manim-slides

name	manim-slides
description	Template skill for creating Manim Slides decks. Discover local paths and wrappers first, then run the knowledge-first generation pipeline.

Manim Slides Skill (Template)

This is a template. Do not assume fixed absolute paths.

Path Discovery (Required First)

Find repo root, then use repo-relative paths:

• scripts/skill-runtime/run-manim-slides.sh
• skills/manim-slides/references
• .manim-slides

Suggested discovery commands:

pwd
rg --files | rg '^scripts/skill-runtime/run-manim-slides.sh$|^skills/manim-slides/references/|^pyproject.toml$'
find ~ -type f -path '*/scripts/skill-runtime/run-manim-slides.sh' 2>/dev/null | head

Update Local Skill Paths

After discovery, update your local/private skill copy only if your runtime differs from repo-relative defaults. Keep this Git-tracked template generic.

Mandatory Phase Order

1. reverse-knowledge-tree
2. manim-code-patterns
3. visual-planner
4. verbose-prompt-builder
5. code synthesis
6. hard-gated render checks

Reference Loading

Load required docs from skills/manim-slides/references/ in this order:

1. reverse-knowledge-tree.md
2. manim-code-patterns.md
3. benchmark-motifs.md
4. visual-planner.md
5. verbose-prompt-format.md
6. manim-slides-api-cheatsheet.md

Runtime

Preferred wrapper flow:

scripts/skill-runtime/run-manim-slides.sh render /abs/path/slides.py Deck
scripts/skill-runtime/run-manim-slides.sh present Deck
scripts/skill-runtime/run-manim-slides.sh convert Deck /abs/path/output.html

Fallback:

uvx --from "manim-slides[manim]" manim-slides --help

Cache Artifacts

Per topic slug under .manim-slides/<slug>/:

• knowledge_tree.json
• concept_plan.json
• visual_plan.json
• verbose_prompt.md
• slides.py
• render_report.json

Default cache behavior:

• Reuse existing artifacts when present.
• Regenerate only when user asks for rebuild/regenerate-fresh.

Phase Protocol

Phase 0: Preprocess

1. Normalize user topic and audience.
2. Compute deterministic slug.
3. Resolve absolute artifact paths.
4. Load existing artifacts if present.

Phase 1: Reverse Knowledge Tree (Subskill A)

Use reverse-knowledge-tree.md.

Requirements:

• Build DAG with target as root.
• Foundation baseline: high-school-level concepts.
• Limits: max_depth=4, max_prerequisites=5.
• Mark foundation concepts.
• Produce topological order from foundations to target.
• Write knowledge_tree.json.

Phase 2: Visual/Math Planning (Subskill B)

Use manim-code-patterns.md and manim-slides-api-cheatsheet.md.

Requirements:

• Map each topo-sorted concept to one scene intent.
• Enforce raw-string LaTeX rules.
• Define consistent palette/typography/spacing policy.
• Choose animation motifs and transitions per concept.
• Write concept_plan.json.

Phase 3: Visual Planner (Subskill C)

Use visual-planner.md and manim-slides-api-cheatsheet.md.

Requirements:

• For each concept, propose 2-3 candidate visual explanations.
• Decompose complex mechanisms into atomic visual beats.
  • Example: FlashAttention tiling should break into Q/K/V chunk creation, tile interaction, partial accumulation, and output matrix assembly.
• Score candidates with explicit rubric dimensions:
  • concept faithfulness
  • explanatory clarity
  • cognitive load
  • temporal coherence
• Iteratively refine until selected plan reaches threshold (>= 8.0/10) or max 3 refinement passes.
• Keep a score trace for each iteration.
• Write visual_plan.json.

Phase 4: Verbose Prompt Assembly (Subskill D)

Use verbose-prompt-format.md.

Requirements:

• Generate complete scene-by-scene verbose prompt.
• Source scene directives from selected entries in visual_plan.json.
• Include timestamps, animation ordering, transition hooks.
• Include explicit math notation and positioning details.
• Include color and pacing constraints.
• Write verbose_prompt.md.

Phase 5: Code Synthesis

Generate slides.py only from verbose_prompt.md + API constraints.

Requirements:

• Use Slide/ThreeDSlide where appropriate.
• Use next_slide() boundaries for presentation control.
• Keep one concept per slide.
• Avoid text walls.
• Ensure code is deterministic and rerunnable.

Phase 6: Hard-Gated Renderability

Final output must be blocked until renderability passes or missing dependency is explicitly diagnosed.

Required checks:

1. Python syntax sanity on slides.py.
2. Manim Slides command sanity using wrapper.
3. Render attempt for target scene.

Failure behavior:

• Capture failing command + stderr.
• Patch only relevant block.
• Retry in bounded loop.
• On dependency failure, stop with explicit missing dependency diagnosis.

Write render_report.json with:

• status: passed|failed|blocked_missing_dependency
• attempts
• failed_command
• errors
• last_updated_at

Visual Quality Rubric

Enforce all items:

• One concept per slide.
• Consistent palette and typography.
• Incremental reveals over dense text.
• Explicit transition intent for each slide.
• Math labels aligned to narration and timing.

Final Response Format

Return:

1. Absolute artifact paths for all 6 artifacts.
2. Rerunnable commands (render, present, convert).
3. Hard-gate pass/fail summary from render_report.json.
4. Short notes on what was reused from cache vs rebuilt.