ml-content

name

ml-content

description

Generate publication-grade ML explainer videos and carousels the way 3Blue1Brown actually builds them — in real manimGL (NOT Manim Community Edition), as a tiny domain DSL of self-arranging Mobjects choreographed into transform-driven beats where every motion carries meaning. Overlap is prevented at construction time, not policed after render. Use for: 3b1b-style ML videos, paper-figure animations, IG carousels, infographics, posters.

ml-content

Generate ML explainer content that looks like 3Blue1Brown, not like AI slop.

⚠️ PRIME DIRECTIVE — treat every video as nuclear. ZERO errors ship.

The content goes public to an audience that will fact-check it. A single wrong number, mislabeled quantity, or overstated claim destroys trust in everything else and gets screenshotted. So: nothing — not even slightly — may be wrong. Every spoken line, every on-screen number and label, every caption, and the thumbnail must be verified against the primary source before it is rendered, and audited again on the rendered video before it ships (§10, the non-negotiable gate). If you cannot cite the exact source line for a claim, you do not say it, write it, or put it on screen. Soften it or cut it. No "dramatic license" on numbers. When in doubt, it is wrong until proven right.

This skill was rebuilt from a full read of Grant Sanderson's actual production code (github.com/3b1b/videos, 503K LOC, 2015→2026) and the real manimGL engine (github.com/3b1b/manim). Every rule below is grounded in that source with file:line citations. Where this skill once guessed, it now measures.

The video engine is manimGL (the 3b1b version), driven by manimgl. Manim Community Edition (CE) is a different library with a different, incompatible API — code written for one crashes on the other. Static IG carousels use HTML/matplotlib (see the Carousel section); everything animated is manimGL.

---

0. Why the old output was slop (read this once)

The previous version of this skill produced overlapping elements, weak animation, no consistency, and infographic-feeling stills. The root causes, now fixed:

1. It shipped Manim CE code while preaching manimGL. The old template used from manim import *, MathTex, ThreeDScene, set_camera_orientation, set_fill_by_value, Create, begin_ambient_camera_rotation — none of which exist in manimGL (grep over the entire engine = 0 hits). It would not even run.
2. 100% of real output was actually built in matplotlib, hand-placing ~57 text calls + ~20 boxes per scene with absolute coordinates. That is the worst possible tool for animation: no relative layout, no transform system, no camera. Overlap is guaranteed.
3. It treated overlap as a validation problem (bbox asserts, frame validators, ffmpeg caption-pads) instead of a construction problem. 3b1b never validates overlap — it makes overlap structurally impossible by building self-arranging objects.
4. It treated motion as decoration. Elements faded in from nowhere as disconnected islands. In real 3b1b, objects are born from the thing they abstract (TransformFromCopy), so every motion teaches.
5. Planning was marketing copy with word-count targets. Real 3b1b planning is an ordered list of named teaching beats that reads top-to-bottom as the narration.

The fix is a different mental model, encoded as the Six Laws below.

---

1. The engine reality (the single most important section)

Start every manim file with:

from manim_imports_ext import *      # in the 3b1b/videos repo
# or, standalone:  from manimlib import *

Scene base class is InteractiveScene (interactive_scene.py:66) — for 2D and 3D. In the entire 2025 corpus: InteractiveScene subclassed 385 times, ThreeDScene (scene.py:930) 0 times — it exists but 3b1b never uses it. 3D is achieved on an InteractiveScene by moving self.frame.

The camera is self.frame (a CameraFrame mobject; scene.py:112). Local alias frame = self.frame appears 114× in 2025 code. self.camera.frame appears 0×.

Render / iterate:

manimgl file.py SceneName              # render
manimgl file.py SceneName -se 120      # drop into IPython at line 120 (the dev loop)
manimgl file.py SceneName -w           # write to file
# inside the embed: checkpoint_paste() runs clipboard code with checkpoint rewind

CE landmines — Table A: these genuinely CRASH on manimGL (absent symbols)

You must NOT emit (CE)	Use instead (manimGL)	Evidence
from manim import *	from manim_imports_ext import * / from manimlib import *	CLAUDE.md:59 (loads the wrong library)
MathTex(...)	Tex(...)	grep MathTex over repo = 0; CLAUDE.md:82
self.set_camera_orientation(phi=,theta=,zoom=)	self.frame.reorient(theta, phi, gamma, center, height)	absent; camera_frame.py:172
self.move_camera(...)	self.play(self.frame.animate.reorient(...))	absent
self.begin_ambient_camera_rotation()	self.frame.add_ambient_rotation(1 * DEG)	absent; camera_frame.py:212
self.add_fixed_in_frame_mobjects(m)	m.fix_in_frame()	absent (no scene-level helper)
Create(m)	ShowCreation(m)	absent; creation.py:48
Unwrite(m)	Uncreate(m) or FadeOut(m)	absent
surface.set_fill_by_value(...)	surface.set_color(c, opacity) / set_color_by_xyz_func(...)	absent
Circumscribe(m)	FlashAround(m)	absent in engine (the one truly-missing indicator)
Wiggle(m) (class)	WiggleOutThenIn(m) or rate_func=wiggle	no Wiggle class; indication.py:355
ease_in_out_*, smoothstep, easeOutCubic	the 15 real rate funcs (§7)	not in rate_functions.py
FadeInFrom, FadeOutAndShift, SpinInFromNothing, AddTextLetterByLetter	FadeIn(m, shift=, scale=), AddTextWordByWord	absent

Self-check (must return nothing):

grep -nE 'MathTex|from manim import \*|set_camera_orientation|begin_ambient_camera_rotation|move_camera|add_fixed_in_frame_mobjects|set_fill_by_value|\bCreate\(|\bCircumscribe\b|\bWiggle\(|\bUnwrite\(|FadeInFrom|SpinInFromNothing|AddTextLetterByLetter' your_file.py

Table B: these RUN on manimGL but are wrong/stale style — don't emit anyway

Avoid (valid but off-style)	Prefer	Why
class S(Scene) / class S(ThreeDScene)	class S(InteractiveScene)	both exist & run, but 2025 corpus is 385× InteractiveScene, 0× ThreeDScene; 3D uses self.frame
TransformMatchingTex(a, b)	TransformMatchingStrings(a, b)	TransformMatchingTex exists (subclasses Strings) but 3b1b uses Strings
eq.set_color_by_tex(tok, c)	Tex(R"...", t2c={tok: c}) or inline eq[tok].set_color(c)	set_color_by_tex is a real Tex method (tex_mobject.py:207) but ~unused; inline set_color dominates
Indicate(m) / CircleIndicate(m)	FlashAround(m) / Flash / FlashUnder	both exist & are used (indication.py:73,142); modern corpus just reaches for Flash* far more
DEGREES	DEG	DEGREES is a live alias of DEG (won't crash); 2025 uses DEG ~520× vs DEGREES 3×

The split matters: don't let a self-check grep reject valid 3b1b code. Indicate, TransformMatchingTex, set_color_by_tex, and DEGREES are style calls, not crashers — only Table A NameErrors.

---

2. The mental model (what makes it 3b1b, not slop)

You are not laying out a frame. You are building a small cast of self-arranging objects and transforming them through a sequence of beats where each motion is the explanation.

Two layers, every video:

1. helpers.py — the domain DSL. 4–10 Mobject subclasses that build and arrange themselves (WeightMatrix, NumericEmbedding, EmbeddingArray, Dial, NeuralNetwork, ContextAnimation — _2024/transformers/helpers.py), plus 2–5 show_*(scene, ...) choreography verbs. Define the cast before writing a single construct().
2. Scene files — thin InteractiveScene subclasses whose construct() is a flat list of # Beat name comments, each assembling DSL objects and animating them.

