M2M2 is not trying to make the rendered MP4 itself the main editing surface. The current design makes the video editable by keeping the pipeline artifacts that produced it: prompt, concept intent, prerequisite graph, curriculum, math packet, storyboard, scene spec, generated Manim code, validation report, render result, review report, and manifest.
- Intermediate artifacts are already the editing contract. A deterministic run
writes inspectable JSON and
generated_scene.pyintoruns/<run_id>/, so a reviewer can see which prompt, storyboard, scene spec, and code produced the output. - Prompt-level editing is supported by rerunning the CLI with a revised prompt, style, quality, or render option.
- Spec/code-level editing is the near-term workflow: edit
scene_spec.jsonorgenerated_scene.py, then rerun validation and render against the same run bundle. - Static validation is already the gate before rendering. Failed validation should stop before Manim runs.
- Render repair is part of the architecture: failed renders can feed Manim
stderr/stdout plus the frozen
scene_specback into the code repair path without recomputing the whole planning chain. - Hermes and Codex are the agent-assisted iteration layer. Hermes can inspect artifacts, patch prompts/specs/code, run CLI smoke checks, review frames or GIFs, and keep the edit history explicit; Codex can be used for generated-code and repair stages when the local Codex CLI provider is selected.
prompt
-> typed planning artifacts
-> storyboard / scene_spec edits
-> generated_scene.py edits or Codex-assisted repair
-> static validation
-> Manim rerender
-> video review artifacts
-> next edit
The important product direction is artifact-first editing:
- Edit the prompt when the explanation goal, audience, style, or duration is wrong.
- Edit the storyboard or
scene_spec.jsonwhen the teaching sequence or visual objects are wrong. - Edit
generated_scene.pywhen Manim implementation details are wrong. - Use the repair loop when validation or rendering fails.
- Use Hermes/Codex to make those edits repeatable instead of treating each video as a one-off render.
- A full browser video editor timeline.
- Direct MP4 frame-by-frame editing.
- A large UI for dragging Manim objects on a canvas.
Those may come later, but the first useful editable workflow is the one that is already aligned with M2M2's architecture: preserve the artifacts, edit the stage where the problem was introduced, and rerender.