Define the deliverable
State where the output will appear, its format, audience and the one idea it must communicate. A social clip, storyboard, product frame and campaign asset need different precision. Clear acceptance criteria stop the prompt becoming a list of unrelated wishes and make review more objective.
Choose the strongest input
Use text for a new scene, an image for identity or composition, keyframes for defined moments and reference video for motion or camera language. Choose the input that contains the hardest information to describe. Extra references only help when each asset has a clear purpose.
Structure the brief
Draft subject, action, environment, composition, camera, light, materials, style, sound and ending separately. Convert those fields into direct natural sentences. Put the most important requirement early. Remove decorative adjectives that do not describe visible or audible evidence.
Assign reference roles
Prepare clean source assets and explain what each should control. One image can define identity, another material, and a clip can define rhythm. State which elements must remain stable and which may change. When sources conflict, establish a priority instead of asking the model to guess.
Control composition
Describe subject position, viewpoint and how much environment is visible. For video, match camera movement to the action and define the pace and final beat. For images, use lens or layout language only when it supports the design purpose. Too many competing camera instructions reduce clarity.
Plan sound with the scene
Connect dialogue, effects and ambience to visible causes. Identify the speaker and keep lines appropriate for the duration. Review timing, pronunciation and continuity. If audio will be replaced in post-production, focus generation effort on visual decisions instead of repeatedly refining temporary sound.
Review systematically
Check prompt adherence before aesthetics. Then inspect identity, anatomy, geometry, typography, materials, motion, camera, sound and the ending. Watch video at normal speed and inspect key frames separately. Record the failure category so the next revision addresses a specific problem.
Revise one variable
Preserve successful instructions and correct one major weakness at a time. Adjust motion without replacing the location, or change lighting without rewriting identity. Controlled iteration reduces drift and reveals cause and effect. Save prompt versions, settings and references beside selected outputs.
Check continuity
For a sequence, repeat stable identity, wardrobe, product, location and lighting descriptions. Compare the end of one clip with the beginning of the next. Review screen direction, camera height, ambience and temporal logic. Continuation helps, but every join still needs human approval.
Verify final details
Confirm names, numbers, claims, logos, packaging and text outside the model. Make sure uploaded references can be used and the result follows destination-platform rules. Apply color, captions, sound and compression for delivery. Preserve the original generation and its inputs for traceable corrections.
Track real cost
Measure the number of attempts needed for an approved result, not only the advertised cost of one generation. Include discarded outputs, extensions, external editing and review time. Availability, queues and rate limits also affect production. Prices on this site are service prices, not direct vendor API prices.
Keep evidence current
Official documentation supports statements about named inputs, output limits and features. It does not guarantee identical quality for every prompt. Record the date and model version of important tests. Early-access products change, so repeat critical checks before a campaign or major purchase.
Prepare source files
Use clean, high-quality references without unrelated overlays or heavy compression. Crop assets so important identity, geometry and style information is visible. Keep originals separate from working files, use descriptive filenames and record which source was used. Careful preparation reduces ambiguity and supports repeatable tests.
Set priorities explicitly
Separate requirements into must keep, should keep and flexible. Identity, product shape or required copy may be non-negotiable, while background detail or color variation may be open to interpretation. Put priorities in the prompt and evaluation sheet. This prevents attractive secondary details from hiding failure on the main objective.
Design shorter test cases
Before generating a complex final scene, isolate difficult requirements. Test identity with simple motion, typography in a stable frame, or dialogue in a close shot. Small tests reveal whether an approach is viable without paying for repeated full compositions. Combine controls only after their individual behavior is understood.
Document failure patterns
Create categories such as identity drift, anatomy, geometry, camera, timing, text, sound and policy rejection. Record which failures repeat across attempts and which respond to prompt changes. Patterns are more informative than a vague quality score. They also help a team decide whether to revise, use another mode or switch tools.