Everything below serves this model.

---

3. The Six Laws (non-negotiable)

Law 1 — Relative layout only. Overlap is structural, not policed.

Real 3b1b layout is overwhelmingly relative: next_to used ~10,500×, arrange ~2,460× across the repo, versus only a handful of absolute move_to([x,y,z]) content placements — and nearly all of those absolutes are the camera, never content. Forbid absolute content coordinates. move_to([x,y,z]) and frame.animate.move_to(...) are for the camera/light source only.

Allowed positioning primitives — content position ALWAYS comes from one of these:

a.next_to(b, DOWN, buff=MED_LARGE_BUFF)     # relative to another object
a.next_to(b, RIGHT).match_y(c)              # compound: x from b, y aligned to c
a.align_to(b, LEFT)                         # share an edge
a.to_edge(UP, buff=LARGE_BUFF)              # to a frame edge (chrome only)
a.to_corner(UL)                             # to a corner (chrome only)
VGroup(*items).arrange(DOWN, buff=MED_SMALL_BUFF, aligned_edge=LEFT)
VGroup(*items).arrange_in_grid(rows, cols, buff=...)
dots = Dot().get_grid(n_rows, n_cols, buff_ratio=0.5)
a.match_x(b) / a.match_y(b) / a.match_width(b) / a.match_height(b)

Boxes and pills are content-sized, never hand-sized:

rect = SurroundingRectangle(label, buff=SMALL_BUFF)   # measures real glyph bounds
brace = Brace(group, DOWN); brace.get_text("12,288")   # fits the span, anchors at tip
under = Underline(word)                                 # width derived from content

SurroundingRectangle is used ~1,400× in the repo and cannot clip or mis-center because its size = target.get_shape() + 2*buff (shape_matchers.py:22). Delete every text_width() character-count heuristic — it has no analog in real code.

The buff ladder is the ONLY source of gaps (default_config.yml:109):

SMALL_BUFF=0.1   MED_SMALL_BUFF=0.25   MED_LARGE_BUFF=0.5   LARGE_BUFF=1.0

Never write buff=0.37. Consistent margins across scenes come for free from this ladder.

Build the group, then place the group. Assemble a VGroup declaratively, .arrange() its internals once, then position the whole thing. This is the dominant pattern across the repo's thousands of VGroups. Never position leaf elements against screen coordinates.

Labels on moving targets follow live: label.always.next_to(target, UP, buff=SMALL_BUFF) (or add_updater). A single-frame bbox assert can't catch a mid-animation collision; a re-running next_to never collides.

Because of Law 1, the old "Five-Layer Defense" (bbox asserts + frame validator + ffmpeg caption pad) is retired. Overlap is prevented at construction. Keep at most a light visual probe-frame check for taste and timing, not collision.

Law 2 — Motion carries meaning. Born-from, never spawn.

The highest-frequency, highest-leverage technique in the corpus: a key object enters by transforming from the concrete thing it abstracts, so the causal link is visible.

# the DALL·E image literally dissolves INTO the numeric vector entries (attention.py:128)
self.play(LaggedStart(*(bake_mobject_into_vector_entries(img, vec) for img, vec in ...)))
# 12,288 numbers collapse into a single symbol E_n  (attention.py:200)
self.play(FadeTransform(entry, sym))
# a formula term is delivered by morphing the data column it denotes (ml_basics.py:642)
self.play(TransformFromCopy(data_column, x_symbols))

Transform-family hard counts (attention.py): TransformFromCopy 46 > FadeTransform 31 > ReplacementTransform 6 > plain Transform ~5 ; TransformMatchingTex 0.

• TransformFromCopy(src, dst) is THE workhorse — source persists, a copy morphs to the destination, so the viewer sees "this becomes that" while "this" is still there.
• For equation retitles, TransformMatchingStrings (not ...Tex).
• Never FadeIn a load-bearing object from nothing. Decorative scaffolding can fade in; the thing the lesson is about must be born from its referent.

time_span=(start, end) choreographs a reveal inside ONE self.play (used 638× in the repo) so a camera move and a multi-part reveal cascade together instead of firing simultaneously:

# the softmax aha — one play, run_time=3, cascaded (attention.py:921)
self.play(
    self.frame.animate.reorient(...),
    GrowArrow(arrow, time_span=(1, 2)),
    FadeIn(label, time_span=(1, 2)),
    TransformFromCopy(ndp_col, softmax_col, time_span=(1.5, 3)),
    run_time=3,
)

generate_target / MoveToTarget (127 refs) is how dozens of objects snap into a new arrangement with zero hand-keyed coordinates:

grp.target = grp.generate_target()
grp.target.arrange(RIGHT, buff=0.15)   # mutate the target with normal layout calls
grp.target.scale(0.65).next_to(anchor, DOWN)
self.play(MoveToTarget(grp))

Law 3 — A tiny domain DSL gives consistency.

Every video defines ~6 custom Mobjects/Animations in a helpers.py plus ~6 scene-local get_*/show_* factories. A vocabulary, not a framework. This is the mechanism for both consistency and no-overlap (the objects arrange themselves).

For ML content, don't start from scratch — vendor 3b1b's transformer DSL (NumericEmbedding, WeightMatrix, EmbeddingArray, ContextAnimation, value_to_color). It is the single fastest path to the authentic "columns of real numbers" look. See §13.

Mobject-subclass recipe (model: Dial, helpers.py:655; MachineWithDials:761):

class WidgetThing(VGroup):
    def __init__(self, value=0, ...):
        # 1. build sub-parts
        body = Rectangle(...); needle = Line(...)
        # 2. lay them out RELATIVELY (ratio buffers, never move_to([x,y,0]))
        ticks = Line(...).get_grid(1, n, buff_ratio=0.5)
        ticks.set_width(body.get_width() - SMALL_BUFF); ticks.move_to(body)
        # 3. assemble, 4. name every meaningful part
        super().__init__(body, ticks, needle)
        self.body, self.needle = body, needle
        # 5. a state mutator that recomputes geometry+style from the logical value
        self.set_value(value)
    def set_value(self, v):
        self.needle.put_start_and_end_on(self.get_center(), self._value_to_point(v))
        self.needle.set_color(value_to_color(v))     # color is a pure function of value
    def animate_set_value(self, v, **kw):            # 6. methods that RETURN animations
        return AnimationGroup(self.animate.set_value(v), ...)

show_*(scene, ...) choreography-verb recipe (model: show_matrix_vector_product, helpers.py:97):

• first arg is scene; the function calls scene.play/scene.wait internally,
• owns its transient highlights via a last_rects/to_fade accumulator so exactly one highlight is ever on screen,
• returns the persistent mobjects it created.

Variants are 2-line subclasses of a shared base overriding one class attribute — never a copy-pasted construct() (model HighlightEarthOrbit(NearestPlanets) with highlighted_orbit = 2, and its sibling HighlightMarsOrbit(NearestPlanets) with = 3; planets.py:2202). Domain numbers live in ALL-CAPS module constants with one conversion_factor (planets.py:4), so sizes can't disagree across scenes.

Large data uses honest ellipsis: render a finite set, swap one element for Tex(R"\dots") (or ellipses_row=-2, default in WeightMatrix), so big arrays read as big without overflowing (helpers.py:478, 620).

Law 4 — Earned 3D only, and 3D is never static.

