Awesome MiniMax H3 Max Prompts — Separating a reference collection from an evaluation
The repository records public examples and their available context. It is a resource for forming hypotheses about motion and inputs; it does not establish that an inferred prompt reproduces a creator’s output.
These records could help someone plan an experiment, but they are not a benchmark. The source videos do not share controlled inputs, a fixed model configuration or a common evaluation protocol. Before comparing models, define those conditions independently and generate new results under them. The repository's categories are useful for selecting a task, while its sources explain the origin of an example. Neither a category label nor an attractive preview establishes a score.
What we actually collected
the open-source H3 Max collection contains 30 cases from 15 public accounts across seven creative categories, with English and Chinese documentation. The categories cover camera movement and action, character performance, animation, short narratives, documentary-style atmosphere, reference continuity, and production workflows. The examples include rooftop combat in the rain, drifting astronauts, clay-style comedy, character showcases, and Blender-based camera previsualization. These are different production questions, so a category should help a reader choose a task rather than imply a ranking of models or creators.
Describe something a camera can observe
For a parkour scene, a useful brief specifies the starting position, obstacle, takeoff, landing, and camera path. That gives a reviewer several separate checks. If the landing is wrong, change the action description before adding more style words. If the camera crosses the subject, revise its path while keeping the action stable. This is an editing method for making experiments understandable, not a measured claim that every detailed prompt performs better. A short clip also needs enough time for each event to register.
A small experiment to keep
Start with one case and prepare the inputs it calls for. Record the model or endpoint, duration, resolution, aspect ratio, prompt, and seed when available. Decide what success means before generating: perhaps the subject stays recognizable, an object remains in the same hand, or the camera reaches a particular final position. Review the result, identify one important mismatch, and change one variable. Keep unsuccessful attempts with their settings. A history of revisions is more useful than a folder containing only the best-looking exports.
Following the evidence
We searched public X posts, inspected creator threads and replies, and looked for disclosed prompts, input materials, and workflow notes. We also reviewed sampled video frames to describe the visible action and composition. Each entry keeps its author and original source. Finding a compelling clip was only the beginning: we tried to separate what the creator actually disclosed from what an editor could reasonably infer. Unknown inputs stay unknown. A missing seed, an undisclosed reference image, or an unclear editing step should not quietly become an invented setting.
Three kinds of entry
An adapted prompt draws on instructions the creator made public, with a link to the original wording. A reconstructed prompt proposes an approach based on the visible result; it is not the recovered original. A workflow entry describes a process involving references, previsualization, editing, or related endpoints that cannot be reduced to one text instruction. This classification matters because the same final shot can result from different inputs and repeated attempts. Copying a reconstruction does not establish that the source clip was generated with that text.
Keeping the collection maintainable
The repository stores cases as structured JSON and generates the individual pages and both READMEs from those records. Automated checks cover duplicate cases, required metadata, local references, and stale generated files. That helps preserve attribution when the presentation changes, but a successful check does not validate the creative output. The large-preview layout lets people scan the shots before opening long instructions. There is also the browser-based prompt gallery, a separate website selection; its contents are not identical to the GitHub collection.
Match the workflow to the model
H3, H3 Max, and related endpoints must be distinguished. A case using several reference assets or an existing-video editing step should not be treated as proof that every H3 Max interface supports those inputs. The repository includes model-boundary notes for that reason. If you try a brief in the video creation workspace, compare the selected model and input mode with the entry first. Parameters in an editorial breakdown are suggested starting points unless the creator explicitly published them; they are not hidden facts recovered from the video.
Ownership and the limits of this release
The collection is maintained by the independent H3Max project I maintain. It is a commercial project, and the collection can be read and downloaded independently without a website account. The original repository contributions use MIT; linked creator videos, thumbnails, and source material retain their existing rights. The previews are the creators' outputs, not independent reproductions using our edited prompts. The launch collection is explicitly marked as not independently generation-tested, and sampled-frame inspection does not verify sound or every frame of a finished clip.
The contribution I would find most useful
The next improvement is a reproducibility record rather than a larger headline count. A useful contribution would pair a result with its inputs, model, settings, attempt count, revisions, and any editing performed afterward. Failed attempts deserve space too. They can show which constraints were difficult and which changes did not help. If you try one of these entries, a documented comparison or a correction to the source attribution would teach the next reader more than an unqualified claim that the prompt works.
