Skip to content

For people who want to make data videos, not just read about them

A finished data reel, in two commands.

47.06s

2160×384060 fps7.9 MB

read out of ffprobe, not the render log

The project behind @reelgorithm.py — 18.5M views across 75 reels. The chart library, the motion layer, and one worked episode.

Regular price $79.99, now $39.99 onceSave $40 until December 1530-day refundLifetime updates

~/reel-starterthe two commands
$ npm install $ npm run render  remotion render src/index.ts Episode \    out/Episode.mp4 --scale=2 --crf=16   bundled   src/index.ts  rendered  Episode   2,820 frames   out/Episode.mp4  2160x3840  60fps  47.06s  7.9 MB

2,820 frames · 13 dependencies · Node 18+ · nothing else

18.5M

views on the channel

75

reels, all measured

80K

Instagram followers

98

charts, 15 families

51,134

words, and the code

measured · 18,515,560 views and 79,738 followers read out of analytics/data on 3 November 2026. What the work did — not a number you are promised.

The work

Eight reels, and what each one measured

Cut from the delivered masters — the files that were posted, not re-renders. The six most recent under the skip ceiling, plus the two that set it.

The team the flags cost most

33K
views
32.1%
skip@3s

Strongest in the final five

35K
views
38.8%
skip@3s

The best bank

49K
views
40.5%
skip@3s

What your team owes you

495K
views
33.4%
skip@3s

The market is right — that's why you lose

31K
views
36.9%
skip@3s

The hardest sport to go pro in

863K
views
36%
skip@3s

The player who shouldn't exist

5.4M
views
40.2%
skip@3s

Ticketmaster

254K
views
22%
skip@3s

Drag the ring, or pick a reel for the rest of its numbers.

source · codepop/out/*.mp4 · figures from analytics/data/history.csv, pulled 2026-11-03 · 18,515,560 views across 75 reels

Who wrote it

@reelgorithm.py

I run a data-storytelling channel. One named dataset per video, the data published with it, and the working shown in an article afterwards. 68.4K people follow it and 15.3M views have gone through it — and neither of those is a number I’m promising you.

Anatomy of a Data Reel is that process written down after 65 of them: what gets picked, what gets thrown out, where the data comes from, and the eight or nine times I got it wrong first.

You have no reason to take any of that on faith, so none of it is behind a login:

The channel is the evidence and the book is the method. If the videos don’t look like something you’d want to be able to make, that’s a good reason not to buy this.

The numbers

The hook doesn’t make a reel work. It decides whether it’s allowed to

Every reel on the channel, plotted on the one measurement that predicts the rest. The interesting part of this chart is the part with nothing in it.

1K10K100K1M10Mempty28 reels, none above 157,932most viewedlowest skip20%30%40%50%60%70%80%left within 3 seconds →views

34×

more views available below 45% skip

157,932

the best of the 28 above it

5.4M

the best of the 47 below

-0.7

correlation, against log views

The most-viewed reel on the channel is not the best hook. Messi sits at 40.2% — worse than eight reels that did a hundredth of its numbers. A hook cannot make anyone care. What it can do is decide whether anyone is still there to find out, and that is why there is no line fitted through this cloud: the shape is a ceiling, not a slope.

source · analytics/data/history.csv · 75 reels with a skip reading · pulled 3 November 2026 · best TicketDrop 22%, median 42%, worst Sorting 84.6%

And they keep going

36 reels, 112 days, nothing new posted

+21.1%

The same set of reels, frozen on 14 July 2026, picked up 2,300,272 views on their own. Nothing in this line is a new post or a reel the tracker found later.

