# 3dviz-pro-max: an agent skill whose capture helper makes the model look at the frames it rendered

> 3dviz-pro-max is an MIT-licensed agent skill for creative 3D visualization, shipping as a Claude Code and Codex plugin with 223 recipes, 440 knowledge records, 22 proved kits and 37 runnable studies. Its central mechanism is a capture helper that returns rendered frames to the agent so it refines from observation, and its published counts are recomputed from the catalog at build time so they cannot drift.

**viettranx/3dviz-pro-max** — Agent skill for creative 3D visualization: turn an idea into a Three.js/Blender scene worth exploring. Claude Code + Codex plugin, 223 recipes, 440 knowledge records, 22 proved kits, 37 runnable studies.

- Repository: https://github.com/viettranx/3dviz-pro-max
- Website: https://3dviz.dev
- Stars: 639 · Forks: 82
- Language: JavaScript
- License: MIT
- Published: 2026-09-17 · Updated: 2026-09-17 · Language: en
- Canonical page: https://hysenlabs.com/projects/viettranx-3dviz-pro-max

## The capture helper closes the loop, so the agent sees what it rendered

The mechanism that separates this from a prompt collection is one component. Alongside the workflow, the recipe catalog and the runnable templates, the skill ships a capture helper whose stated purpose is to let the agent inspect the frames it rendered instead of describing them. That is the whole argument: a model asked to improve a scene will otherwise optimise its own description of the scene, and the description is exactly the thing that is already wrong. Feeding back actual pixels turns refinement into an observation loop. The instructions reinforce it by saying the agent should use the existing project stack, choose an appropriate representation and refine the scene from observed results, rather than specify a renderer in advance. There is a second rule in the same spirit, and it is about claims rather than pixels: factual explanations need sources, while animated decoration does not become a simulation just by moving. That is a precise and unusual line to draw, because the failure it prevents is the specific one where a decorative wave function gets narrated as fluid dynamics.

## The interface is one sentence, and there is no renderer dropdown

The documented way in is a single sentence. Ask it to build a small fantasy village you can explore, choose a distinctive art direction, add a few moving creatures, and inspect the result. What the author explicitly tells you not to do is choose a renderer or fill out a specification first. That is a deliberate interface decision for an agent skill rather than a human tool, and it has consequences. You are trusting the agent to pick a representation appropriate to the subject, which means the output quality depends on the catalog's coverage of your subject rather than on your ability to configure it. The catalogue is what absorbs that responsibility, and it is large: recipes across all twenty-four registered directions with at least six per direction, plus a set of knowledge records and a set of proved kit blueprints. What you give up in exchange is control over the specific technique. What the framing assumes is that you know what you want to look at but not how it should be built, which is usually the right way round for the fantasy-village-shaped problems this targets.

## Every published number is recomputed from the catalog at build time

The most transferable practice in this repository is one sentence in the shipping table. The landing page recomputes every count from the skills, docs, evaluation and evidence directories at build time, so a number cannot drift from the data behind it. Given how many skill repositories quote statistics in a readme and then quietly stop updating them, that is worth more than any individual feature. The counts themselves are read from a catalog summary file, and as of the tenth of September 2026 they read as follows: 223 recipes across twenty-four directions with at least six each, 440 knowledge records across eighteen populated families, 22 kit blueprints, 37 style profiles including five named looks, 451 sources backed by 3,599 evidence events in a public archive, 37 example studies across five runnable chapters, 21 gallery frames, and 249 retrieval cases passing. The audit trail is unusually visible too, with an evidence directory in the tree and a public archive referenced from the readme. One number is incomplete: 249 cases pass but the total number of authored cases is never stated.

## The flagship village frames come from a retired revision, and the readme says so

The gallery section carries the disclosure most projects would leave out. The frames are runtime captures on a real GPU from the skill's own kits, rigs and recipes, and all twenty-one of them carry a host, a tier and a hash in the gallery. Then it says which ones came from where: the village frames come from a retired pilot revision at mixed tiers, while the kit frames use the blueprint proof harness at the tier named for each. Two sentences follow that function as a quality standard: nothing is cropped or retouched, and no frame claims a tier its record calls missing. A separate note above the table goes further, describing an author-supplied project recorded before this skill existed, with a Vietnamese-language interface, as visual inspiration rather than a benchmark, historical footage rather than an English demo or a runtime test of the skill. That is a project saying plainly which of its images are not evidence, which is the difference between a gallery and a marketing page.

## Twenty-two kits times four captures is eighty-eight, and the gallery shows twenty-one

The arithmetic of the proof system is worth doing, because it shows the ratio between verification and publication. Each kit blueprint is described as a record plus a module plus four real captures, re-checked by hash, so twenty-two kits imply eighty-eight captured frames existing somewhere in the evidence base. The public gallery holds twenty-one. So roughly a quarter of the verification output is published, which is a reasonable editorial choice but also means the gallery is a curated subset rather than the complete record. The same pattern appears in the retrieval evaluation, where 249 authored known-intent cases pass and the total is not given, so you cannot compute a pass rate. Both numbers are the kind of thing a sceptical reader will notice immediately, and both are recoverable from the evidence archive if you want to do the work yourself. That is the point of having the archive public: the summary table is a convenience, not the only source of truth.