3D is earned only when a quantity's dimensionality is the payload (a vector space, a surface f(x,y)=z, a volumetric field). Roughly half of even the attention video is intentionally flat. A flat data series in 3D is the #1 faux-3D tell — and the old matplotlib "isometric stack of parallelograms" cube is exactly the slop to never produce.

Production lesson (§13): 3D actively HURTS a stack of thin layers (interleaved attention layers, a residual tower) — viewed at an angle they collapse into one solid block and the per-layer colors vanish. Render those flat, face-on. Reserve 3D for clouds, surfaces, and collapsing volumes.

Counter-lesson — over-flattening is its own slop (learned shipping the DeepSeek/Qwen attention series). Apply "flat by default" too zealously and every scene becomes the same colored square grid. A viewer's note, verbatim: "it just seems like a single block matrix throughout the whole video, we're lacking elements." Flatness is honest, but monotony kills retention, and the cure is variety + rigor, not decoration:

• Rotate the visual primitive across the reel. Don't reuse the row-of-cells / n×n grid in more than ~half the scenes. Across ~10 beats budget at least: one numeric-matrix scene, one on-screen equation, one earned-3D scene — plus vectors / a plot / a code view.
• Earned-3D is earned more often than "half flat" implies — use it. Three patterns that shipped clean (and read as not faux-3D because dimensionality IS the payload):
  • A memory/field as a z=f(x,y) surface that DEFORMS under its update. e.g. the gated-delta-rule state: forget = scale heights down, erase = remove a bump, write = raise a new bump. srf = ThreeDAxes((-3,3),(-3,3),(0,3)).get_graph(lambda x,y: sum_of_gaussians(x,y,bumps)); srf.set_opacity(0.7); srf.always_sort_to_camera(self.camera) then Transform(srf, make_surf(new_bumps)) between states; frame.add_ambient_rotation(0.5*DEG) to orbit.
  • A multi-dimensional tensor as a Cube volume that grows or collapses. The KV cache is tokens×layers×heads (grow it: cube.animate.set_width(w, stretch=True, about_edge=LEFT)); attention compute is L×L×d collapsing to L×k×d (set_width(0.8, stretch=True)). box3d(w,h,d)=Cube(opacity=0.3).set_width/height/depth(.., stretch=True).
  • An embedding space as 3D vectors / a GlowDots point cloud for relevance-as-alignment or nearest-neighbour selection.
  • A feature/representation geometry as a polytope of unit vectors (the superposition reel): Line(ORIGIN, v) rays + GlowDot tips + faint hull Lines, set_floor_plane('xz'), and escalate the dimension — show the 2-D cases face-on (antipodal pair, then a triangle in-plane), then reorient into 3-D for the tetrahedron with add_ambient_rotation. The directions in space ARE the payload (a genuine vector space), so it never reads as faux-3D. Sequence the corner readout swaps (D=\frac{2}{3} → D=\frac{3}{4}) with non-overlapping time_span or the speed-fit stretch garbles them.
• Show the REAL math, not abstract blocks — the single biggest density win. Replace a colored grid with the actual numbers + the actual equation:
  • Numeric matrices/vectors with exact, consistent values: WeightMatrix(values=...), NumericEmbedding(values=...) (compute the numbers so Δ = v − Sk literally checks out on a pause), then show_matrix_vector_product(scene, M, v) to animate o = S q row-by-row.
  • Animate a dot product accumulating pair-by-pair: for i: FlashAround(q[i]); FlashAround(k[i]); ChangeDecimalToValue(running_sum, partials[i]).
  • Pin the actual update/score equation as Tex (S_t = S_{t-1}\alpha_t(I-\beta_t k_t k_t^\top)+\beta_t k_t v_t^\top, I_{t,s}=\sum_h w_h\,\mathrm{ReLU}(q_h\cdot k_s)) with fix_in_frame() — but above CAP_FLOOR: a to_edge(DOWN) HUD equation lands inside the burned-caption band, so place it at y ≈ −2.4, not −3.4.

There is one camera verb. All five DOF specified together:

self.frame.reorient(theta_deg, phi_deg, gamma_deg, center_tuple, height)
#                   azimuth     polar    roll       look-at      zoom(=visible height)
# static set, or animated inside a play that ALSO moves content:
self.play(self.frame.animate.reorient(-21, 79, 0, (1.13, 0.35, 0.88), 3.81), run_time=5)

• Angles in degrees. reorient(0, 0, 0, center, height) is a flat pan/zoom used constantly even in 2D — the camera is the pointer.
• Camera args must look hand-tuned, not round. reorient(-178, 9, 178, (2.15, 1.12, 0.56), 6.84), never reorient(45, 60, 0). (Grant copies exact numbers out of interactive mode.)
• A 3D scene is never static. Either chain multiple reorients with multi-second run_times, or self.frame.add_ambient_rotation(0.5 * DEG) during holds. Call self.frame.clear_updaters() before a new scripted reorient so ambient + scripted motion don't compound.
• self.set_floor_plane('xz') when Y is up and Z is depth (makes hand-tuned angles compose intuitively).

3D legibility kit (from the holograms video, the 3D gold standard):

• Vectors/arrows in 3D get vector.always.set_perpendicular_to_camera(self.frame) so they don't foreshorten edge-on (geometry.py:798 — it's a Line method; every real call site applies it to a vector, never a Text). Text labels in 3D are kept readable a different way: either lay them into the plane (label.rotate(PI/2, RIGHT) + set_backstroke(BLACK, 3)) as spatial labels that move with the camera, or pin them as a HUD with fix_in_frame() (titles/equations/readouts). That two-layer split — spatial vectors face-camera, text labels either in-plane or fixed-in-frame — is the legibility rule.
• Light sources/sparks/dots are GlowDot/GlowDots/TrueDot/DotCloud (additive glow), never a shaded Sphere (a Lambert ball reads as a solid ball — faux-3D tell). A beam = ~50 stacked GlowDots of rising radius, ~3/n opacity.
• Coplanar stacks separate by tiny graded offsets along IN/OUT (1e-3…2e-2) to kill z-fighting.
• surface.always_sort_to_camera(self.camera) on any translucent surface that will be orbited.
• Real 3D solids use real depth (Cube().set_shape(w,h,depth)), never an extruded flat rect — so a side view doesn't collapse.
• The 3D should not be able to lie: drive the render from a real point array / real projected data (basis @ model[word]), so the picture is a computed consequence of the geometry.

Law 5 — The metronome (measured pacing).

These numbers are counted from the source, not guessed.

