agentleFS
MAX-786MAX-786/claude-3d-harness/CLAUDE.md

claude-3d-harness

CLAUDE.md6 starsChanged 57 days agoMITView on GitHub

A harness for building 3D scenes in Blender with Claude Code. The 3D skills live in library/: five skill libraries, vendored, security-reviewed and checksummed. Four ship an open license at their origin; library/kb's origin has none, so its MIT grant was given directly to the maintainer. Around them sit the registry that routes between them, the workflows, the complexity profiles and the install tooling. Blender is driven through one MCP server named blender. Invoke the blender-harness skill before doing anything…

The file

We split this CLAUDE.md into the rules it gives an agent

6 rules. About 0 are common, 0 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈0

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Read only the library files the load plan names.

    line 22
  2. Very specific:

    Every SKILL.md in library/ is cataloged.

    line 51
  3. Very specific:

    Every file in library/ matches library/SHA256SUMS.

    line 52
  4. Very specific:

    Every library has an entry in registry/libraries.yaml and ships its LICENSE.

    line 55
  5. Very specific:

    Run verify after any edit to library/, registry/, workflows/, SKILL.md or .claude-plugin/, and claude plugin validate .

    line 66
  6. Very specific:

    After editing scripts/harness.py, run the tests: uv run --with pytest --with pyyaml pytest -q.

    line 68

≈ “Common” is an estimate: at least 1 in 5 of the 28,029 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
CLAUDE.md

claude-3d-harness

A harness for building 3D scenes in Blender with Claude Code. The 3D skills live in library/: five skill libraries, vendored, security-reviewed and checksummed. Four ship an open license at their origin; library/kb's origin has none, so its MIT grant was given directly to the maintainer. Around them sit the registry that routes between them, the workflows, the complexity profiles and the install tooling. Blender is driven through one MCP server named blender.

When the user asks for anything 3D

Invoke the blender-harness skill before doing anything else, including for requests that never mention Blender ("make a product shot of a watch"). It points to the entry skill, SKILL.md at the repository root, which classifies the job, picks a workflow and gets a load plan from the registry:

uv run scripts/harness.py resolve -w <workflow> -p <fast|standard|cinematic>

Read only the library files the load plan names. Do not pick skills by browsing library/: the libraries overlap and use different MCP tool names. The registry already made those decisions.

When maintaining the harness

PathWhat it holds
library/<key>/The skills, one folder per library, each with its LICENSE. library/SHA256SUMS records every file as reviewed
registry/libraries.yamlThe libraries: origin, imported commit, license, MCP dialect, what was omitted or changed
registry/skills.yamlEvery SKILL.md in library/, as active, chained or excluded
registry/capabilities.yamlCapability to provider routing, with fallbacks and variants
registry/profiles.yamlfast, standard, cinematic: always-on capabilities and budgets
registry/mcp.yamlThe MCP providers, the pinned release, the dialect translation tables
workflows/*.yamlOrdered stages per job type
orchestrator/*.mdClassifier, workflow and skill selection rules, QA loop
scripts/harness.pyThe engine: setup, doctor, verify, resolve, outdated, audit, checksums
tests/Tests for the engine. pytest.ini keeps pytest out of library/, where collecting a file would run it
.github/workflows/ci.yml (three OSes, plus the extension install into Blender 4.2 on Linux) and the weekly mcp-watch.yml
docs/security-review.mdWhat the library was reviewed for, what was found, what was changed. Add to it when a library or a risky change comes in
notes/lessons.mdLessons from real jobs. resolve prints the ones that concern a load plan's skills, so keep the entry form the file describes
SKILL.mdThe entry skill. Installed as a plugin, it is the /claude-3d-harness command
.claude-plugin/Plugin and marketplace manifests; bump version in plugin.json to release, and move the "Unreleased" notes in CHANGELOG.md

Invariants, all enforced by uv run scripts/harness.py verify:

  • One provider per capability (or per variant). Alternatives go in fallbacks.
  • One MCP server, named blender. .mcp.json is generated from registry/mcp.yaml (harness.py mcp-config --write); do not edit it by hand.
  • Every SKILL.md in library/ is cataloged. Uncataloged files are drift.
  • Every file in library/ matches library/SHA256SUMS. After editing a skill: uv run scripts/harness.py audit --changed, read what it flags, then uv run scripts/harness.py checksums --write, and commit both together.
  • Every library has an entry in registry/libraries.yaml and ships its LICENSE. Only vendor what a license or a permission lets you copy; a library with license: permission needs a license_note, which verify prints.
  • library/ is never edited during a 3D job; lessons go to notes/lessons.md. Improving a skill is a separate change, reviewed like code, because Claude follows these files with code-execution rights inside Blender.
  • The entry skill stays at the root as SKILL.md, and there is no root skills/ directory. That layout makes the plugin's command /claude-3d-harness; the plugin, marketplace and skill names must match. The plugin reuses the root .mcp.json.

Run verify after any edit to library/, registry/, workflows/, SKILL.md or .claude-plugin/, and claude plugin validate . after editing the manifests. After editing scripts/harness.py, run the tests: uv run --with pytest --with pyyaml pytest -q. CI runs verify --strict and the tests on Linux, macOS and Windows. The library follows nobody's branch: to take an improvement from a library's origin, port it by hand as a skill change (never re-import a folder, which would undo the security fixes).

library/kb's origin ships no LICENSE file. It is vendored MIT on the author's statement, given directly to the maintainer, that the work is MIT licensed (library/kb/LICENSE carries that text under the author's name). Do not copy anything else from that origin (its blender-toolkit/ is another author's work), and never copy from any source without a published license or a clear grant that allows it.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.