博文

Write a five-second video brief for Seedance 2.5

A short video brief works best when every instruction contributes to one visible moment. Seedance 2.5 is available on RouterBase for text-to-video and image-to-video. Use a small scene to explore the workflow before building a longer sequence. Start with a subject and an action Choose something easy to describe: a cup with steam rising, a notebook page turning, or a slow reveal of a desk lamp. Write what the viewer should see happen. A clear action provides an acceptance test: did the page turn, did the steam move, or did the lamp remain the focus? Give the camera one job For a first prompt, specify a simple movement such as a slow push toward the subject. If the scene also includes a pan, an orbit, and several cuts, it becomes harder to understand a disappointing result. You can add complexity once you know which parts of the smaller brief are working. An illustrative prompt is: “A ceramic cup on a quiet café table. Steam rises from the cup. The camera slowly moves closer. Soft m...

A practical product-image brief for GPT Image 2.5

A product-image prompt is easier to evaluate when it reads like a small photography brief. GPT Image 2.5 Flare and Sunburst are available on RouterBase for creating images and editing existing ones. Here is a simple way to prepare a first experiment. Describe the placement first Suppose you need a square concept image for a handmade mug. Decide where the mug sits, what the background looks like, and which side needs empty space. These decisions matter because a pleasing picture can still be unsuitable for the page where you plan to use it. Try a brief like: “A matte blue ceramic mug on a warm beige surface. Soft light from the left. Eye-level view. Leave clear space above the mug for a heading. No additional objects.” This is an illustrative prompt, with no generated result attached. Make the requirements observable Write a short acceptance checklist before submitting. Is the handle visible? Does the composition leave the requested space? Is the surface free of distracting objects...

A meeting-note checklist for trying Gemini 3.8 Flash

Gemini 3.8 Flash is now available on RouterBase. Here is a small evaluation exercise from the RouterBase team: turn meeting notes into a useful follow-up without changing what the participants actually agreed. This article was drafted with AI assistance. Separate decisions from suggestions Use a short synthetic note with three kinds of information: one agreed decision, one proposal that was not approved, and one open question. Give each a clear label in your reference answer before sending the note to a model. For example: the team agreed to review a prototype on Friday; someone suggested moving the launch to Monday; no launch date was approved. A summary that says “Launch on Monday” has changed the decision, even if the rest reads smoothly. Ask for a bounded output A useful prompt is: “List decisions, proposed actions, and open questions separately. Include an owner or date only if the notes state it. Do not turn suggestions into commitments.” Compare the answer with the reference. Ch...

A support-ticket extraction exercise for DeepSeek V4.1 Flash

DeepSeek V4.1 Flash is now available on RouterBase. Here is a small exercise for evaluating it in a support workflow. This is a proposed test, not a report of measured model results. Use a fictional ticket Start with this input: My CSV export failed twice today. I am using the web app. I have not tried another browser, and I do not know the error code. Ask the model: Extract a short summary, the reported failure, known environment details, and missing information. Return JSON with the keys summary, failure, environment, and missing_information. Use null for unknown facts. Do not diagnose a cause that the ticket does not establish. The input is fictional and contains no customer information. Keep the prompt identical while comparing candidate models. Check the answer in two stages First, parse the output as JSON and check the required keys. Second, check meaning: the result should preserve the two failed attempts, mention the web app, and treat the browser and error code as unknown. Val...

A First Prompt for MiniMax H3 Max: Describe the Change

MiniMax H3 Max is now on RouterBase, with separate text-to-video and image-to-video workflows. Both pages list clips from 5 to 15 seconds with synchronized audio, in 480P or 768P. A useful first prompt describes a change that can happen within the requested duration. One subject and one clear action are easier to evaluate than an entire story compressed into a few seconds. Start with text Here is an untested example prompt: “A ceramic cup sits beside a rainy window. Steam rises slowly as the camera gently pushes toward the cup. Soft daylight, quiet rain against the glass, no dialogue.” This specifies a subject, action, camera move, lighting, and sound. Start with six seconds. After reviewing the result, change one instruction at a time so you can tell which revision helped. Start with an image If you already have a cup photograph, use image-to-video and focus on motion: “Steam rises from the cup. Rain trails down the window behind it. A gentle camera push, quiet rain, no dialogue.” Tre...

MiniMax H3 Prompting Starts Before the Prompt: Pick the Right Input Mode

A better MiniMax H3 result often starts with a product decision, not a longer prompt. RouterBase currently offers three H3 routes—text-to-video, image-to-video, and reference-to-video—with 2K output and 6- or 10-second clip options. Before writing, decide what the input asset is supposed to do. For text-to-video , describe five things: subject, action, environment, camera, and ending. Avoid stacking unrelated ideas. A prompt such as “a product bottle on wet stone” becomes more useful when it also specifies the motion, lighting, camera distance, and final composition. For image-to-video , treat the image as the first-frame anchor. Do not spend most of the prompt redescribing what is already visible. Explain what moves, what stays fixed, how the camera behaves, and what should happen by the final second. For reference-to-video , describe the relationship between the reference and the target. Should the model preserve the subject, borrow the style, follow the palette, or carry a product s...

A Small Reliability Checklist for OpenAI-Compatible AI APIs

Most AI API reliability work is not about adding another provider. It is about making ordinary failure states visible before a Friday deploy turns them into a support thread. Here is the small checklist I keep next to any model change. 1. Resolve the exact model ID before deploy A model name in a launch post is not an API identifier. Copy the exact ID from the live model catalog or models endpoint, put it in configuration, and log the value used by each request. This makes a rollback boring: change one reviewed config value, not several strings scattered through the codebase. 2. Give every request a client-side correlation ID Generate a request ID before calling the API. Carry it through your application logs, background jobs, and user-facing error reports. A client-side ID will not replace provider logs, but it gives your team one stable handle when the request crosses several services. 3. Separate retryable failures from fix-the-request failures Do not put every non-200 respons...