• play : wait ≈ 1.6 : 1. Stable across the corpus (attention 216/148, 2025 aggregate 2287/1414). ~3 animated actions per 2 holds. Flag scenes below 1.2 or above 2.0.
• The default hold is self.wait() with no argument = 1.0s, and it is 80% of all waits. Write the bare call. Don't emit wait(1) or wait(1.7).
• Timed waits come from a small rounded set: {0.5, 2, 3, 4, 5} for content; {8,10,12,15,20,30} only for ambient/3D holds. (One wait(1.5) and one wait(0.25) exist in the entire transformers corpus.)
• 95% of plays use the default run_time=1.0. Emit a bare self.play(...). Override only with purpose: run_time=2 is the deliberate-reveal workhorse (~2× the next value), 3–5 for important/complex reveals, 0.5 for a snappy punch.
• run_time ≥ 8 is exclusively camera moves / axis Rotate / slow enumerations, almost always rate_func=linear. Never on a text FadeIn.
• lag_ratio is an intent dial: 0.1 = standard gentle stagger (the default, ~2.3× the next value, usually with FadeIn); 0.25–0.5 = readable "enumerate, read each" cascade; 0.01–0.05 = near-simultaneous bulk reveal (matrix entries, dot fields); 1.0 = strict succession; 0 = simultaneous. Scale to count with lag_ratio=1.0/len(items) to hold total duration constant.
• rate_func is left default (smooth ease) ~97% of the time. The only overrides that earn their place: there_and_back (emphasis pulse returning to rest), linear (continuous camera/rotation), there_and_back_with_pause, rush_from (snappy entrance, often run_time=0.5). Never set rate_func unless the motion's meaning requires it.
• The authoring unit is the labeled beat: one # Capitalized beat comment ≈ one self.wait() (attention: 149 comments / 148 waits). Each beat = 1–2 plays then a wait. Avoid firing 5+ animations before a hold.
• self.add() is the silent reveal (1 instant add per ~3.5 plays): drop in scaffolding/re-layered elements with zero runtime; animate only the element the eye should follow.
• Ensembles always stagger. Many similar objects enter via LaggedStartMap(FadeIn, group, shift=0.25*DOWN, lag_ratio=0.1), never a simultaneous FadeIn(a), FadeIn(b) (that stacks — the #1 slop tell). Nest LaggedStart + .shuffle() for a field of many things so timing isn't mechanical.

Law 6 — One color per concept; figure/ground discipline.

• Default everything WHITE on BLACK. Greys (GREY_A..E) for chrome/de-emphasis. Spend a saturated hue only on a load-bearing role, bound to a named variable reused everywhere (value_color = RED; Q=YELLOW, K=TEAL, V=RED, input=BLUE_B). Never flood the frame with color.
• Color is assigned inline per part (part.set_color(NAMED) — 226× in transformers) and bound to a concept; t2c= is fine at construction but used far less. Do not color by hand-counted glyph slices — use substring indexing eq["\\omega"].set_color(PINK) or t2c.
• Signed numeric grids use value_to_color (helpers.py:51): positive→blue (BLUE_E→BLUE_B), negative→red (RED_E→RED_B), magnitude→lightness, blended in HSL. Unsigned data (raw embeddings) uses magnitude-only GREY_C→WHITE — never a good/bad diverging scale on neutral data. Forbid hand-setting per-element colors on number grids.
• Real palette is manimGL's default_config.yml (see §8 for the full A–E ladder + correct hexes). YELLOW is pure #FFFF00, not the old skill's invented #F7D96F.

---

4. Workflow

User has...                          → Start at...
─────────────────────────────────────────────────────────
Topic/paper, no plan                 → Recon → Beat-sheet → DSL → Build → Ground
A beat-sheet                         → DSL → Build → Ground
A helpers.py + beat-sheet            → Build → Ground
A finished video                     → Grounding pass

Always ask what they have and what they need; don't assume the full pipeline.

1. AHA + Recon. One sentence: the shift in perspective the viewer leaves with. If it's a fact, not a shift, stop. Then a lean recon (see §5) — primary source + 2 independent writeups for anything recent; flag every claim DIRECT/INFERRED/UNCERTAIN.
2. Beat-sheet (§6) — the ordered named beats that read as the narration. This replaces the old word-count brainstorm.md.
3. Gate A — claims ledger (§10). Before writing scene code, verify every number/label/claim the beats will assert against the primary source. Fix or cut anything not CONFIRMED. Cheaper to fix on paper than in a rendered scene.
4. DSL — write helpers.py (and vendor tb_helpers.py, §13): the self-arranging Mobjects + show_* verbs the beat-sheet needs.
5. Build — thin InteractiveScene files, beats as # comments. The continuous-VO + speed-fit + captions pipeline is §13.
6. Render → read frames → fix loop (§13): every scene, every overlap/garble. 3–8 cycles.
7. Gate B — audit the rendered video (§10). Sample frames across the whole cut, read each, confirm on-screen numbers/labels/captions/thumbnail are all correct and clean. Nothing ships until this passes.
8. Caption + thumbnail + ship.

The render→critique→re-render loop still applies: render, sample 5–10 key frames (manimgl -w then ffmpeg -ss T -frames:v 1, or manim -ql -s --format=png for a scene's last frame), read each frame, fix, re-render. 3–8 cycles. But you are now checking taste, timing, and motion, not hunting collisions — Law 1 already prevents those.

---

5. Recon (lean)

For the paper/topic, produce concise working docs (no word-count targets — quality is whether they're usable, not long):

• paper-summary.md — title + arXiv ID + authors; per-claim DIRECT/INFERRED/UNCERTAIN; the thing studied; method; the numbers that matter; lineage; limitations; why-now; hook angles in tension.
• related-work.md — lineage tree, predecessor/competitor deltas, what it does and does NOT contribute.
• discussions.md — the social reception: camps, surfaces to cite, which hook will latch.
• moodboard.md — references mined from the audience's visual world (Distill.pub, Anthropic Circuits, 3b1b itself), each with one line on what to steal. For manim, the strongest moodboard is the relevant _2024/transformers/*.py scene itself — point at the exact scene that already solves a similar visual.

Discipline: never quote hard citation counts ("widely cited", not "348 citations"); avoid generic adjectives; flag every recent (≤6 mo) claim for re-verification at publish time. Use BrightData (mcp__brightdata__scrape_as_markdown, search_engine) for primary + 2 secondary sources; fall back to WebFetch/WebSearch. Large scrapes (a full paper/blog) overflow the tool result and save to a file, returning the path — don't try to read the whole thing inline; spawn a fact-extraction subagent with the path + a precise brief ("read all N lines in chunks; return verbatim quotes + exact numbers for these claims"), so the raw text stays out of your context. WebFetch's small model silently drops detail on long pages — prefer scrape-to-file + subagent when you need exact figures. Gate B tip: run a final adversarial-reviewer subagent — hand it the rendered video + the claims ledger and tell it to refute every on-screen number and recompute the arithmetic; it catches what your own frame-sweep rationalizes away.

---

6. The beat-sheet (replaces the old brainstorm.md)

The real 3b1b planning artifact is an ordered, numbered list of named beats that mirrors SCENES_IN_ORDER and the in-construct() section comments — it reads top-to-bottom as the spoken script. Model: ml_basics.py, attention.py section comments, supplements.py chapter banners.

beat-sheet.md format — one line per beat:

NN. BeatName — <imperative narration verb: Show/Ask/Label/Contrast/Highlight> + the single idea
     · operates-on: <which persistent object this beat moves>
     · [REVEAL]  (if this is THE aha: tag reorient + run_time≥2-6 + wait≥2 after)
     · objection: <predicted viewer confusion here> → <the interjection beat that resolves it>
     · caveat:    <if this beat simplifies, name the honesty beat that flags it>

Mandatory structure:

• ESTABLISH first: nest the topic in known categories (Transformer → Deep Learning → ML → "learn from data"), zooming down to the one concrete object the video manipulates. Never open on a labelled diagram or a definition. Open on the phenomenon (the crime scene).
• GAME-PLAN beat: an on-screen ordered checklist of the chapter's atomic teaching units.
• One persistent worked example, locked (the "fluffy blue creature" / "Michael Jordan → Basketball"), carried by the same mobjects from setup to payoff. If any beat introduces a NEW concrete example mid-thread, flag it.
• Every formula is grounded: the formula beat sits after its motivating concrete beats, and is immediately dismantled back to concrete arrays with color-coded symbol→object mapping (DescribeAttentionEquation, attention.py:1850).
• Motivate-by-violation for any normalizing op (softmax/normalization/masking): show the raw output failing the desired property first, then introduce the op as relief.
• Tag exactly one beat [REVEAL] and give it the reveal rhythm. Setup beats stay at default time.
• Objection + caveat columns: in-lesson, per-beat objection handling and scheduled honesty beats wherever you simplify (translate the recon's DIRECT/INFERRED/UNCERTAIN flags into on-screen caveats).
• RECAP beat mirrors the opening game-plan with units checked off; mark PREVIEW→PAYOFF pairs.

