Rivya Prompt Writing Guide

Start with the result you need and the details that must be correct. A useful prompt gives you something concrete to check afterward: the mug stays blue, the camera stays still, or the answer uses only the supplied facts. These examples are starting points, not tested outputs or promises of model performance.

Choose a model and one task

Use the model catalog to find an available model that supports your input and output. A prompt cannot enable an unsupported attachment, make a paused model available, or turn an image generator into a video editor. Check model availability when an option is missing.

Separate a larger project into requests: draft a script in chat, review it, then take the approved words into speech generation. Asking for a script, finished narration and edited video in one text box does not create that production workflow.

Choose supported duration, aspect ratio, resolution and voice options in the form. Mentioning them in prose does not replace those controls. Fields and length limits vary by model and mode; see model fields and parameters.

Write a concrete image or video brief

Include the subject, setting, composition or action, and the few constraints that matter. Replace vague praise such as “premium” with visible details.

Image example: “Create a product photo of a blue ceramic mug on a pale gray tabletop. Show the whole mug at a three-quarter angle, with soft light from the left. Use a plain background. Add no props, logos or text.”

This describes an invented mug. If you need to depict an actual product, attach a suitable reference where supported and check the result against the original. A detailed prompt alone does not establish the product’s real shape, dimensions or materials. See image workflows.

Video example: “A blue ceramic mug stands on a pale gray tabletop. The camera stays still as a hand slowly lifts the mug by its handle and moves it out of frame. Keep the background and lighting unchanged. Use one continuous shot.”

Choose a supported duration separately. Start with one action you can review through the entire clip; add more scenes only when the selected model and controls support that request. See video workflows.

Give each reference a clear role

Explain what you want to retain: “Use the uploaded photo for the mug’s shape and blue color; replace its background with a pale gray tabletop.” Naming that role expresses your intent, but does not guarantee exact product identity or consistent frames.

Use the file inputs the model actually provides. An image reference, a motion video and an audio recording are different inputs. A sentence describing a file does not attach it. Wait for uploads to finish and confirm the attachments before submitting; see references and uploads.

If several references disagree, resolve the conflict in your brief or remove the unnecessary reference. More files are not automatically more useful. For required text or logos, inspect every character and consider adding the approved artwork in an external editor instead of repeatedly asking the model to reproduce it.

Separate spoken words from an audio description

For text-to-speech, enter the words to be read. Speech example: “Your order is ready for collection. Please bring your order number.” Choose a supported voice in the form. Do not put a direction such as “write a friendly announcement” into the spoken-text field when you want the finished announcement read aloud.

Dialogue models may use a separate text field and voice for each line. An uploaded recording is only usable where that model provides the corresponding input; it is not a universal voice-cloning control. See audio workflows.

Music generation uses a description of the intended music rather than a chat instruction to write a script. Music example: “A quiet acoustic guitar piece with a slow, steady rhythm and a gentle piano melody.” Select the instrumental option where provided if you want no singing, and check the returned audio. For lyrics and other music tasks, follow music workflows.

Give chat the facts and the answer format

A useful chat request includes the task, relevant source material and the desired response format. Chat example: “Write three short product captions using only these facts: blue ceramic mug. Do not invent capacity, dishwasher suitability or delivery promises. Return a numbered list.”

For an image question, attach an image only to a model that supports it, and state what to examine. Do not refer to an attachment that is absent from the pending message. Ask for missing information to be identified rather than guessed, then verify the answer yourself. See chat workflows.

Review one change at a time

1. Check model availability, required files, prompt length and the current form settings. Remove conflicting instructions before submitting.

2. Review the credit estimate and send one request. Each new attempt can cost credits; a completed result you dislike is not automatically a failed task. See credits and billing.

3. Inspect the full image, watch the whole video or listen through the audio. Compare product details with the original reference and check any factual claims in text.

4. Keep a copy of the prompt and settings that produced a useful result. For the next attempt, change one important detail so you can judge its effect. Do not assume an unsent draft is saved; see saved sessions.

An example from Image Prompts or Video Prompts can be a useful starting point. Replace its subject and references, check the selected model’s controls, and set your own acceptance criteria before generating. Reusing the wording does not guarantee the same result.