one means for a shot

Technique	What they say	What it means for a shot
Be clear and direct	State the task plainly and say what the output should contain. Vagueness is filled in by the model, not by you.	The single highest-yield habit in a shot prompt. "Cinematic" is a wish; "85mm, eye level, single key from camera left" is a brief.
Add context	Scope, audience, constraints and purpose change the answer, so state them.	Where the shot will run decides how it should be framed. A vertical cut-down has half its height reserved by interface, which is a constraint the prompt should carry.
Use examples	Showing two or three examples of a good answer teaches a pattern that description alone does not.	This is why a reference image outperforms an adjective, and why reusing identical phrasing across a batch keeps a set looking related.
Structure with XML tags	Delimit the parts of a prompt so the model can tell instruction from content from example.	Media prompts arrive at the same idea from a different direction: the labelled block form, where SCENE, MOTION, AUDIO and NEGATIVE are separate lines rather than one paragraph.
Break complex tasks into steps	Numbered steps guide the reasoning and stop a task being answered in one undercooked pass.	The beat-blocked scene is exactly this, applied to time: 0-3s, 3-7s, 7-11s, each with its own action rather than one instruction for the whole clip.
Give it a role	Naming a role shapes tone and approach.	Weak on its own for images. Naming the craft is what carries: writing as a cinematographer means naming lens, height and light, and it is the naming that does the work, not the job title.
Ask it to think first	Space to reason before answering produces more considered output.	Applies to the writing of the prompt, not to the render. Use a language model to draft and interrogate the shot list, then send a finished brief to the media model.
Ask the model to help you prompt	When you cannot phrase the request, ask for help phrasing it. Anthropic's handout calls this "perhaps the most powerful technique of all".	Works well for media because the vocabulary is the barrier. A model that knows the words can convert what you can picture into terms a generator responds to.

OpenAI

Publishes prompting guidance as API documentation and as cookbook guides, including one written specifically for Sora 2 that is the most directly useful of the three publishers for anyone writing a shot. Notably, no prompting PDF: see the note on formats below.

The originals

Prompt engineering guide Docs
The API-side reference. Names eleven strategies, of which the transferable ones here are message roles, markdown and XML structure, few-shot examples, and the split between explicit step-by-step instruction for general models and high-level guidance for reasoning models.
Sora 2 prompting guide Guide
OpenAI's own guide for its video model, updated March 2026. First-party instruction on how to write a shot, and explicit about which parameters a prompt cannot reach.
ChatGPT Enterprise: practical prompt engineering for everyday work Guide
By Sophia Qin and Aaron Wilkowitz, April 2026. Teaches a three-part frame — context, instructions, additional information — and the habit of treating a prompt as a handoff to a new colleague rather than a command.

Their techniques, and what each one means for a shot

Technique	What they say	What it means for a shot
Structure the prompt, do not narrate it	Use markdown headings or XML tags so the model can tell instruction from content from example.	The same reason the labelled block form works for shots: SCENE, MOTION and AUDIO as separate lines beat one flowing paragraph.
Context, instructions, additional information	Background and persona first, then the specific task, then constraints, format and acceptance criteria.	Maps almost exactly onto scene, action, then the technical constraints a shot has to satisfy — aspect ratio, duration, where it will run.
One deliverable per prompt	Scope tightly. A prompt asking for several things produces several half-answers.	One clear action per shot. A model given three beats in one clip compresses them into one.
Clarity over cleverness	Treat the prompt as a handoff to a new colleague who has no context.	The single most useful sentence in any of these guides for media work, and the one most often ignored in favour of adjectives.
Meta-prompting	Ask the model to improve your prompt before you run it.	Cheap, and it works: the barrier in media prompting is vocabulary, and a language model has the vocabulary.

The formula they publish

[Scene description] + [Cinematography] + [Actions] + [Dialogue]

Part	What goes in it
Scene description	Plain prose: characters, costume, setting, weather.
Cinematography	Framing, camera motion, depth of field, lighting and lens.
Actions	Specific beats or gestures, written as a numbered list.
Dialogue	In its own block below the prose, in a <dialogue> section, with speakers labelled consistently.