Self-check before building: read the beats top-to-bottom. Do they form a coherent spoken narration where (a) phenomenon precedes formalism, (b) one example carries throughout, (c) the AHA is tagged and paced, (d) every formula is grounded, (e) objections/caveats sit at their trigger points, (f) a recap mirrors the open? If it reads as a feature list, it fails.

---

7. manimGL grammar reference (the real catalog)

Animations that exist (use these by exact name)

• Reveal: FadeIn(m, shift=0.25*DOWN, scale=2) (most common), Write(m) (text/equations; default rate_func=linear, auto run_time), ShowCreation(m) (strokes/paths), DrawBorderThenFill(m), GrowFromCenter/GrowFromPoint/GrowArrow, FadeInFromPoint(m, point).
• Morph: TransformFromCopy(src, dst) ⟵ workhorse, FadeTransform(a, b), Transform(a, b), ReplacementTransform(a, b), TransformMatchingStrings(eq1, eq2, key_map={...}), FadeTransformPieces. Use path_arc=PI/2 for an arced morph (~1,200× in repo).
• Compose: LaggedStart(*anims, lag_ratio=0.1), LaggedStartMap(FadeIn, group, shift=, lag_ratio=), AnimationGroup(*anims), Succession. Schedule sub-anims with time_span=(a,b) and/or per-target .set_anim_args(run_time=, rate_func=).
• Emphasis (the real set): FlashAround(m, time_width=1.5, run_time=2), Flash(point), FlashUnder(m), VShowPassingFlash(path, time_width=1.5, run_time=4) (glowing pulse along a connection), FocusOn, WiggleOutThenIn.
• Numbers: ChangeDecimalToValue(dec, target), CountInFrom(dec, start) — dec must be a DecimalNumber/Integer. For bespoke motion, UpdateFromAlphaFunc(m, lambda m,a: ...) (the escape hatch Grant actually uses) over inventing Animation subclasses.
• Move/rotate: MoveToTarget(m) (+ m.generate_target()), Restore(m) (+ m.save_state()), MoveAlongPath, Rotate(m, PI) (discrete) vs Rotating(m, TAU, run_time=5, rate_func=linear) (continuous), m.animate.method(...).

The 15 rate functions (the ONLY ones that exist)

linear, smooth (default), rush_into, rush_from, slow_into, double_smooth, there_and_back, there_and_back_with_pause, running_start, overshoot, not_quite_there, wiggle, squish_rate_func(f,a,b), lingering, exponential_decay. smooth is a quintic smoothstep. squish_rate_func is a factory returning a rate func (windows an ease into part of an animation).

Text & math

• Tex(R"...") — pure math (raw strings always). TexText("words $math$ words") — prose with inline math. Text("plain", alignment="LEFT") — non-LaTeX (Pango).
• Substring access: eq["Q"] (all occurrences, a VGroup), eq["Q"][0] (first), eq[R"\sqrt{d_k}"][0][1:3] (specific glyphs). Mark selectable substrings with isolate=[...] or t2c keys (which auto-isolate). Substrings can't partially overlap.
• Color: Tex(R"e^{i\omega t}", t2c={R"\omega": PINK, "t": BLUE}) at construction, or inline eq["\\omega"].set_color(PINK).
• Animate a number inside an equation: dec = eq.make_number_changeable("0.00"), then drive dec with an updater/ValueTracker — don't rebuild the Tex.
• Braces: Brace(mob, DOWN, buff=0.2) then brace.get_text("A.U.") / brace.get_tex(R"2R_E"). There is no brace_text() free function. BraceLabel/BraceText when brace+label move as a unit.
• Legibility: m.set_backstroke(BLACK, 5) (width 3 small / 5 labels / 8 headline) on any text over busy content — 42× in transformers, preferred over background rectangles.

Building blocks

• Matrices/vectors: DecimalMatrix, IntegerMatrix, TexMatrix, MobjectMatrix; address via .get_rows()/.get_columns()/.get_entries()/.get_brackets()/.get_column(i). Entries auto-tile; brackets auto-fit. (Prefer the video DSL WeightMatrix/NumericEmbedding.)
• Graphs: place data-space objects with axes.c2p(x,y) / number_line.n2p(x); labels via add_coordinate_labels()/get_axis_labels(). Never compute pixel positions.
• Bars are rectangles: Rectangle(prob * width_100p, bar_height) (width is positional, first arg = value); VGroup(*bars).arrange(DOWN, aligned_edge=LEFT); set_submobject_colors_by_gradient(TEAL, YELLOW). No BarChart/Axes for distributions — length-encoding is the point.
• Connectors: pass node mobjects as endpoints — Arrow(boxA, boxB, buff=0.1) — so endpoints resolve to boundaries. For data flow, transform the data object between stages rather than drawing arrow glyphs.
• Animated scalar: one ValueTracker + always_redraw/updaters + DecimalNumber.set_value, so one number drives all dependents and nothing drifts. (455 ValueTracker, 253 always_redraw in repo.) Two-tier sugar: m.f_always.method(getter) and m.always.method(args) over raw add_updater where it reads cleaner — pass bound getters (orbit.get_start), not calls.

Camera / 3D

self.frame.reorient(theta, phi, gamma, center, height) (degrees) · frame.animate.reorient(...) inside a play · frame.add_ambient_rotation(1*DEG) · frame.to_default_state() (snap front-on) · frame.set_height(h) (zoom) · self.set_floor_plane('xz'). Surfaces: ParametricSurface(lambda u,v: [...], u_range=, v_range=, resolution=(nu,nv)) or subclass Surface; built-ins Sphere, Torus, Cylinder, Cube, Prism, Square3D, ThreeDAxes, TexturedSurface. Color a surface with set_color(c, opacity) + set_shading(refl, gloss, shadow) (flat 2D mobjects use set_shading(0,0,0)); always_sort_to_camera(self.camera) for translucent. Light: self.camera.light_source.move_to(...).

---

8. Brand baseline (manimGL default_config.yml)

Canvas is pure black #000000; default mobject color WHITE. manimGL's engine default bg is #333333 — so to get the brand-black canvas, drop a custom_config.yml in the project dir (the way the videos repo does). Verified working (corner pixel renders (0,0,0)):

# custom_config.yml — manimgl reads it from the working directory
camera:
  background_color: "#000000"

Core hues (each has a full A=light → E=dark ladder; C is the base alias):

BLUE_A #C7E9F1  B #9CDCEB  C/BLUE #58C4DD  D #29ABCA  E #1C758A
RED_A  #F7A1A3  B #FF8080  C/RED  #FC6255  D #E65A4C  E #CF5044
GREEN_A ...     ...        C/GREEN #83C167 ...        E ...
TEAL_C  #5CD0B3   PURPLE_C #9A72AC   GOLD_C #F0AC5F   MAROON ladder
YELLOW_C/YELLOW  #FFFF00 (pure!)   YELLOW_D #F4D345   YELLOW_E #E8C11C
ORANGE  #FF862F (no ladder)
GREY_A #DDDDDD  B #BBBBBB  C/GREY #888888  D #444444  E #222222
GREY_BROWN #736357   PINK #D147BD   WHITE #FFFFFF   BLACK #000000
COLORMAP_3B1B = [BLUE_E, GREEN, YELLOW, RED]   # the canonical 4-stop heatmap

Use named constants and their A–E steps — never invent hexes. value_to_color (signed blue/red) and color_gradient(colors, n) / interpolate_color_by_hsl for ramps.