Use human review gates
Assign explicit approval points for the brief, references, first output, revisions and final delivery. Different reviewers may own factual accuracy, brand consistency, legal risk and creative quality. A generation should not move forward merely because it looks polished. Review gates keep errors from becoming more expensive later in the workflow.
Plan accessibility
Provide meaningful alt text for published images and captions or transcripts for video where appropriate. Check contrast and readability when generated text is part of a design. Do not rely on audio alone to communicate essential information. Accessibility work belongs in the production brief rather than being added only after an asset is approved.
Protect sensitive material
Avoid uploading confidential, personal or restricted references unless the service terms and project policy allow it. Remove unnecessary metadata and obtain permission for identifiable people and protected assets. Generated content can still create privacy, likeness and intellectual-property concerns. Treat source governance as part of creative preparation.
Test delivery formats
Review the asset after the crop, resize and compression used by the destination. Fine text, faces and rapid motion can change noticeably on social platforms or small screens. Export a representative draft early. A result that works in a large preview may need different framing, pacing or typography for the final placement.
Create a reusable template
Turn successful briefs into templates with fields for goal, subject, action, references, camera, light, sound, ending, exclusions and approval criteria. Keep examples of good inputs and common failures beside the template. Reuse the structure, not the exact creative content, so new tasks benefit from previous learning without becoming repetitive.
Know when to stop iterating
Set a maximum number of attempts or a credit budget before starting. Stop when the output meets the defined criteria, when improvements become marginal, or when repeated failures indicate that another workflow is needed. Unlimited revision can cost more than external editing or a simpler production method. A stopping rule makes experimentation accountable.
Record reproducibility data
Store the complete prompt, mode, duration, resolution, aspect ratio, reference order and generation date with every shortlisted output. Note any external edits separately. Reproducibility does not mean a stochastic model will return identical pixels, but it gives the next operator the same starting conditions and makes meaningful comparison possible.
Plan team handoff
Summarize the creative goal, locked elements, flexible elements, rejected directions and remaining risks when work moves between people. Include links to approved references and the latest prompt. A concise handoff prevents a new operator from repeating failed experiments or unintentionally removing constraints that protected identity, geometry or brand consistency.
Maintain version labels
Use simple version names for prompts, references and outputs, and never overwrite the approved source file. Connect each revision to a reason such as camera correction, text correction or audio timing. Version labels make stakeholder feedback precise and allow the team to return to the last stable result when a later experiment introduces new problems.
Review after publication
Check the asset in its real destination after publishing. Confirm that platform cropping, autoplay, captions, compression and volume behave as expected. Record audience or stakeholder feedback that reveals a production issue. A short post-publication review improves the next brief and turns one generation task into durable operational knowledge.
Update the workflow
When a model, interface or pricing rule changes, revise templates and checklists instead of relying on memory. Remove instructions that no longer help and document new controls with a dated test. A maintained workflow keeps long-form guidance accurate and prevents teams from repeating advice that belonged to an older model or access tier. Review this process before every important production cycle.
Run a like-for-like test
Use the same subject, action, reference pack, aspect ratio and duration target in both products. Preserve meaning even when parameter names differ. Generate several outputs instead of selecting one lucky sample, and retain failures. This reveals consistency as well as peak quality.
Score adherence separately
List every requirement and mark whether it appears correctly. Separate adherence from visual taste: an attractive result that changes the product or omits the action has not completed the brief. Weight identity, text and required events more heavily than decorative details.
Compare temporal quality
Review faces, clothing, objects, architecture and text across the timeline. Look for shape changes, lighting jumps and camera moves that contradict the scene. Longer duration is valuable only when the requested content remains coherent throughout the stated generation window.
Choose by scenario
Exploration values speed and range; brand work values identity, typography and geometry; narrative work needs continuity, camera and audio. Build a scorecard for the real use case rather than averaging unrelated strengths. The best model for one production category may be a poor fit for another.
