What does this skill do, and when should you use it?
The CAD skill is one of 13 skills in the earthtojake/text-to-cad repository, covering 3D modeling of mechanical parts and assemblies. It instructs the agent to write decorated Python model scripts (build123d-based) that a pinned cadgen CLI builds locally into STEP, with maintainable STL, 3MF and GLB mesh outputs. The workflow also covers project organization, resolving geometry references from prompts, measuring and validating geometry, reviewing snapshots of its own work, and opening existing STEP/STP, STL, 3MF and GLB files in a local CAD Viewer. The repository is MIT-licensed and runs everything through uv, downloading its runtime on first use.
- Reads/edits decorated build123d Python model scripts and runs python <model>.py to regenerate STEP outputs
- Exports meshes via @stl/@threemf/@glb decorators or one-off cadgen stl/3mf/glb build commands
- Resolves prompt references like /path/assembly.step#o1.2.f7 with read_scene + scene.resolve() and measures properties such as face area
- Writes Python geometry checks for dimensions, clearances and topology, and renders snapshots via cadgen step snapshot for self-review
- Shows models to the user with cad_show or a local CAD Viewer started with cadgen viewer --detach
- Diagnoses install/kernel issues with cadgen doctor and explains unexpected rebuilds with cadgen store why
- A mechanical engineer asks the agent to produce a parametric STEP bracket from a verbal description or an annotated image
- A robotics team organizes a multi-model CAD project (src/ layout, model catalog, formatted output folders) and maintains assemblies
- A 3D printing user one-off exports an existing STEP part to STL or 3MF
- An agent receives a CAD Viewer Quick Edit note (file path, references and a marked-up sketch) and edits the corresponding geometry
- A reviewer verifies dimensions and clearances via snapshots and geometry checks before handing off to the user
- Users needing 2D DXF drawings — SKILL.md explicitly defers these to the $dxf skill
- Users needing URDF/SRDF/SDF robot description files — the skill requires the corresponding robot-description skills
- Offline-restricted environments without uv — the first run must download the runtime and a headless browser
How do you install this skill?
- Static review only; no commands were executed. All capability claims rest on documentation.
- cadgen sends anonymous usage stats and crash reports (PostHog) by default; run `uvx cadgen telemetry off` or set CADGEN_TELEMETRY=0 if unwanted.
- The CAD Viewer local server is unauthenticated (loopback only); do not bind a non-127.0.0.1 host or expose it beyond the local machine.
- First run downloads runtime and a headless browser from the internet; in mainland-China networks uv/PyPI/browser downloads may be unreachable or slow, with no mirror guidance provided.
- No Chinese-language support; Windows 11 Smart App Control blocks the unsigned OCP native module — disable it or use WSL.
- Publisher is unverified; the repository contains 13 skills and this score covers only skills/cad/SKILL.md.
- Shell / CLI
- Network access
- Local filesystem
uv / uvxPython 3.13 (uv-managed)cadgen 0.7.20 (pinned via uvx)
This skill ships with the text-to-cad collection (13 skills; this one lives at skills/cad/SKILL.md). uv must be installed first.
Claude Code
claude plugin marketplace add earthtojake/text-to-cad#latest
claude plugin install text-to-cad@earthtojakeCursor
git clone --depth 1 --branch latest https://github.com/earthtojake/text-to-cad ~/.cursor/plugins/local/text-to-cadGrok Build
grok plugin install earthtojake/text-to-cad@latest --trust
grok plugin enable text-to-cadGemini
gemini extensions install https://github.com/earthtojake/text-to-cad --ref latest --consent --auto-updateOther agents (Skills CLI)
npx skills add earthtojake/text-to-cad#latestClaude Desktop has no skill-install command; only the MCP server config (cadgen mcp, pinned cadgen==0.7.20) is documented, added manually to the app config, then restart.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Create a 40mm wide, 20mm deep, 6mm thick bracket and export it as STEP and STL
- Export the assembly in /work/robot/STEP/assembly.step to GLB so I can view it in the browser
- Here are the dimensions from this sketch — build a parametric model and state your assumptions
- Open the viewer and show me the bracket.step you just generated, then check the hole spacing against the drawing
Trigger it by describing a modeling task; the skill routes to task-specific reference docs. The core flow: find or create a decorated Python model script (e.g. src/bracket.py with @step(out="../STEP/bracket.step")), edit and run:
python src/bracket.pyBoth python and cadgen are pinned uvx invocations (--python 3.13 --from cadgen==0.7.20). One-off mesh exports:
cadgen stl build STEP/bracket.step STL/bracket.stl
cadgen 3mf build STEP/bracket.step 3MF/bracket.3mf
cadgen glb build STEP/bracket.step GLB/bracket.glbGeometry checks are hand-written Python (no inspect CLI; exploratory checks in tmp/, reusable ones in checks/). After creating or changing geometry you must render and personally read at least one snapshot:
cadgen step snapshot STEP/bracket.step tmp/review.pngShow results with the cad_show tool if available; otherwise run cadgen viewer --host 127.0.0.1 -- --detach (always pass --detach), read url from the JSON line, and return url?file=<URL-encoded absolute path> per file. Assemblies call child models and place results with .moved() or Location * shape; search $step-parts for named purchasable parts before making placeholders.
