Straightedge

Checked SVG figures and Manim animations for technical explanations.

Explore all figures and videos →

Straightedge turns structured intent into deterministic visuals. It provides two independent output lanes:

Both lanes are designed around the same constraint: a visual can render successfully and still be wrong. Straightedge validates inputs before drawing and exposes geometry and findings that callers can use to reject or repair visible defects.

Install

From a checkout:

python3 -m pip install -e .

SVG figures use only the Python standard library. To build and render animations:

python3 -m pip install -e '.[render]'

Make an SVG figure

from pathlib import Path

from straightedge.diagrams import render_diagram

svg = render_diagram(
    {
        "type": "unit_circle",
        "params": {"angle": 45, "show_triangle": True},
    }
)
Path("unit-circle.svg").write_text(svg, encoding="utf-8")

render_diagram() needs no browser, network, or headless renderer. An unknown diagram type returns an empty string so a missing optional figure does not abort an entire document build.

The registry currently contains 35 templates across several domains:

• Math and data: function graphs, coordinate planes, Riemann sums, unit circles, polar graphs, matrices, step functions, heatmaps, and tables.
• Computer science: binary trees, linked lists, stacks, queues, hash tables, call stacks, dynamic-programming tables, architecture diagrams, and state machines.
• Projects and business: Gantt charts, work-breakdown structures, project networks, timelines, flow diagrams, and T-accounts.

Inspect straightedge.diagrams.DIAGRAM_REGISTRY for the exact registered names. Each renderer accepts a compact, serializable hint and returns a complete SVG string.

Make an animation

Every shipped animation is reachable by name, in any language, with no LLM — name a template and render it:

straightedge list-templates                                   # what exists
straightedge render --template calculus/derivative_tangent    # the hero animation, in English
straightedge render --template conic/ellipse_foci --qc        # and check the frame
straightedge render --template calculus/riemann_integral \
  --params '{"expression": "x**2 + 1"}'                        # refine with parameters

--template takes any id from list-templates and skips the keyword router entirely — it is how the animations in the gallery above are drawn.

A formula is another language-neutral path to the deterministic scenes:

straightedge render "y=x^2-4*x+3" --language en

scaffold writes the scene without rendering; render streams Manim's progress and prints the final media path — the usual low-quality output is media/videos/scene/480p15/GeneratedScene.mp4.

The formula parser accepts y= and f(x)=, the variable x, arithmetic, implicit multiplication, powers, common constants, and common elementary functions. It validates expressions against a strict allowlist before generating code.

Useful render controls:

# Vertical composition for short-form video
python3 -m straightedge.cli render "y=sin(x)" --aspect 9:16

# Match scene beats to externally produced narration
python3 -m straightedge.cli render "y=sin(x)" --beat-seconds beats.json

# Choose a Manim quality preset and media root
python3 -m straightedge.cli render "y=sin(x)" --quality m --media-dir build/media

--language {en,zh} controls on-screen labels; English is the default. --aspect {16:9,9:16} changes both the composition frame and pixel resolution. Beat files map IDs to durations, for example {"b01": 2.4, "b02": 3.1}.

What gets checked

The checks are deliberately usable without Manim. straightedge/qc.py works against plain geometry values, so callers can apply the same policy to both figures and scenes.

• Preconditions reject malformed or unsupported structured input.
• Diagram tests reject blank output and verify that meaningful data marks were drawn.
• Scene builders report overlaps, off-screen content, untranslated labels, and other visible risks as structured findings.
• Example simulations assert their mathematical or systems claim before they animate it.

The gallery labels those standalone dataflow examples separately because they are written by hand and do not use Straightedge's prompt pipeline. Their checks are useful demonstrations, not generated-library output.

Narration-driven timing

Hand the renderer the measured length of each narration clip and every step runs for exactly as long as the sentence spoken over it:

straightedge render "riemann sum of x squared" --beat-seconds beats.json

{ "b01": 3.4, "b02": 5.1, "b03": 2.8 }

Straightedge does not synthesise speech — durations arrive as data, so the same scene renders identically from a cloud TTS clip, a local model, or a human recording, offline and without an API key. A step with no measurement keeps the timing it was written with. See docs/narration-timing.md for the walkthrough, the two pacing helpers, and the silent failure worth knowing about.

Prompt-driven scenes

For concepts outside the deterministic templates, straightedge/agent/ provides a writer, reviewer, executor, and bounded repair loop against an OpenAI-compatible API. See docs/agent-design.md for the design.

⚠️ This lane runs model-written Python. The generated scene is syntax-checked, scanned for disallowed imports and interpreter escapes, reviewed, and executed with a timeout — but that is defence in depth, not a sandbox. An allowlist over an AST is not a security boundary. Run the agent lane in a container or VM whenever the prompt or the model is untrusted. The deterministic template lane (render, --template) and the figure lane do not execute model output and carry no such caveat.

export OPENAI_API_KEY="..."

# Run it isolated when the input or model is not fully trusted:
docker run --rm --network=none -v "$PWD/out:/out" straightedge-render \
  agent-render "Show why the focal-distance sum of an ellipse is constant" \
  --language en --output-dir /out

# …or directly, only when you trust the prompt and the model:
python3 -m straightedge.cli agent-render \
  "Show why the focal-distance sum of an ellipse is constant" --language en

Language and voice adapters

The figure renderer, geometry checks, scene builders, and English output do not depend on Chinese input. The first natural-language teaching adapter was built for Chinese-speaking teachers, so its keyword planner and optional local Whisper transcription remain useful value-adds in the repository. They are one input adapter, not Straightedge's product boundary.

python3 -m straightedge.cli scaffold \
  "用单位圆展示正弦函数" \
  --language en

Audio transcription is local-only and opt-in:

python3 -m straightedge.cli plan --audio lesson.wav

Development

python3 -m pip install -e '.[dev]'
python3 -m pytest -q

The gallery is a static GitHub Pages site under site/, published at https://scimigo.github.io/straightedge/. It intentionally keeps the library-generated visuals separate from the hand-written, assertion- backed examples.

Related open-source work

• ManimCommunity/manim provides the animation engine.
• makefinks/manim-generator inspired the writer/reviewer/retry shape; Straightedge's agent implementation is written from scratch.
• ManimCommunity/manim-voiceover is a natural future integration point for narration synchronization.

License

MIT