0+1.3M+2.5M14 Jul3 Nov

source · analytics/data/history.csv · the 36 scenes already tracked on 2026-07-14, carried forward to 2026-11-03 · views gained, so the baseline is a true zero

In the box

A repo that runs, and about 51,000 words on why it runs that way

One zip, 0.27 MB, 72 files, Node 18 and nothing else — and the writing is a web page you open in a browser, not a PDF and not a video course.

reel-starter.zip0.27 MB · 72 files
reel-starter/
  src/
    viz/          98 charts, 15 families
      core/       scales, marks, morph
    episode/      the worked 6-beat reel
      beats/      one file per beat
      layout.ts   throws on collision
    motion.tsx    easing, springs, air
    grid.ts       the 8pt law
    Watermark.tsx ships as "yourchannel."
  data/
    pro-pipeline.csv   the episode's dataset
  qa/ql.mjs       the gate
  scripts/
    sheet.mjs     24 stills, one bundle
    make-data.mjs one real bug, on purpose
5 galleries13 dependenciesTypeScriptRemotionno API keys

Three families are held back. geo, three, iso3d are the only files pulling three.js, deck.gl and topojson — roughly 200 MB of install, and where every one of the four silent black-frame traps lives. Leaving them out is a decision, not an oversight.

ranking14

BarH · Dumbbell

time10

Line · Area

distribution9

Histogram · Violin

relation9

Scatter · ScatterFit

flow8

Sankey · Chord

partwhole7

Treemap · Waffle

stat7

ForestPlot · QQPlot

advanced6

ParallelCoordinates · RadialBar

matrix6

Heatmap · CorrelationMatrix

single6

BigNumber · Bullet

spatial5

Choropleth · PointMap

hierarchy4

Dendrogram · CirclePack

atlas3

SystemMap · DecisionTree

showcase3

Globe · IsoCity

sport1

SportHeatmap

One contract, 98 charts. Every one takes the same props, so any chart swaps for any other in a scene — and a bar can morph into a scatter, because the marks are matched by a stable key rather than by array position.

Twelve chapters, twelve questions

about 51,000 words7 format blueprints10 fill-in sheets

  1. 01

    Is this idea even worth making?

  2. 02

    Where do I get data that isn't already on everyone else's page?

  3. 03

    How do I know my numbers aren't quietly wrong?

  4. 04

    How long should each part of the video actually be?

  5. 05

    Why does my narration sound like a robot reading a list?

  6. 06

    Why does nobody watch my chart?

    free to read below
  7. 07

    What's about to quietly waste my entire day?

  8. 08

    Did it work, and how would I actually know?

  9. 09

    What does all of this look like on one real video?

  10. 10

    How do I get a project rendering on my own machine?

  11. 11

    How is one of these actually put together, beat by beat?

  12. 12

    What's between a rendered file and a posted video?

source · codepop/starter — tree from find, counts from the library’s own VizMeta registries, zip from scripts/pack-starter.mjs · chapter questions from content/copy.ts

FreeChapter 62,291 wordsNo email

Making a chart carry an argument

The whole chapter, not an excerpt. It’s the one I’d read first if I were deciding whether to buy the rest — if it isn’t useful to you, nothing else here will be either.

This is the free chapter. It's the one I'd read first if I were deciding whether to buy the rest.

A chart in a video has a different job from a chart in a report. A report's reader can look twice. A video's viewer gets one pass, at speed, often on mute, and cannot read a value off a moving mark.

That constraint drives almost everything in this chapter, and it has an academic backing that's worth knowing about because it's counterintuitive: animation is excellent for presentation and one of the worst formats for analysis. Static traces and small multiples win when someone needs to read exact values. Reels are presentation, so animation is correct — but it means every analytical beat has to settle on a clear static state and the value has to be annotated explicitly. Never make the audience eyeball a quantity from motion.


1 · One chart is not an analysis

The blunt version, delivered after a video whose entire statistical content was "a radius got smaller", shown on a single line chart: "doesn't dig into any of the in-depth metrics, uses one boring chart, this just isn't a well made video."

Three rules came out of it:

  • A chart per analytical beat, each answering a different question.
  • At least one must show a distribution, not a summary statistic. On that video I computed a 67th percentile and threw the distribution away — when the percentile was literally the thing being explained.
  • Before locking a script, ask: "what is the most interesting number in this dataset that I have not put on screen?" For that episode it was the along-track versus cross-track error split — the reason the forecast cone is becoming an ellipse rather than a circle, which was on disk and never used.