What are this skill's strengths and limitations?
- Fully local workflow — model files never leave the machine; MIT license
- STEP-first output with maintainable STL/3MF/GLB exports covering mainstream downstream uses
- Complete documented workflow: modeling, project layout, reference resolution, validation, snapshot review, viewer handoff, repair loop
- uv-pinned cadgen==0.7.20 and Python 3.13, with cadgen doctor for environment diagnostics
- Hard dependency on uv and a first-run network download (runtime plus a headless browser) — not viable offline on first use
- The collection sends usage stats and crash reports by default (opt-out available), plus a daily version-check request
- On Windows 11 with Smart App Control enabled, every cadgen command fails due to the unsigned OCP native module; requires disabling it or using WSL
- The skill prescribes process, not correctness — geometry accuracy depends on the model code, and snapshot/validation review is performed by the agent itself with no independent guarantee
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| CAD (text-to-cad's CAD skill) this page | 55 · Use with care | ★ 19k | 1d ago | MIT |
| DfAM Check — 3D Printability Analysis | 67 · Recommended | ★ 19k | 1d ago | MIT |
| 3D Morph PPT Skill | 48 · Use with care | ★ 32k | 5d ago | Apache-2.0 |
| Academic Figure Prompt — Modern ML Airy Style | 68 · Recommended | ★ 2.6k | 6mo ago | MIT |
| Banana Claude — Gemini Image Creative Director | 59 · Recommended | ★ 1.1k | 1mo ago | MIT |
The README positions this skill within a 13-skill library: it complements step.parts (off-the-shelf STEP parts), DXF (2D drawings), URDF/SRDF/SDF (robot descriptions) and manufacturing skills like DfM/SendCutSend/G-code; this skill owns 3D parametric modeling and export itself. The source names no third-party competitors.
How did FollowSkills review this skill?
Evidence: the skill operates purely on local files; no credential access, covert exfiltration or destructive defaults; versions pinned (cadgen==0.7.20, uv --no-config); telemetry on by default but fully disclosed with opt-out (README); viewer binds loopback and SECURITY.md explicitly states the unauthenticated trust boundary. Deductions: telemetry sends by default, local viewer is unauthenticated, first run downloads runtime and a headless browser from the internet, publisher identity unverified, rollback/recovery only partly described. Not full marks.
Documentation is highly self-consistent: explicit command contracts, doctor diagnostics, migration guides, and well-designed failure feedback (teaching errors, named refusals). Deductions: static review cap of 10; key paths (cadgen build/snapshot/viewer) have no executed reproduction evidence; test suite not present in this evidence.
The task table clearly covers modeling/export/inspection/animation/migration, and boundaries are thoroughly declared (closed-loop linkages out of scope, mesh-only limits, etc.). Deductions: core function depends on uv/PyPI downloads, ffmpeg and network first-run, with no statement on mainland-China reachability; no Chinese-language support; semantic trigger precision is only weakly evidenced.
Layered docs (SKILL.md plus references) with progressive disclosure, migration guides, MIT license, version pinning and doctor checks are all strong. Deductions: no standalone changelog/version history for the skill itself, maintenance responsibility only implied via the repository README, unverified publisher, and several referenced reference files are absent from this evidence.
Goals, methods and alternatives (script vs step build, snapshot vs viewer) are clearly argued; outputs are directly usable STEP/STL/3MF/GLB with mandatory self-review snapshots. Deductions: static review cap of 7; no executed representative output verifies direct usability; all benefit claims come from documentation only.
README shows CI test and PyPI version badges and links external migration docs — auditable primary material. Deductions: static review cap of 5; this evidence contains no actual test files or third-party execution results for the skill's key paths, and cross-source corroboration is thin.
Open a dimension to read why it scored that way
Evidence confidence:Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →