Thanks for considering a contribution. This is a small project; the workflow below keeps the bar honest and low-friction.
Open a GitHub issue using the Bug template. Include:
- Your Blender version (
bpy.app.version_string) and OS - The exact prompt you sent + the skill that triggered (or didn't)
- Whether the failure was numerical (Python error, file not created) or visual (render looks wrong)
- For visual bugs: attach the rendered output
The validation pattern documented in docs/test-results/ shows how the project's tester→patcher loop works. Bug reports that follow that pattern get acted on faster.
Open a GitHub issue using the Feature template. Useful new content:
- New scene-class validation (e.g. character, vehicle, kitchen) — surfaces recipe gaps
- New material recipe (e.g. Neon glow, anodized aluminum) — extends
blender-materials - New dimension reference for a common subject — extends
references/common-object-dimensions.md - Cross-version compat report from a Blender 4.x install
- Trigger-eval queries that surfaced real false-positive / false-negative behaviour
Welcome but not required. If you do open one:
- Branch from
main - Keep the change scoped to a single concern
- Update
CHANGELOG.mdwith the version it lands in - If the change affects skill behaviour, add an
evals/evals.jsonentry covering the new trigger / no-trigger boundary - If the change adds a new scene class, add the proof render to
plugin/skills/text-to-blender/assets/v<version>-validation/— including failure-state renders if the path was non-trivial. Honest validation > cherry-picked best result.
Every version of this project so far has shipped with explicit "what doesn't work" documentation. Maintain that. If a change has trade-offs or known limits, surface them in the PR description and the relevant skill's failure-modes table.
Semantic Versioning. Pre-1.0 (unstable) was through v0.9.x; v1.x is the stable line. Breaking changes warrant a major bump; recipe additions are minor; doc-only fixes are patch.
By contributing, you agree your contribution is licensed under the project's MIT license.