And a related trap: don't let already-fetched assets set the ceiling. I had the position data and the wind radii locally, everything needed to animate a storm moving and growing, and rendered one frozen frame of it because the still was already downloaded. Fetching is cheap. A dead chart isn't.


2 · Lie factor near one

The graphical element's size should be proportional to the quantity. That's Tufte's rule and it's the floor, not the ceiling.

The version that actually bites in practice isn't bar-area distortion, it's axis machinery. A charting library that "nices" an explicit domain will round [1, 9] outward to 0 → 10 and put ticks at rank 0 and rank 10, neither of which exists. For an ordinal quantity that isn't cosmetic — it invents categories. Full write-up in chapter 7 §5.

Check the rendered axis end labels against what you asked for, in a still, every time. An explicit domain is a request, not a guarantee.

Three more that fall in the same bucket:

  • Real dimensions must be real. Any pitch, court or field has to use true proportions with equal pixels-per-metre on both axes. A squashed rectangle is a lie factor problem wearing a sports jersey.
  • Place event markers to match the real description of each event — near-post finish, thirty-yard strike, cut-in from the left, and credit them honestly ("shot spots match each goal's description"), never as exact coordinates you don't have.
  • Normalise to one direction. Teams switch ends at half-time, so a first-half and second-half goal must both be drawn attacking the same goal. Don't split them onto opposite sides of the pitch because that's where they physically happened.

3 · The honest statistics layer

For anything that claims a relationship:

  • Plot the real points. All of them, if you have them. Never pad a cloud with generated dots to make it denser.
  • Draw the fit, and draw it as its own beat rather than as chrome.
  • Show the interval. A shaded 95% band around the fit.
  • Print r, n and p on the plot. On-plot, not in a card, so they can't be separated from the picture they describe.
  • Units in the axis titles. Always. "Response time" is not an axis label. "Ambulance response time, minutes" is.
  • Keep any on-screen sigma and any derived rarity computed from the same numbers, so they can't disagree with each other inside one frame.

And the method has to fit the distribution — count data isn't normal, a sigma on a small sample is descriptive only, and rarity gets quantified as a frequency or with a model that fits. That's chapter 3 §5, and it's the difference between rigour and the appearance of rigour.


4 · Emphasis by de-emphasis

The single most-cited colour rule in data journalism, and the one people underdo: grey the crowd and light one thing.

Reserve exactly one saturated accent for the subject. Everything else is low-contrast. And when you highlight the subject, actively dim everything else — don't just brighten the hero, darken the field. The delta is what the eye reads.

Density encoding for the cloud. Let opacity encode local density so the mean blob glows solid and the tails fade. The shape itself then shows the relationship before any line is drawn, which means your first analytical beat has already landed before the fit appears.

Keep the entry a quick coordinated rain-in, not a long cascade. Heavy staggering gives negligible or negative tracking benefit — that's a measured result, not a taste preference.

A caution on one-accent discipline. It's easy to reach for a second accent when two things need emphasis in the same frame. If two things need emphasis, you have two charts.


5 · Stage the reveal

Never show a finished chart. A great animated chart is a sequence of reveals, each synced to a narration beat, building one analytical layer at a time and ending on a single surprising point.

The universal order:

axes/frame  →  the cloud rains in  →  the fit draws on  →
a threshold or highlighted band  →  the ONE hero point

Start simple, add overlays one at a time, and give the eye time to read each before the next.

Distribution first, outlier last. Establish what "normal" looks like before the surprise. The famous version of this move shows years of typical variation and then drops the outlier year on top of it — the outlier only means something against an established baseline.

The design principles behind staging, from the controlled experiments, and they're worth applying verbatim:

  1. Maintain valid data graphics through the transition. Every in-between frame is still a real chart.
  2. Object constancy. A dot representing a datum keeps its identity across the whole animation. Never re-spawn it — the eye is tracking the same object.
  3. Use staging for complex changes. Never move everything at once; sequence the sub-transitions. Staged animation beat simultaneous animation in almost every condition tested.
  4. Group similar motions. Animate like objects together with shared timing — fewer things to track keeps it inside working memory.
  5. Minimise occlusion and travel distance. Prefer simple, predictable paths.
  6. Slow in, slow out, on every move. It signals start and end, and it feels physical.
  7. Avoid ambiguity. Don't let points cross or swap in ways that break correspondence.

Two further constraints from the animation-comprehension literature: apprehension — the motion has to be perceivable, so segment it and don't over-densify, and congruence — the motion must correspond to a real change in the data. Decorative motion raises cognitive load and buys nothing.

The camera is the narrator's finger. Open wide showing the whole cloud, then a slow S-curve push-in to the region carrying the story, then pull back wide for the payoff. Ease it; never a linear zoom. And clip the plot to its own band so a zoom can't spill over the rest of the layout.


6 · Annotate on a static state

Because viewers can't read values off motion, every analytical beat settles and then labels.

  • A drawn ring or arrow tied by a leader line to a two-to-four-word label. Animate the stroke on rather than fading it in.
  • A guide line from the point to the axis, so the viewer can literally read the value off the scale.
  • Label the value at the element, at the end of the line or beside the point. Never make them find the axis.
  • Number readouts count up rather than hard-cutting.

Draw the gap you state out loud. If the narration says one thing is twice another, there should be a visible measured span on screen showing it. An asserted comparison is a claim; a drawn one is evidence. This is the same coupling idea as the caption rule in chapter 4 §6 — the words label the mark, the mark proves the words.


7 · The minimalism target

The aesthetic bar is dark-field macro-finance minimalism, and it comes down to six things:

  • Near-black field, ruthless minimalism, one message per chart. Almost no chart junk, few or no gridlines, huge negative space. If an element isn't carrying the point, delete it.
  • A bold takeaway title that states the conclusion. Not "Income vs response time" — "Your ZIP decides how fast help comes."
  • Muted field, one accent.
  • Threshold lines and regime shading. A dashed horizontal line at a meaningful benchmark and a shaded zone for the danger quadrant. For an ambulance-response chart: dash the real eight-minute standard across the plot and shade the quadrant of poor ZIP codes that blow past it. Then the story is "look how many dots live in the red zone" rather than a scatter you have to interpret.
  • Smooth draw-ons with values labelled at the element.
  • Confident pacing. Let a beat breathe on a static, fully-labelled state before moving.

8 · The banned chart, and what replaces it

Slope charts are banned outright. Two ranked columns joined by diagonals turn into spaghetti the moment more than about three entities cross, and the crossings are the entire reason you drew it. The chart destroys its own subject.

Use a dumbbell instead. One row per contender, both endpoints legible, and the length of the connector is the size of the disagreement. Same finding, measurable instead of tangled.

Pie is banned everywhere except one slot: a set of weights, animated from equal slices to actual ones. Weights genuinely are angular shares of a whole that sums to one, which is the single case a pie encodes honestly, and the equal-to-weighted animation supplies the baseline a static pie never has. A pie anywhere else is a bug.

And never let big text be the visual. A sentence set large is not a chart. If your beat's visual is a headline, the beat has no visual — you have narration with a caption. Every beat needs something that encodes a quantity.


9 · The chart is not the hook

This is where chart craft collides with distribution reality, and the collision is expensive if you don't know about it.

Every static-chart opener on this channel lives in the 40–50% skip band. A rigorous 43-chain scatter with logos doing a physics drop-in measured 44.7% skip and about 4,500 views on the same week that a moving-simulation opener measured 21.4% and 218,000. The chart data was real and rigorous, and that did not save it. It read as "a fact", not as a spectacle.

A settled scatter, beeswarm or bar chart at frame 0 is a chart no matter how many marks it has. It is not motion.

But charts aren't banned — a chart opener that runs works. Within chart-openers alone the correlation between skip and views is −0.78, and one running-chart opener measured 27.1% skip and 132,000 views.

The recipe that reliably drags a chart concept out of the death band: find the physical process behind the number and animate that at frame zero, then resolve to the chart. An income-versus-ambulance-response scatter opened on two clocks ticking side by side — one freezing when its ambulance arrives, the other still climbing, and posted 36.3% skip, the best any chart concept had ever opened on that channel, clearing the whole death band.

So: lead with the process, never with the axes. Then everything else in this chapter applies to the chart the process resolves into.


10 · Anti-patterns

Each of these has shipped here at least once.

  • Dumping the whole finished chart at once. The most common one.
  • Rainbow or multi-accent colour. Two accents means neither is an accent.
  • Decorative motion that reveals nothing. Fails congruence; costs attention.
  • Over-staggered dot entry. Measured as no better and sometimes worse.
  • Linear or robotic zooms. Every camera move eases.
  • A text sentence used as the visual.
  • An annotation anchored off the top bar of a sorted bar chart. The top bar is the longest by construction, so the callout lands on the data. Anchor into real white space.
  • Labels stacked below a dot in a dense region. Side-place them — left-half events label left, right-half label right, so they never collide.
  • A settled scatter at frame zero. §9.

11 · The checklist version

Every chart in a video should be able to answer these before it's built, not after:

  1. What single question does this chart answer, and is it different from the previous chart's question?
  2. Is a distribution visible somewhere in this video, or only summary statistics?
  3. Does the mark size match the quantity, and do the axis end labels match the domain I asked for?
  4. Is the crowd grey and is exactly one thing lit?
  5. Does it build in layers, one per narration beat, ending on the hero?
  6. Does every analytical beat settle on a static state with the value labelled at the element?
  7. Is every stated comparison also drawn?
  8. Are units in the axis titles, and are r, n and p on the plot?
  9. Is anything moving that doesn't correspond to a change in the data?
  10. If this is frame zero — is it a process, or is it axes?

The printable version is in the pack: the chart-craft checklist.

10 more failures · free · no email

That was one chapter of 12.

It stays free whether or not you buy anything — nothing on this page moves behind a paywall later. The other 11 are the same depth, and they’re $39.99 once with 30 days to change your mind. That’s the founding price — it goes to $79.99 on December 15.

Get the other 11 — $39.99

30-day refund · lifetime updates · no subscription

Price

One price, once

$39.99 until December 15, then $79.99 — the same book either way, and 30 days to change your mind. The chapter you just read stays free whatever you decide; nothing on this page moves behind a paywall later.

Regular price $79.99, now $39.99Save $40

One time · Lifetime updates · No subscription

Update 2: the motion-graphics build-along arrives by December 15, free: how the three newest reels were made, with the ASCII, motion and sound code behind them.

Founding price. It goes to $79.99 on December 15 and stays there. Buy before then and you keep every future update at no extra cost — including the five formats still marked as hypotheses.

Anatomy of a Data Reel

  • reel-starter — the project — runs with two commands; renders a 47.0s reel at 2160×3840 · 60fps
  • The chart library — 15 families, one shared contract, staged reveals built in
  • One worked episode — 6 beats, its dataset, its script, its beat clock — and its bugs
  • 12 chapters — 33,013 words, every number measured. The last three build a reel
  • 7 format blueprints — 2 proven in full, 5 marked as hypotheses
  • 10 fill-in sheets — 386 tick-boxes, 279 blanks, all persisting
  • Lifetime updates — text and code, same link, no second purchase

30 days, no questions.

Not a form, not an interrogation. If the project doesn’t run on your machine, or a chapter isn’t delivering, that’s my problem to fix — tell me and take the refund anyway.

  1. Lemon Squeezy takes the payment — card details never touch this site.
  2. You land back here on a page with the link on it. No waiting for email.
  3. The same link is emailed to you. It doesn’t expire, and updates arrive there.

$39.99 once — the founding price until December 15, after which it is $79.99. Lemon Squeezy is the seller of record and handles any VAT or sales tax, so the price above is the price you pay. By buying you agree to the terms and privacy policy.

