NEWTON: Command for Claude Code

You are helping the user run the Newton physics-video generation loop. Follow each step below.

Step 1 — Pre-flight checks

Verify the environment is ready:

1. Check that .env exists in the project root. If not, tell the user to run /setup first.
2. Scan .env for placeholder values (lines still containing <your-). Warn if any required keys are missing:
   • PLANNER_API_KEY and PLANNER_BASE_URL — required
   • SEEDANCE_API_KEY and SEEDANCE_HOST — required
   • Others are optional but enable more tools (tell the user which tools will be unavailable).
3. Check that the newton conda environment is active. If not, activate it.

Step 2 — Scenario input

Ask the user: "What physics scenario do you want to generate?"

Encourage them to describe a concrete physical interaction — objects, materials, and what happens. Give a few examples to inspire them:

• "A glass marble rolls down a wooden ramp and launches off the end"
• "A water balloon drops onto a table and bursts"
• "A pendulum swings and knocks over a row of dominoes"

Once they provide a description, help refine it if needed — the prompt should be specific about:

• Object types and materials
• Initial positions and motions
• The key physical interaction (collision, deformation, fluid flow, etc.)
• Expected outcome

Confirm the final scenario text with the user before proceeding.

Step 3 — Configure run options

Ask the user if they want to adjust any settings, or use defaults:

Flag	Default	Description
--max-turns	8	Maximum planner/executer/verifier rounds
--duration	5	Generated video length in seconds
--out-dir	outputs/loop	Output directory
--baseline	(none)	Path to an existing baseline mp4 to skip baseline generation

Most users should use the defaults. Only ask about these options once — don't repeat.

Step 4 — Run the loop

Execute the inference loop:

python loop/run_loop.py "<scenario>" --max-turns <N> --duration <D> --out-dir <dir>