## Twenty-four directions, and half of them are a mathematics curriculum

Read the direction list as a syllabus rather than as a mood board and the project's real shape appears. Environments, characters and motion art, products and assemblies, architecture and interiors, operations and logistics are the creative half. The other half is a linear algebra course rendered as geometry: basis maps, determinants and volume, the cross product, eigenvectors, the singular value decomposition, orthogonal projection, and then a single-qubit Bloch ball, two-qubit Bell correlations, attention aggregation and small-network activations. Calculus and fields adds directional derivatives, double integrals, divergence and flux, field integration, curl and circulation, Hessian critical points, finite Fourier synthesis, Mandelbrot and quadratic Julia escape orbits, and predator-prey phase time. Probability contributes covariance ellipsoids, Markov chains and Monte Carlo volume; topology contributes a Moebius band, torus cells, trefoil diagrams, mesh orientability and the Riemann sphere. Nobody calls that creative visualization in a job description, which is what makes the pairing interesting.

## Numeric defaults are mandatory for styles and optional for knowledge records

Two authoring conventions live side by side in the catalog and they disagree about how concrete a record has to be. Of the thirty-seven style profiles, thirty-three carry numeric defaults, about nine in ten, which is the right ratio for anything that sets a look. Of the four hundred and forty knowledge records, sixty-nine carry authored numeric defaults, roughly one in six. So most knowledge records are prose and style profiles are parameters, which makes sense if the knowledge records are explanations and the profiles are settings. What it means for a user is that the catalog has two different levels of machine-readability and you cannot tell which one a given record is without looking. The knowledge records are further split by population, with eighteen families populated out of a larger taxonomy, and sixty-nine of them authored with numbers rather than inherited. The style profiles also carry a handful of named looks, five of them, which are the presets you would actually reach for when starting from the one-sentence brief.

## It is a skill definition with an evidence base, not an application, and it ships for two agents

The tree tells you what kind of thing this is. There are two plugin directories at the root, one for Claude Code and one for Codex, and the skill itself is a markdown file in a skills directory. Alongside them sit documentation, an evaluation directory, a public evidence archive, an examples directory, a schemas directory, scripts, a site directory and tests. There is no application in that list, and the language is recorded as JavaScript because the runnable studies and the capture helper are browser code. Thirty-seven example studies live across five runnable chapters, and the examples directory contains one subdirectory per study alongside an index page and an asset manifest. The operational shape is worth stating plainly for anyone deciding whether to use it: this is a prompt-and-catalog project with a verification apparatus attached, published under MIT with no releases and a single open issue against 620 stars. The one open issue on that many stars is either very good maintenance or an unmonitored tracker.

## Conclusion

3dviz-pro-max is interesting for a reason that has nothing to do with the pictures: it closes the loop on pixels. Most agent skills for visual work let the model describe a scene and trust that the description matches what rendered. This one ships a capture helper so the agent inspects the frames it produced and refines from what it saw, and it applies the same discipline to claims, requiring sources for factual explanations while refusing to let animated decoration count as a simulation. The counterweight is that the discipline is uneven in places. The catalog's published counts are recomputed at build time, which is the strongest anti-drift practice here, yet the gallery of twenty-one captured frames sits alongside a proof harness that captures four frames per kit across twenty-two kits, and the flagship village imagery is disclosed as coming from a superseded revision. The stated purpose is creative visualization and a good half of the shipped directions are linear algebra, calculus and topology, which tells you what the author actually wanted to build. Before you adopt it, confirm three things: whether the retrieval pass rate means anything without the denominator, since 249 cases pass and the total is not stated, whether the catalog's academic bias matches the scenes you want, and whether you can drive it from a one-sentence brief rather than a specification, since that is the interface it is designed around.

## FAQ

### What does 3dviz-pro-max do?

It is an agent skill for creative 3D visualization that takes a one-sentence brief, picks a representation itself, and builds a Three.js or Blender scene you can explore. You are told not to choose a renderer or fill out a specification first; the agent uses your existing project stack and refines the scene from what it observes.

### How does 3dviz-pro-max know a scene is finished?

It ships a capture helper so the agent inspects the frames it rendered instead of describing them, which turns refinement into an observation loop. The skill also draws a line between claims: factual explanations need sources, and animated decoration does not become a simulation just by moving.

### What is in the 3dviz-pro-max catalog?

223 recipes across 24 registered directions with at least six each, 440 knowledge records in 18 populated families of which 69 carry authored numeric defaults, 22 kit blueprints each verified by hash, and 37 style profiles of which 33 carry numeric defaults including five named looks. It also ships 451 sources, 37 runnable example studies and 249 passing retrieval cases.

### Which tools does 3dviz-pro-max work with?

It ships as a plugin for two coding agents, with separate plugin directories at the root for Claude Code and for Codex, and the skill itself is a markdown file. The rendered output targets Three.js and Blender, and the studies are browser code captured headlessly in Chromium at 1920 by 1080 on a GPU.

### Can I see evidence that 3dviz-pro-max renders what it claims?

There is a public evidence archive and a frame gallery of 21 images, each carrying its host, tier and hash, with a note that nothing is cropped or retouched and no frame claims a tier its record calls missing. The gallery is a subset, since each of the 22 kits is verified with four captures and the readme also discloses that the village frames come from a retired revision.

## Sources

- [Issues](https://github.com/viettranx/3dviz-pro-max/issues)
- [License: MIT](https://github.com/viettranx/3dviz-pro-max/blob/main/LICENSE)
- [Project website](https://3dviz.dev)
- [README](https://github.com/viettranx/3dviz-pro-max/blob/main/README.md)
- [viettranx/3dviz-pro-max on GitHub](https://github.com/viettranx/3dviz-pro-max)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/viettranx-3dviz-pro-max