Color roles (bind each to a named variable, reuse everywhere): orange = hero/innovation · yellow = single aha pulse (one per video) · blue = data/process · green = correct/converged · red = error/caveat · purple = secondary · grey = chrome/de-emphasis.

Typography (manim scenes): math Tex renders in mathastext (3b1b's look; \minus macro available). Text/TexText default to CMU Serif, alignment="CENTER" in the videos repo. Inter Tight / JetBrains Mono are NOT used in manim — they belong only to HTML/poster assets. font_size ladder (4K canvas, frame_height=8): 120/96/90 hero · 72/60 headline · 48 (default) /42/36 body · 30/24 caption/inline · 16 fine print. Pass integers from this ladder.

Chrome (paper title, episode number, brand handle, scene dots) is pinned with fix_in_frame(), not an ffmpeg pad.

---

9. Audio sync (manimGL has no voiceover plugin — do it yourself)

For the proven, shipped pipeline → see §13. It supersedes the sketch below: one continuous ElevenLabs VO with character timestamps, scenes speed-fit to narration spans, and burned .ass captions. Use scripts/pipeline/. The notes below explain why.

manim-voiceover's VoiceoverScene is a Manim CE plugin and does NOT subclass manimGL's InteractiveScene — don't use it here. The entire videos corpus contains zero voiceover code (grep voiceover = 0 hits), so 3b1b records VO out-of-band and times animations to it. For an automated reels pipeline, do the DIY equivalent that works on manimGL:

1. Generate one ElevenLabs clip per narration line, cache to audio/lines/NN.mp3 (skip if exists → zero re-cost).
2. Measure each clip's duration once (ffprobe -show_entries format=duration or pydub), store in a dict.
3. Drive run_time from that duration in the scene:

DUR = json.load(open("audio/durations.json"))   # {"l01": 2.4, "l02": 3.1, ...}

# Beat: "Picture the attention matrix."
self.play(FadeIn(matrix), run_time=DUR["l01"]); self.wait(0.3)
# Beat: "Compression collapses every m columns."
self.play(TransformFromCopy(matrix, comp), run_time=DUR["l02"]); self.wait(0.3)

4. Concatenate the line clips (with the small inter-beat gaps) and mux onto the rendered video with ffmpeg. Because each block's run_time equals its line's audio length, picture and voice stay aligned by construction.

TTS copy rules: spell out acronyms with spaces ("M L P", "key value cache") so the engine reads letter-by-letter; spell out numbers for cadence ("eighty-seven cents"); the on-screen caption can still use the compact form. ElevenLabs SDK: export both ELEVEN_API_KEY and ELEVENLABS_API_KEY; if the key lacks voices_read, call the raw TTS endpoint with your voice_id directly (the SDK's voices() validation is what trips on that scope).

---

10. The zero-error gate (NON-NEGOTIABLE — the Prime Directive made concrete)

Every video is nuclear (see the top of this file). This gate runs twice: once on the script + planned visuals before rendering, and once on the rendered video before shipping. Nothing publishes until both pass clean.

A claim is anything a viewer could check, spoken OR on-screen OR captioned OR on the thumbnail: every number, ratio, label, model identifier, layer/expert count, benchmark, price, date, "first to X", and every quantity an axis or readout asserts.

Gate A — the claims ledger (before rendering)

Build a table of every claim from the VO script + every planned on-screen number/label + captions + thumbnail. For each: grade CONFIRMED (with the exact primary-source quote + line/section), WRONG, or UNVERIFIED. Then: fix WRONG, soften UNVERIFIED (≈, "reportedly", "the released config", "see Table 1"), and cut anything you cannot cite. For 10+ claims, delegate to a subagent: read script + on-screen text → extract every checkable claim → verify against the primary source → CONFIRMED/WRONG/UNVERIFIED with the source line → literal edits.

Gate B — audit the RENDERED video (before shipping)

The script being right is not enough — a label or a dramatized number can still be wrong on screen. Sample frames across the whole video (ffmpeg -ss T -frames:v 1, ~every 3–4s), read each one, and verify: on-screen numbers match the claims and each other; labels name the quantity correctly; captions match the VO word-for-word; the thumbnail's claims are true; and no title/transition is garbled. Distinguish a real error from a harmless mid-animation still (a counter ticking, a morph in progress) — but a sustained wrong frame ships as a wrong frame.

The four failure modes that actually shipped a wrong video (learn these by name)

1. Dramatized numbers. "Stack a thousand layers" — the model was ~61 layers. Never inflate a number for effect. If the real number is undramatic, describe the phenomenon without a number ("layer after layer", "go deep enough"), don't invent one.
2. Mislabeled quantities. A relevance score in [0,1] labeled "q · k" (a raw dot product, which would be in the hundreds). The concept was right, the label lied. Name on-screen quantities by what they actually are.
3. Scope overstatement. "Each weight is reduced to four bits" — FP4 was only on the MoE expert weights + the indexer. Say exactly what the source says; "each/every/all" is almost always too broad.
4. Thumbnail hooks that aren't true. "for free" when the model is open-weights but priced. A hook may compress, never falsify.

Standing rules

• Decimal precision past the source reads fabricated (source says ">91%" → write ">91%", not "92.3%").
• Config-specific values (top-k, layer counts, prices) are not universal law — anchor to the specific model/version ("V4-Pro", "Claude Opus 4.6") and note it's the released config.
• Date-drift: model versions/prices change weekly in 2026+; re-verify within 48h of publishing.
• If it's already posted and an error is found: write the corrected re-export AND a short, honest pinned-comment correction (own it; it builds trust).

---

11. Static carousels & posters (matplotlib / HTML — NOT manim)

Carousels are static, so manim is the wrong tool; matplotlib + HTML are right. Use scripts/carousel_template.html (brand baseline, 1080×1350 slides) → scripts/render_carousel.py (weasyprint → pdftoppm), and scripts/render_3d.py for any earned static 3D PNG. The same Laws apply where they can: relative layout (CSS flex/grid, not absolute px), one color per concept, content-sized boxes, earned-3D-only (no faux-isometric cubes), 3 phone-readable annotations max. The matplotlib palette in render_3d.py is the manimGL palette; YELLOW is #FFFF00 (the old #F7D96F was wrong).

For the pixel discipline on a hand-placed matplotlib frame, still probe-render mid-scene frames and read them — but prefer building with relative transforms (axes.transAxes fractions, arrange-like helper functions) over absolute pixels.

---

12. Scripts in this skill

• scripts/manim_scene.py — the real manimGL starter: from manim_imports_ext import *, InteractiveScene, a small self-arranging DSL, transform-driven beats, self.frame.reorient 3D, set_backstroke, SurroundingRectangle, beat-comment structure. Copy and adapt; never write CE.
• scripts/helpers_template.py — the DSL pattern: a self-arranging Mobject subclass + a show_*(scene, ...) choreography verb, annotated with the recipe from Law 3.
• scripts/pipeline/ — the proven render-to-ship audio/caption/assembly pipeline (see §13): vo_continuous.py, timing.py, build_captions.py, assemble.py.
• scripts/thumbnail_template.py — 9:16 thumbnail with the 1:1 / 4:5 safe-crop bands baked in (§13).
• scripts/render_3d.py — matplotlib 3D for static carousel PNGs only (palette fixed).
• scripts/carousel_template.html + scripts/render_carousel.py — static IG carousel pipeline.

