Threejs agent skills: 24 graphics systems and deliberately no API reference

The pack exists to skip the API reference

The README pre-emptively rules out the thing most people would want from a three.js pack. It says this is not a three.js API cheat sheet, and that it skips basic three-dimensional production fundamentals on the grounds that any decent model already has that internal knowledge, along with three.js API technicalities, for which the reader is pointed at the documentation or at existing API-oriented skills. The argument behind that choice is the more interesting part. You cannot simply prompt the agent for good graphics and expect it to produce them; the agent needs to see the exact implementation. What the pack offers is described as the vocabulary of sophisticated graphics implementations, and as a skill pack with an attached example library, meant to teach not just what to do but exactly how.

Every effect has to survive with post-processing switched off

The section titled operating model sets out six things every graphics system in the pack is expected to expose, and the last one is the hard constraint. Deterministic or reproducible inputs come first, then named controlling fields and perceptual parameters, then diagnostic outputs, then scale, distance and temporal stability rules, then an intentional mechanism-backed quality or resolution tier when the system defines one at all. The sixth is a no-post baseline that still reads. That means a shader, a scattering model or a water surface has to look correct on its own before any bloom, grading or exposure pass is applied, which rules out the common shortcut of authoring for the final composite.

MIT AND GPL-3.0-only, with provenance files in the tarball

The manifest declares the licence as the SPDX pair MIT AND GPL-3.0-only, which is the correct way to signal that the pack combines material under two terms rather than choosing one. The published file list backs that up. Alongside the code directories it ships the GPL-3.0 licence text, a third-party notices file, and a trace manifest. Shipping a trace manifest inside the npm package rather than only in the repository is the detail that matters: a consumer who installs the pack gets the record of where the borrowed material came from, without needing to clone anything. The author field is a single name, Scott, and the package is public.

One file in that provenance directory is deliberately left out

The published file list ends with an exclusion. `source_materials/README.md` is listed with a leading negation, so the README that would explain the provenance directory is not shipped to npm, while the licence text, the notices and the manifest are. That is a reasonable packaging decision, since the exclusion keeps the tarball smaller, and it also means an installed copy cannot answer why a given file is in the pack. The repository root does hold the source_materials directory in full. A consumer auditing the licence position therefore needs the repository rather than the installed copy, and the two artefacts do not carry the same information.

Six test stages, and one of them tests agent routing

The test script is a chain of six, not a single runner: validate, then routing, then references, then installer, then assets, then a geometry kit. Each maps to a named script, so the pack tests its own installer, its reference examples, the origins of its example assets, and the geometry quality assets belonging to one skill. The unusual one is routing, exercised by a dedicated script, which for a pack of prompt material is the load-bearing check: it verifies that a given request selects the intended skill rather than a neighbouring one. The geometry stage reaches inside a skill directory rather than the shared scripts, which suggests quality is asserted per-skill where the artefacts exist.

Five install targets in the commands, seven in the keywords

Installation is a single binary invoked with npx, taking an agent name:

npx threejs-awesome-graphics-agent-skills@latest install --agent codex
npx threejs-awesome-graphics-agent-skills@latest install --agent claude-code
npx threejs-awesome-graphics-agent-skills@latest install --agent cursor
npx threejs-awesome-graphics-agent-skills@latest install --agent github-copilot --scope project
npx threejs-awesome-graphics-agent-skills@latest install --agent custom --path designated-agent-skills-dir

A user scope and a project scope are both offered, and a custom agent is handled by pointing at a directory. The keyword list in the manifest names seven agent identifiers against the five in these commands, adding gemini-cli and windsurf, neither of which appears in an example. A table of supported targets follows, with user scope under `~/.agents/skills`, so the package is designed around a shared directory layout rather than per-agent conventions.

The reinstall command and the universal project path both stop early

Two places in the document end before their content does, and both matter to a reader trying to follow along. The install section's final entry is introduced by a comment about forcing a reinstall of the exact currently installed version, but the command under it stops partway through the package name and never reaches the install verb. Separately, the supported-targets table gives the universal target a user scope of `~/.agents/skills` and then a project scope beginning with `.ag`, which is where the text ends. The consequence is that the reinstall path and the universal project directory are both documented by implication rather than by a complete line.

Twenty-four skills and one time claim with nothing behind it

The skills table has twenty-four rows, and the expertise cells are written in a shared compressed vocabulary that is itself part of the product: Jacobian foam, submerged Snell windows, total internal reflection, hybrid FFT and Gerstner water, finite-volume beach waves, curved-ray and null-geodesic integration, traversable wormhole throats with transported observer frames, GTAO-style horizon sampling, encoded luminance metering, asymmetric adaptation, generated three-dimensional lookup tables, deformable granular sand with mass-coupled grain shading. One skill is described as validated FFT synthesis, which is the only quality adjective applied to a method. The single performance claim in the document is also the vaguest: a realistic ocean of the kind shown would have taken hours if not days to create and finetune, and is now available out of the box. No measurement accompanies it.

Editorial conclusion

This pack suits someone building a three.js scene or game who has an agent that can read files and who wants worked implementations of expensive effects rather than another reference. It does not suit someone looking for API coverage, for the pack states outright that it skips fundamentals and technicalities on the assumption the model already holds them. Check three things first. Read the operating model, because the pack commits every effect to looking acceptable with post-processing disabled, and that constraint will rule out effects you had in mind. Read the third-party notices and the trace manifest before redistributing anything. And check whether your agent is one of the install targets, since the command-line examples and the package keywords do not name the same set.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. scottstts/Threejs-Awesome-Graphics-Agent-Skills on GitHub