Vocabulary

Group	Words they list
Framing	wide establishing shot, medium close-up, eye level, low angle, overhead
Camera motion	slow dolly-in, tracking left to right, handheld, tilting
Depth of field	shallow focus, deep focus
Lighting	key direction, fill sources, rim light, practical lights
Lens	anamorphic, spherical primes, 32mm, 50mm, 85mm

What the prompt cannot reach

Resolution, clip length and character references are API parameters, not prose. Asking for a longer clip in the prompt does nothing.

model
size
seconds (4, 8, 12, 16 or 20)
characters (up to two references)

Negatives

This guide does not address negative prompting.

OpenAI states the structure is not a one-size-fits-all recipe, and that leaving detail out invites the model to interpret.
The guide does not address negative prompting at all, which is itself worth knowing.
It recommends generating a still first to lock composition and style, then letting the text define the action.

Google

Publishes both general prompting guidance and, unusually, first-party guidance for its own video and image models. The media-specific pages are the ones worth reading here.

The originals

Ultimate prompting guide for Veo 3.1 Guide
Published 2025-10-16 by Khulan Davaajav and Hussain Chinoy on the Google Cloud blog. The most directly useful document in this list for anyone writing a shot.
Image generation prompt guidance Docs
Gemini API documentation. Carries fill-in templates for photographic, illustrated, product and text-in-image work.
Writing effective prompts for Gemini for Workspace Guide
The four-part frame Google teaches to everyone: persona, task, context, format. Written for documents and email rather than for pictures.
Gemini for Workspace prompting guide (PDF) PDF
Served from Google's own file host. A large document aimed at office work; the transferable part is the persona-task-context-format frame.
Prompt Engineering (whitepaper) PDF
By Lee Boonstra, published with Google on Kaggle, November 2024 and updated since. The most academic of these: zero-shot and few-shot, system and role prompts, chain of thought, self-consistency, tree of thought, and configuration such as temperature and top-p.

The formula they publish

[Cinematography] + [Subject] + [Action] + [Context] + [Style and ambiance]

Part	What goes in it
Cinematography	Camera work and shot composition, stated first.
Subject	The who or what the shot is about.
Action	What the subject does.
Context	Environment and background.
Style and ambiance	Aesthetic, mood and lighting.

Vocabulary

Group	Words they list
Camera movement	dolly shot, tracking shot, crane shot, aerial view, slow pan, POV shot
Composition	wide shot, close-up, extreme close-up, low angle, two-shot
Lens and focus	shallow depth of field, wide-angle lens, soft focus, macro lens, deep focus

Sound

Kind	How to write it	Example
Dialogue	Put the spoken line in quotation marks.	"We have to leave now."
Sound effects	Label them and describe them plainly.	SFX: thunder cracks in the distance
Ambience	Name the background soundscape.	Ambient noise: the quiet hum of a starship bridge

Negatives

Describe what to exclude affirmatively rather than as a prohibition.

Instead of “no man-made structures”
write “a desolate landscape with no buildings or roads”

Workflows

Workflow	What it does
First and last frame	Supply both ends and let the model make the transition between them.
Ingredients to video	Reference images carried across shots to hold a scene together.
Timestamp prompting	Direct a multi-shot sequence with [MM:SS-MM:SS] notation so pacing is stated rather than hoped for.

Their fill-in template for stills

A photorealistic [shot type] of a [subject] in a [setting]. [Light description]. Shot from a [angle] with a [lens].

Illustration work wants the style named, plus the qualities that carry it, such as bold outlines or cel shading.
For text inside an image, state the exact words, describe the font rather than naming one, and say where it sits.
Product work wants the studio lighting setup, the surface, the angle and what is in focus.

Which of these exist as a PDF

Asked directly, because a downloadable file is easier to keep than a URL. The answer differs by publisher and it is worth knowing before you go looking.