$0Already published

What’s free, and stays free

  • 41 datasets — 79,924 rows across 64 files
  • 14 articles — one per shipped video
  • Chapter 7 — complete, on this page, no email

The raw material behind every video, with its methodology note. Not a sample of the paid thing — it was free before this existed.

No bragging, just real mistakes and lessons and how to learn from them, which makes it feel trustworthy. The content is deep and detailed, like actual hard-earned knowledge.
Pre-release readerRead a free pre-release copy
Go to the kit

The code

Three decisions you can disagree with

Verbatim from the zip — not a simplified version for the page. The whole of the 98-chart library works the way the first one does.

the contractsrc/viz/core/contract.ts

One shape, 98 charts

Every chart in the library takes the same props, so any chart swaps for any other in a scene. The rule that makes it work is the one that is hard to retrofit: every datum carries a stable key. A scatter point for Kansas City and a bar for Kansas City are the same object in two arrangements — which is why the transition between them is a translation and not a cross-fade.

/** The universal datum. Every field except `key` is *  optional — a bar needs {key, value}, a scatter needs *  {key, x, y} — but both are Datum, so either can *  morph into the other. */export type Datum = {  /** stable identity. Survives across chart types. */  key: string;  x?: number | string;  y?: number;  value?: number;  group?: string;  /** THE subject, drawn in the accent colour. Normally   *  exactly one per chart: more than one accent and the   *  frame no longer has a point. */  accent?: boolean;};
the clocksrc/episode/beats.ts

Measure the read. Don't budget it

The runtime is an output, not an input. Pick 45 seconds first and divide by six, and the day you record the narration every beat is wrong at once — not by a little, and not fixably, because every animation inside a beat is timed against a start that just moved. The cap is a constraint on the script.

// The mistake almost everyone makes: you pick "45// seconds", divide it into six, then record the// narration and discover beat 4 needed nine seconds.// Now every beat is wrong.//// So it goes the other way round://   1. Write the script.//   2. Speak it, and MEASURE each line.//   3. Each beat is its own clip, plus a little air.//   4. The runtime is whatever that adds up to. export const BEAT_SECONDS = [  6.0,  // 1 · the question  7.5,  // 2 · what we measured  8.5,  // 3 · the field  8.0,  // 4 · the obvious answer, and why it's wrong  9.0,  // 5 · the answer  8.0,  // 6 · the finding, and the ask] as const;
the call sitesrc/episode/beats/Beat6Finding.tsx

What using it looks like

No axes to build, no reveal to stage by hand. plotBox() encodes the Reels safe band so the chart cannot land under the caption, and the six reveal stages — frame, scaffold, marks, structure, emphasis, annotation — are timed by the library from one start.

import {ScatterFit, plotBox, PALETTES} from '../../viz'; const BOX = plotBox(PLOT_SHORT); <ScatterFit  frame={frame}  start={start + 0.3}  box={BOX}  data={POINTS}  palette={PALETTES.pink}  xLabel="years of training"  yLabel="chance of going pro"/>

source · codepop/starter/src — verbatim, as shipped in the zip

The machine

An agent writes most of this repo. Nothing ships until a gate says so

Speed was never the hard part. The hard part is knowing, without watching a 47-second render, that the thing you just changed didn’t quietly break a frame three beats later.

CLAUDE.mdthe router
## 2 · viz library = main tool // hand-build an axis, or hand-stage a reveal⊥ hand-build axes. ⊥ hand-stage reveals. - V1: ∀ chart takes VizProps<D> = frame, start, out?, box,      data, labels, palette? ∴ any chart swaps for any other;      morph keyed on Datum.key.- V2: six stages, always this order — frame → scaffold →      marks → structure → emphasis → annotation.- V3: box ∀ chart with plotBox() / panelBox() / rowBox(),      ⊥ raw margins — they encode the Reels safe band.

A map, not a rulebook

One file at the root says where everything is and which single document to open for the task at hand. An agent that reads it doesn’t grep 80 files to find the chart library, and doesn’t hand-build an axis that already exists.

