Blender Pipeline

Start from the Codex client

This is the user-facing Codex launcher. If the requested output is unclear, ask whether the user wants only a tutorial or a complete Blender reconstruction; wait for their choice before launching. Do not ask again when the goal is explicit.

Read references/codex-launch.md, interpret the user's natural-language inputs, and invoke this skill's scripts/launch_from_codex.py yourself. It calls the same maintained pipeline as the API entrypoint. Tutorial-only stops after preparation; full reconstruction continues through knowledge retrieval → Blender generation → render and review.

A video with an accompanying .blend, preview, texture, or dependency bundle retains those inputs in either selected route; it is not an independent edit. A supplied Markdown tutorial skips video extraction: prepare it for tutorial-only, or continue knowledge-backed generation for reconstruction. No RW1 is required.

Use the user's requested output location or established data root. Ask only for missing information needed to run, not command-line flags or API credentials. The Codex path uses the signed-in Codex CLI; the user need not run a Python launcher manually. run_codex.py remains a compatibility entrypoint, not the client interaction. Requests to explain, inspect, or dry-run do not authorize an actual reconstruction.

Extract a tutorial from video

For tutorial-only, pass --extract-only and do not start reconstruction. Read tutorial-extraction/SKILL.md for illustrated Markdown, learner inputs and the separate rubric; both replay routes reuse it.

Reproduce a maintained showcase

Read reproduction/SKILL.md only for an explicitly requested catalog recipe or hybrid-delivery workflow. A tutorial's subject also appearing on the showcase does not change ordinary replay routing. Public previews do not authorize downloading the underlying assets.

Generate from zero

For ordinary video/tutorial requests use the client launcher above. Use the specialized generation mode for an explicit from-zero reference/text task or low-level generation work, with no authoritative starting project.

Read generation/SKILL.md, then load knowledge/generation.md, knowledge/evidence-and-routing.md, and only the relevant sections of knowledge/rendering-and-dynamics.md.

Edit an existing asset

Use editing for a standalone change request on an existing asset, without a video/tutorial reconstruction workflow. A generated replacement object within that scoped change remains part of the edit.

Read editing/SKILL.md, then load knowledge/editing.md and only the relevant sections of knowledge/rendering-and-dynamics.md.

Shared rules

• Keep mode-specific code, tests, manifests, and outputs separate. Share only stable utilities and contracts.
• Preserve verified source intent and non-target dimensions. A successful file write is not a visual or semantic pass.
• Use explicit input/output paths and environment-backed credentials; never embed workstation, mounted-volume, or server paths in committed code.
• Load knowledge/operations-and-knowledge.md only for orchestration, paid calls, knowledge maintenance, or publication.
• Start knowledge lookup at knowledge/index.md; do not load every reference for an ordinary single-mode task.

Package maintenance uses scripts/validate_package.py, not every ordinary launch.