---

13. The proven production pipeline (learned shipping reels)

The Laws above are the craft; this is the assembly line that actually shipped a 2-minute manimGL reel end to end. manimGL renders fine — pip install manimgl (1.7.x), then manimgl scenes.py SceneName -w --hd --video_dir out --file_name SceneName writes a flat out/SceneName.mp4 at 1080p. The --file_name flag is what keeps the output flat-named (so assemble.py finds out/<Scene>.mp4); without it manimGL nests under a media/quality subdir. To re-render one fixed scene, just re-run that one command + re-run assemble.py (timing/spans come from the VO and don't change). The render → read-the-frame → fix loop is not optional — sampling a frame (ffmpeg -ss T -frames:v 1) and actually looking at it caught a real overlap on nearly every scene.

Fast path to 3b1b density: vendor their transformer DSL

For ML content, do NOT hand-build matrices and vectors. Copy 3b1b/videos/_2024/transformers/helpers.py into the project as tb_helpers.py, switch its top import to from manimlib import *, and drop the few functions that need external datasets/images. You instantly get the authentic look: NumericEmbedding(length=N) (a column of value-colored numbers), WeightMatrix(shape=(r,c)), EmbeddingArray, ContextAnimation (curved attention rays), value_to_color, show_matrix_vector_product. Scenes built from these read like 3b1b; scenes built from bare cubes read like AI slop.

3D vs flat — the decision that actually matters

Earned-3D is real (Law 4), but 3D hurts some concepts. Hard-won rule:

• 3D wins: point clouds (a vast token field via GlowDots/DotCloud), f(x,y)=z surfaces (a loss landscape with descent paths), volumes that collapse (a KV-cache cube shrinking to 1/10), genuine vector spaces. Camera reorient + ambient drift.
• 3D loses → use flat face-on instead: any stack of thin layers (interleaved attention layers, a residual-layer tower). Viewed at an angle, thin wide slabs collapse into one solid block and the per-layer colors disappear. A flat, head-on striped panel reads instantly. Never 3D a stack.
• But don't let "flat" become monotony. A reel whose scenes are all the same colored grid reads as slop regardless of correctness (real viewer note: "a single block matrix throughout"). Per reel, hit the density bar: real numbers (numeric matrices, an accumulating dot product) + a real on-screen equation + at least one genuine earned-3D scene (a deforming surface, a growing/collapsing tensor volume, a vector space). Rotate the visual primitive scene-to-scene. See the Law 4 counter-lesson for the copy-paste patterns.

Audio + captions + assembly (scripts/pipeline/)

One continuous VO drives everything (manim-voiceover's VoiceoverScene is CE-only — don't use it):

1. vo_continuous.py — ElevenLabs /v1/text-to-speech/{voice}/with-timestamps → vo.mp3 + vo_alignment.json (per-character times). A single continuous read gives natural prosody and zero per-clip silence. Working config: model_id="eleven_multilingual_v2", voice_settings={stability:0.38, similarity_boost:0.85, style:0.20, use_speaker_boost:True}; the endpoint returns base64 audio + a per-character start/end alignment you keep verbatim for the captions. The script joins paragraph breaks to single spaces, so write one paragraph per scene.
2. timing.py — atempo-fit the VO to your length cap (ElevenLabs length varies run-to-run; nudge tempo, capped ~1.12 so it stays human — don't re-generate to chase a number, it burns credits; for a fixed-length series just set TEMPO=1.0 and let the scenes speed-fit), scale the alignment, and emit each scene's narration span by matching an anchor phrase. Anchor gotcha: anchors match on words normalized to [a-z0-9] (lowercased, punctuation + hyphens stripped), so "grouped-query" becomes one token groupedquery and "down-project" becomes downproject — pick each scene's anchor as a short verbatim word-run that survives that normalization, and never lean on a hyphen as a word boundary.
3. Render scenes --hd, then assemble.py — SPEED-FIT each scene's video to its span via setpts (time-scale, so nothing is cut and nothing freezes — no dead air). Keep factors 0.8–1.2×. Two-sided rule: rebuild any scene that plays >1.28× (rushed — trim it), AND enrich any scene that plays <0.6× (draggy — its VO span ≫ its animation): add holds, a second emphasis pass, or an extra beat to raise the scene's native duration so the factor climbs back toward ~0.8 — never speed the VO up to fill the gap. assemble.py prints every factor; scan that list before shipping. Then concat → burn captions.ass → mux VO → loudnorm to −16 LUFS.
4. build_captions.py — phrase-level .ass captions, char-offset-timed from the alignment (read tempo from scene_spans.json so captions and the tempo-fitted VO stay in sync — a mismatch desyncs them). Bottom-center, Inter Tight, semi-transparent box. Keep all in-scene content above the bottom caption band (y ≳ −2.7 at 1080p, the CAP_FLOOR) so captions never collide.

The VO copy register (human, professional, TTS-safe)

• Not robotic: kill mechanical enumeration ("Bet one… Bet two… Bet three…"). One through-line, a real voice. Use the worldbuilder-writing skill.
• Not slangy either: a polished technical-narrator register beats casual slang. "almost counterintuitive" not "almost cheeky"; "a familiar danger" not "a nasty habit"; "the accuracy loss is negligible" not "barely flinched".
• TTS landmines: ElevenLabs read "V4" as "deep 5 four" — never make TTS speak a model version like "V4"; say "the latest model" or spell it out. Spell acronyms ("M H C") or use full words. Numbers like "1,024" and "128" read fine.

Thumbnail (scripts/thumbnail_template.py)

Match the series: ONE bold 3-line headline (last line the accent color) + ONE hero visual + a serif (EB Garamond italic) payoff, on the dark dotted canvas. Build at 9:16 (1080×1920) but keep the core (headline + hero + payoff) inside the 1:1 safe band y[420,1500] (also inside 4:5 y[285,1635]); brand / eyebrow / CTA / footer live outside it. The template renders all three crops — verify the 1:1 (the IG-grid view) reads on its own.

The IG caption

worldbuilder-writing method: open inside the reader's world with one concrete verifiable claim, ride a single leverage point (not a feature list), ground every claim with specifics. Keyword-dense for IG search, ≤2 paragraphs, ≤5 hashtags, no em dashes, no AI-slop phrases. Ship it twice: ig-caption.md (with a # title line) and a plain output/caption.txt for copy-paste — strip the md header with awk 'NR==1&&/^# /{next} NR==2&&/^$/{next} {print}' ig-caption.md > output/caption.txt, then verify the .txt has no em dashes and exactly 5 hashtags.