The gates

  • ql-sec83 passed

    webhook signatures, open redirect, CSP, consumer law

  • ql-starter52 passed

    what may leave the building in the paid zip

  • ql-chartsgreen

    safe band, edge clip, label collision

  • ql-layoutgreen

    the 8pt grid, statically

135assertions green

run 2026-09-18 · counted from the run, not from memory

in the boxstarter/src/episode/layout.ts

A comment saying “these must not overlap” is a comment

The worked episode’s vertical budget throws at import time. Make a chart taller without moving the caption and the build stops, rather than shipping a frame with two layers of text on top of each other. Four beats of six collided the first time this was built, and the studio showed none of it.

src/episode/layout.tsships in the zip
for (const [name, plot, caption] of [  ['PLOT_TALL', PLOT_TALL, CAPTION_TALL],  ['PLOT_SHORT', PLOT_SHORT, CAPTION_SHORT],] as const) {  if (bottomOf(plot) > caption) {    throw new Error(      `${name} ends at ${bottomOf(plot)} but captions start ` +      `at ${caption} — the source line would render on top.`,    );  }}

07:00, every day

Four desks. No model in the loop.

A launchd job wakes four desks — six keyless public sources, SEC Form 4 filings, House roll calls, federal contract awards — and each one ships a carousel. The scoring that picks the day’s angle is a σ calculation, not a prompt. Deliberately: a generator that can be surprising can also be wrong in a way no gate catches, and this one runs while nobody is watching.

  • daily

    6 keyless sources

  • insiders

    SEC Form 4

  • floor

    House roll calls

  • contracts

    USAspending

ASK → TEST → ANSWER · 4 carousels a day · zero LLM calls

To be exact about what that means for you: the $39.99 ships the repo and the book — the chart library, the motion layer, the worked episode and its gate. It does not ship the agent setup, the router, the daily engine or the analytics stack. Those are how this studio runs, not a product, and the full exclusion list is further down this page.

source · codepop/CLAUDE.md · qa/ql-sec.mjs and qa/ql-starter.mjs, run 2026-09-18 · codepop/daily/README.md · starter/src/episode/layout.ts

The anatomy

A ranking episode is six slots and a stopwatch.

Not a style. A spec — 2,850 frames at 60fps, six beats, 185 words of voiceover, locked so an episode becomes a spec file plus a dataset instead of a new build every time.

This is the real table out of the format spec, at true scale. One beat takes 38% of the runtime, and the chapter explaining why is the part you’re buying.

47.5s

2,850 frames · 185 words · 6 beats

  1. Hook4.0s16w
  2. The Ruler7.5s29w
  3. The Field6.5s25w
  4. The Examples18.0s70w
  5. The Podium8.0s31w
  6. CTA3.5s14w

Where the format spent its blood

Four episodes of rebuilding one four-second beat.

Measured skip at 3 seconds — lower is better

Twenty-seven points of improvement across four episodes, and views moved with it.

The direction of travel is the whole finding: wide, dense, everything at once, camera moving out. The metro episode showed a few rows racing. The bank episode showed the entire country extruded, every state carrying the mark of the bank it actually uses.

Read those skip numbers as roughly comparable: none of these four was ever pushed to a cold audience, and measured skip inflates when a reel breaks out. These are the least-contaminated readings on the channel.

Both tables above are printed from formats/01-the-ranking.md, one of the seven blueprints. What isn’t here is the analysis around them — why beat four absorbed two beats that were deleted, and the measurement showing every comma costs about half a second of voiceover.

Readers

Five people read it before it went on sale

Each was sent the whole thing free and asked to be honest. Their words are trimmed for length and nothing is reworded — including the two who said it reads long, which is answered underneath rather than left out.

  • It’s an honest, well-argued manual on making data reels that guides you through all the process and its iterations to get the best result, and it knows exactly what it is — failures costed, untested claims labeled as guesses. If you also do analysis and decks, there’s a solid chapter of verification discipline and a handful of chart rules worth lifting.

    Pre-release reader

    Read a free pre-release copy

  • No bragging, just real mistakes and lessons and how to learn from them, which makes it feel trustworthy. The content is deep and detailed, like actual hard-earned knowledge.

