Working with the agent

Matemium’s agent is a production collaborator. It can inspect project files, propose edits, work through roadmap phases, compile scenes, and respond to render evidence. It should not be treated as an unquestionable source of facts.

Give intent, not vague approval

Useful direction describes the viewer’s experience:

Keep eliminated values visible but subdued so the viewer can see why the search space shrinks.

Weak direction gives no evaluable goal:

Make it better.

When answering a project question, explain the reason behind a preference when that reason should influence later decisions.

Use project questions

The agent may offer:

A recommended choice
Alternative production choices
A custom response
Decide for me

Choose Decide for me only when the decision does not express important subject knowledge, tone, or creative intent.

Review edits

Agent-generated code and prose remain your responsibility. Before accepting a substantial change:

Read the explanation of what changed.
Inspect the relevant artifact or source diff.
Check whether the change respects the current roadmap phase.
Render when the change affects visuals.
Verify factual and subject-matter accuracy yourself.

Report visual evidence precisely

Good repair requests identify:

The visible object
The moment or phase
The observed defect
The intended result

Example:

During the first camera focus, the right comparison plot clips against the portrait boundary. Reduce the focus zoom enough to keep the entire plot and border visible.

Avoid requesting a large rewrite before confirming the actual cause.

Provider choice

External AI uses your selected provider account. Different models vary in:

Code-editing reliability
Long-context planning
Visual reasoning
Cost and speed
Tool-use behavior

Use a capable model for roadmap decisions and nontrivial repair. A smaller or local model may be sufficient for short edits and explanations.

Sensitive material

When you invoke cloud AI, the relevant prompt and project excerpts can be sent to the provider you selected. Do not include secrets or material you are not authorized to process.

Rendering and full media files remain local. The cloud agent does not need your rendered video uploaded as a normal part of rendering.

When to work manually

Edit scenes.py directly when:

The change is small and obvious
You know the exact API call
You want deterministic control
The agent repeatedly misunderstands a visual requirement

Agent and manual edits use the same project source; they are not separate modes of authorship.