Publisher	PDF?	What exists
Google	Yes	Two: the Gemini for Workspace prompting guide on Google's own file host, and the Prompt Engineering whitepaper published with Kaggle. Both linked above.
Anthropic	One, and it is a handout	"6 Techniques for Effective Prompt Engineering" on Anthropic's own CDN, CC BY-NC-SA 4.0. The full guidance is documentation, not a book.
OpenAI	No	Its prompting guidance is documentation and cookbook guides. The ChatGPT Enterprise guide notes a markdown version is available by appending .md to the page URL, which is the closest thing to a downloadable file. Several PDFs of OpenAI guidance circulate on document-sharing sites; those are re-uploads and are not linked here.

Where the publishers and this site agree

Independent sources converging is the strongest signal available in a field where most advice is one person's habit.

Where they agree	The detail	Why it matters
State the negative as a positive	Google's Veo guide says to write "a desolate landscape with no buildings or roads" rather than "no man-made structures". The concept pages on this site make the same point from the mechanism: attention has no natural operator for not, and a banned word is still a word in the prompt.	Two independent sources and a mechanical explanation agreeing is about as close to settled as prompt advice gets.
Time should be stated, not implied	Google documents [MM:SS-MM:SS] timestamp prompting for multi-shot sequences. The practitioner reference on this site independently arrived at beat-blocked scenes written 0-3s, 3-7s, 7-11s.	Both are solving the same failure: a model given one instruction for a whole clip compresses a three-beat idea into one, or pads one beat with drift.
Cinematography goes first	Google's formula opens with the camera before the subject. The mood recipes on this site also start from angle and shot size rather than from what is in frame.	The first clause of a prompt gets the most weight, and the camera decides what the subject even looks like.
Examples beat adjectives	Anthropic's guidance is that two or three examples teach a pattern description cannot. In media that is the reference image, and the same logic explains why reusing identical phrasing keeps a batch consistent.	It is the one technique that transfers unchanged from text to picture.
Structure beats prose	All three now say it. Anthropic uses XML tags, OpenAI recommends markdown headings or XML and a context-instructions-information frame, and Google's Veo formula is five named slots in a fixed order. The practitioner pack on this site independently arrived at the labelled block form.	Four sources converging on the same shape is the strongest signal in this whole document: a prompt is a form to fill in, not a paragraph to write.
The prompt does not control the parameters	OpenAI is explicit that resolution, clip length and character references are API parameters and that asking for them in prose does nothing. Google says the same about Veo's upscale ceiling and duration.	It explains a whole class of apparent model failure. "Make it longer" is not a prompt instruction, and no amount of rephrasing turns it into one.
Timestamped beats are now corroborated four ways	Google documents [MM:SS-MM:SS] notation for Veo, OpenAI's Sora 2 guide structures a prompt around timed actions, the practitioner reference here arrived independently at beat-blocked scenes, and a contributed set of eleven hand-drawn VFX shorts gives every beat an explicit timestamp and reports what happened when it did not: at twelve transformations in fifteen seconds the model silently dropped one.	The failure mode is the useful part. A model given too much for the time does not error and does not visibly truncate — it quietly omits a beat, which you only notice by watching for something that never arrives.

Where we take a different line

Where this site differs	The detail	Our position
Role prompting	Anthropic recommends giving the model a role. For media that mostly does nothing on its own: "act as a cinematographer" does not change an image the way naming a 85mm lens and a key light does.	Name the decisions a cinematographer would make, not the job title.
Longer prompts	General guidance leans toward adding context. Media models distort with very long prompts, and adding tokens spreads attention across more of them.	Add context that changes a decision. Cut context that only adds atmosphere.

What is deliberately not linked

Several highly-ranked results for these guides are third-party re-uploads on document-sharing sites, and one is a paid resale of freely published material.

Only URLs served by the publisher's own domain are linked here. A re-upload cannot be trusted to be current, and linking one would send readers to a copy that its author never updated.