Four 1,024 by 1,536 layers become one editable card
RuiC Card Skill starts with artwork rather than a finished flat image. Its normal canvas is 1,024 by 1,536 pixels, split into subject, background, line art, and typography, with an optional effects layer. The agent writes a card configuration, validates transparency and alignment, builds a Blender scene, exports glTF geometry, and assembles a Three.js viewer. The delivery includes the .blend file, source layers, renders, browser assets, and a verification folder.
Depth is the useful trick. The subject and background receive different parallax values while typography stays fixed to the card edge. Blender builds the editable version; the browser reconstructs the material in GLSL because glTF cannot carry Blender's custom node graph. That gives you two editable representations with similar behavior, though the skill correctly warns that they are not pixel-identical. Four finishes and browser sliders make the result more than a looping foil animation.
The 37 MB base install hides the heavy production tools
Our installed environment used 37 MB after adding 35 packages. A real card also needs Python 3.9 or newer with Pillow, Node and npm, a Chromium-family browser, and Blender 4.5. The skill can fetch an official portable Blender into each project and check its SHA-256 before execution. Generated images, Blender binaries, model exports, and renders sit outside the small repository footprint.
The model requirement is just as important as the software list. A host must generate transparent artwork layers, inspect frames, notice registration errors, and rerun failed work. The scripts do not call one named image API or carry a credential. That avoids vendor lock-in, but it also means output quality depends on the agent and image tool you pair with the skill. A text-only model cannot complete the documented workflow.
What happened when we ran it
Our fresh Debian sandbox installed commit 712fc9f in 13 seconds on 3 CPUs with 8 GB of RAM and no secrets. Installation succeeded with 35 packages and a 37 MB disk footprint. The build step completed successfully in 3 seconds. Pip-audit reported 0 known vulnerabilities in the installed Python dependency set.
The checkout held 34 files, about 34,806 lines of source, and occupied 5.7 MB. There was no test script or target, so the lab skipped tests. We found 0 CI workflow files, no Dockerfile, and no tests directory. The repository does include standalone regression scripts, but our measured command did not run them. Our result covers installation and build mechanics, not card generation, Blender download, rendering, or the 33-step browser verification described by the project.
One stale-artifact bug can turn a failed build green
Open issue 10 identifies the most consequential defect. Blender 4.5 background mode can print a script traceback yet return exit code 0. The current pipeline checks that card.blend exists, so a file left by an earlier successful build passes that condition. Export then continues from stale geometry, and the run can appear successful even though the requested card was never rebuilt.
That problem deserves a release gate because batch users may not notice a stale hero image. The issue proposes checking modification times for card.blend and the exported GLB. Until such a guard lands, use a clean output directory or record timestamps before each run, then confirm that every requested artifact changed. The verifier exercises browser behavior; it cannot prove that the browser contains the newest requested artwork if the pipeline fed it an old model.
Two open verifier defects complicate the green report
Issue 9 shows that absent optional layers can cap browser verification at 32 of 33 checks. The verifier counts a 1 by 1 placeholder texture alongside real layers, then reports mismatched canvas sizes. That is a false failure rather than a broken card. Issue 8 affects an earlier stage: Blender release discovery uses a bare Python request that the download host can reject with HTTP 403, even though the package downloader sets a User-Agent.
Both reports include narrow fixes, but they remained open on October 1, 2026. The repository had 9 open issues and 7 open pull requests, with no closed issue or pull-request history returned by the API. Recent activity is real: the last push was September 30, and several reports include controlled reproductions. The missing part is maintainer throughput. Detailed outside contributions only help users after they are reviewed or folded into main.
Host independence still requires an agent that can see
The repository provides a standard SKILL.md, plain Python scripts, and a Node verifier rather than a host-specific plugin binary. Its full English README also documents installation without tying the skill to one vendor directory. An issue reports a successful 33 of 33 run under a different agent harness on headless Linux after environment fixes. That is credible evidence of portability across skill-capable hosts.
Portability has a clear floor: the agent must make images, inspect images, run local programs, keep a server alive, and return to failed steps with judgment. Claude Code users with those tools can get an unusual editable artifact from one instruction. Unattended card factories should wait. With no tagged release and the stale-output path still present, the safe use is one card at a time, in a clean directory, followed by the automated report and an actual look at both tilt directions.