harleycoops/hermeslearnsmanim/hermes-learns-manim

Hermes Learns Manim

Use this skill when the user wants a mathematical, scientific, or newsworthy concept turned into a Manim animation.

Hermes is the only agent harness for this workflow. Do not introduce another agent SDK.

Required routine

1. Call initialize_run_workspace with the user's request.
2. If the topic is a current event, run the news-researcher subagent first to ground the narrative in verified facts and canonical citations.
3. Use Hermes subagents to rebuild the reasoning tree and downstream artifacts.
4. Save every artifact back through MCP tools (save_text_artifact, save_json_artifact) so the workspace stays canonical.
5. Validate (validate_manim_scene, estimate_scene_complexity, validate_latex_block) before rendering.
6. Render only if the user explicitly wants video output, and start at quality l for iteration.

Recommended subagent split

Delegate only when it helps materially. Default split:

• news-researcher (optional): gather verified facts and citations for newsworthy topics; write research.md
• concept-analyst: fill analysis.json
• prerequisite-explorer: build knowledge_tree.json
• math-enricher: tighten equations, definitions, and numerical values inside the tree or notes
• narrative-composer: write narrative.md as a scene-by-scene cinematic plan
• code-generator: produce scene.py and call validators before handing back

Keep artifact ownership clear so subagents do not overwrite each other casually.

Artifact contract

The MCP workspace seeds these files:

• request.md
• workflow.md
• analysis.json
• knowledge_tree.json
• narrative.md
• scene.py
• run_manifest.json

Additional artifacts the coordinator may write:

• research.md — for current-events topics (facts, dates, numbers, sources)
• validation.json — merged output of the validators

Keep all of these in sync as the workflow progresses.

Cinematic 3D defaults

When the user asks for a "3D", "cinematic", "best-in-class", or current-events animation, the code-generator should default to:

• ThreeDScene with set_camera_orientation(phi=65-75°, theta=-45° to -55°)
• Dark space background (e.g. #050510) with a Dot starfield for depth cues
• Ambient camera rotation at 0.08-0.15 rad/s during wide shots
• A coherent color language declared at the top of the file (title / accent / data / warning)
• add_fixed_in_frame_mobjects for all titles, equations, captions
• Labeled, color-coded key terms (MathTex parts) so the viewer can follow symbolically
• A 3-to-6 act structure: hook → concept → math → key insight → counterfactual → payoff
• Raw strings (r"...") for every MathTex and Tex literal
• Self-contained: no ImageMobject or external asset loads

Validation gate

Before rendering:

1. validate_manim_scene on the full scene.py source
2. estimate_scene_complexity — if complexity == "high", confirm the user still wants to render
3. validate_latex_block on any equation the narrative marks as load-bearing

Fix issues in scene.py and re-save through MCP before rendering.

Current-events protocol

For news-driven topics the coordinator MUST:

• Anchor the animation in verifiable facts (dates, distances, speeds, crew names)
• Record sources in research.md with URLs and a last-checked timestamp
• Use honest numbers on screen (e.g. 252{,}756\ \text{mi}) and preserve units
• Never invent quotes, statements, or unconfirmed mission outcomes

Provider note

Hermes handles login and provider selection with hermes setup and hermes model. This skill is provider-agnostic.

References

• references/mcp-config.yaml
• references/subagents.md