---

Render-critique gotchas (add to this after every project)

Bugs the frame-reading loop caught, with the fix — check these before re-rendering:

Symptom	Cause	Fix
Only VMobjects can be passed into VGroup	3D Cube/Prism are SGroups, not VMobjects	put 3D solids in Group(...), not VGroup(...)
'Cube' object has no attribute 'set_stroke'	3D surfaces have no stroke	use .set_color(c, opacity) / .set_shading(...)
A fix_in_frame Integer renders gigantic	manual add_updater(set_value) on a fixed decimal	animate with ChangeDecimalToValue, no manual updater
could not broadcast (15,3) into (53,3)	two animations on one DecimalNumber in one play (move + value change)	move via an add_updater(next_to), change value with ChangeDecimalToValue
Stacked 3D layers look like a solid block	thin slabs viewed top-down hide their faces	flat face-on (see 3D-vs-flat above)
Label sits on a curve / another element	placed in an occupied region	move it to a measured-empty region; set_backstroke(BLACK, 5)
A risen/moved mobject covers a label	element animated over fixed text	don't move it over the label, or move the label out first
Captions overlap content	in-scene mobject below the caption band	keep content above CAP_FLOOR (y ≳ −2.7)
A scene plays rushed in the final	scene video ≫ its narration span (speed-fit >1.28×)	rebuild that scene shorter; don't rely on speed-fit alone
Garbled title at a scene's title swap	FadeTransform(old_title, new_title) morphs two different texts at the same position → ~0.5s of scrambled, illegible glyphs (looks like a render bug when paused)	never morph two unrelated titles in place. Do a clean FadeOut(old) then FadeIn(new), or fade out/in with opposite shift=. Same for any in-scene annotation that swaps text (a label that becomes another).
Two texts overlap mid-cross-fade in the FINAL only (looked fine in the silent render)	a FadeOut(a) + FadeIn(b) at the same spot in one self.play — speed-fit stretches it (a 1.5× slow scene makes both co-exist ~50% longer), so the brief overlap becomes a readable garble	sequence them inside the one play with non-overlapping time_span: FadeOut(a, time_span=(0,0.5)), FadeIn(b, time_span=(0.5,1.0)) (fractions of run_time). Survives any stretch.
HUD equation/label collides with the burned captions	a fix_in_frame() Tex/label placed with to_edge(DOWN) sits at y≈−3.4, inside the caption band	place HUD math at y ≈ −2.4 (above CAP_FLOOR −2.7); put step-labels for a 3D scene in the clear band above the surface, not below it
On-screen number is wrong even though the script is right	a label/number hand-set in a scene, never grounded	Gate B (§10): every on-screen number/label is a claim — verify it on the rendered frame, not just in the VO
A small/exact WeightMatrix shows a phantom "…" dots row or column	WeightMatrix/NumericEmbedding default to ellipses_row=-2, ellipses_col=-2 (a deliberate ellipsis to imply a big array)	pass ellipses_row=None, ellipses_col=None for any matrix you want shown in full. For a 1-wide NumericEmbedding the col-ellipsis can blank the whole column — set ellipses_col=None, or build vectors from a plain cell-column helper
A near-identity / small-valued WeightMatrix renders dark, low-contrast	default value_range=(-9.9, 9.9) maps a value of 1.0 to ~10% intensity	pass a tight value_range (e.g. (-1.5, 1.5)) + explicit high_positive_color=/high_negative_color= so the diagonal reads bright and off-diagonals stay dim
An auditor flags a matrix-product / dot-product entry as "wrong" on a sampled frame	show_matrix_vector_product and dot-product accumulation animate the running partial sum — a mid-frame entry is a correct intermediate, not the full product	only the settled END/hold frame must equal the full product; verify that frame, not a mid-accumulation still. (Gate B: distinguish transient from sustained — a partial sum counting up is not an error)
timing.py: "anchor not found"	the anchor isn't a verbatim contiguous word-run after [a-z0-9] normalization (hyphens merge tokens)	re-pick the anchor from the exact VO words: "grouped-query"→groupedquery, "down-project"→downproject
A scene plays draggy in the final (speed-fit < 0.6×)	the scene's narration span is much longer than its animation	enrich the scene (add holds / a second emphasis pass / an extra beat) to raise native duration — do NOT speed the VO up
A stroke-only ring/shape renders as a filled disk (esp. faux-3D blobs)	.set_opacity(x) on a Circle/VMobject sets BOTH stroke AND fill opacity, so Circle().set_stroke(c,2).set_opacity(0.5) fills it	use .set_stroke(c, w, opacity=x) + an explicit .set_fill(opacity=0) for a hollow ring
Tex throws a cryptic LaTeX error (\UseTextAccent has an extra }, \bBigg@ has an extra })	fragile constructs — \big(/\bigg, {+} spacing hacks, doubled \;\;, or coloring via fragile substring indexing tex[r"..."]	use plain ( ) (nesting \cos(w(a+b)) is fine); \begin{pmatrix}…\end{pmatrix} works; split mixed math+prose into Tex(...) + Text(...); color with t2c= at construction or color the whole mobject, not a post-hoc substring slice

---

Project folder convention

NN-short-slug/
├── paper-summary.md  related-work.md  discussions.md  moodboard.md   # lean recon
├── beat-sheet.md                                                     # ordered named beats
├── tb_helpers.py                                                     # vendored 3b1b transformer DSL (§13)
├── helpers.py                                                        # project-specific DSL (Law 3)
├── scenes.py                                                         # thin InteractiveScene files
├── custom_config.yml                                                # black canvas (§8)
├── .env                                                             # ELEVENLABS_API_KEY, ELEVEN_VOICE_ID
├── audio/  vo_script.txt  vo.mp3  vo_alignment.json  scene_spans.json
├── captions.ass
├── thumbnail.py   ig-caption.md
└── out/FINAL.mp4

---

Appendix — copy-paste manimGL idioms (verbatim grammar)

# Ensemble entrance — always staggered, small drift
self.play(LaggedStartMap(FadeIn, things, shift=0.25 * DOWN, lag_ratio=0.1))

# Born-from morph (the workhorse — source persists)
self.play(TransformFromCopy(data_column, x_symbols), run_time=2)
self.wait()

# Highlight one symbol in an equation
eq = Tex(R"\text{softmax}\left(\frac{K^T Q}{\sqrt{d_k}}\right)V")
self.play(FlashAround(eq[R"\sqrt{d_k}"], time_width=1.5, run_time=2))

# Content-sized box (never hand-sized)
box = SurroundingRectangle(label, buff=SMALL_BUFF).set_stroke(YELLOW, 2)

# Snap many objects into a new arrangement with no hand coords
grp.target = grp.generate_target()
grp.target.arrange(RIGHT, buff=0.15).next_to(anchor, DOWN, buff=MED_LARGE_BUFF)
self.play(MoveToTarget(grp))

# The aha reveal — one play: camera + born-from + cascade, then breathe
self.play(
    self.frame.animate.reorient(-110, 12, 0, payoff.get_center(), 6.7),
    TransformFromCopy(setup, payoff, time_span=(1.5, 3)),
    run_time=3,
)
self.wait(2)

# Real 3D hold — never static
self.frame.add_ambient_rotation(0.5 * DEG)
self.wait(8)
self.frame.clear_updaters()

# Animated number driving dependents
tracker = ValueTracker(100.0)
num = DecimalNumber(0).add_updater(lambda m: m.set_value(tracker.get_value()))
self.play(tracker.animate.set_value(10.0), run_time=2)

# Section-comment teleprompter (each block ≈ one narration sentence)
def construct(self):
    # Add the sentence
    ...
    self.play(LaggedStartMap(FadeIn, words, lag_ratio=0.25)); self.wait()
    # Box the adjectives
    ...
    self.play(LaggedStartMap(DrawBorderThenFill, adj_rects)); self.wait()
    # The reveal — why this works
    ...
    self.play(..., run_time=3); self.wait(2)

Add to this appendix after every project. Named idioms compound; named failures don't recur.