    Pre-release reader

    Read a free pre-release copy

  • There’s quite a few amazing concepts. I’ve always been captivated by the reels and I can see why… I’m getting loads of gems. It’s applicable to one of my side projects.

    Pre-release reader

    Read a free pre-release copy

  • The dataset article has great decision-making philosophy without the boringness of a textbook, but still carries a lot of the important information. Note that an immediate solution definitely does not come, as there's lots to unpack — but quite worthwhile.

    Pre-release reader

    Read a free pre-release copy

All four received a free copy in exchange for an honest review. None were paid, and none were shown a draft of what to say.

The most common criticism, before you buy it

Early readers said it reads long — a wall of text with the density turned all the way up. That is fair, and the rebuild into modules with diagrams and inline checklists is underway. Everyone who buys now gets it without paying again.

The line

What you get, and what you don't

Printed in the same words the book uses. The right-hand column is everything that is genuinely not in the box, and I'd rather you knew now than in thirty days.

  • A runnable Remotion project: the frame, the grid, the motion layer, the atmosphere
    My API keys, tokens and environment — you bring your own
  • The chart library — 15 families, one contract, staged reveals built in
    The three GPU families (geo, 3D, iso) — 200 MB of install and four ways to render black
  • One complete worked episode: source, dataset, script, beat clock, and the bugs in it
    The analytics pull, the funnel scripts, the posting automation and the scheduler
  • The build-along — empty folder to rendered file, then to a delivery encode
    The licensed voice model, and any audio at all. The chain is taught; the asset isn’t mine to give
  • The QA tooling: the chart gate, and the contact-sheet sweep that found four real bugs
    My channel’s own analytics data, and anything identifying me
  • How ideas get picked, and the two kill gates that reject most of them
    Photographs, logos and third-party imagery — every one is somebody’s licence
  • Data sourcing routes that are hard to find, and the traps in each
    A guarantee that any of it will work. See the no-guarantee section
There’s quite a few amazing concepts. I’ve always been captivated by the reels and I can see why… I’m getting loads of gems. It’s applicable to one of my side projects.
Pre-release readerRead a free pre-release copy

The rule it holds in every ambiguous case

If it’s a decision or a tool, it’s in; if it’s a credential, an account or somebody else’s licence, it’s out.

You get the code

You get the code. It’s a real project — `npm install`, `npm run render`, and a 47-second reel comes out. One thing that’s on you: Remotion, the renderer underneath it, has its own licence and needs a company licence above a certain size. Nothing you pay me covers that.

Fit

Who it’s for, and who should skip it

The right-hand column has stopped more wrong purchases than anything else written here. It is not modesty and it is not a disclaimer.

Buy it if

  • Someone who wants to make these — data-driven short video — and would rather start from a working project than from an empty folder.

  • Someone making them already who keeps ending up with a chart nobody watches.

  • Someone who wants the decisions as well as the code: what to measure, what to withhold, how long a beat runs, why the hook keeps failing.

  • Someone doing analysis for an audience who needs the honesty apparatus: sensitivity, residualisation, tie handling, the difference between a claim and a measurement.

  • Someone who has been burned by a tool that returned a plausible number without doing the work.

Don’t buy it if

  • Anyone who won’t open a terminal. The project is TypeScript and React, run from a command line. You don’t have to be good at either — the worked episode runs with two commands — but there is no visual editor here and there isn’t going to be one.

  • Anyone looking for growth tactics. There’s one chapter about measurement and it spends most of its length explaining why three things I believed turned out to be false. If you want a list of hooks that go viral, this will annoy you.

  • Anyone who doesn’t have or want a dataset. The whole method is built around one real, nameable dataset per video. Without that, most of this doesn’t apply.

  • Anyone who needs it to work on a platform I haven’t measured. All the retention numbers here come from one channel on one platform. The craft transfers. The specific thresholds might not.

  • Anyone in a hurry. The shortest chapter is about two thousand words and the useful parts are in the details.

The dataset article has great decision-making philosophy without the boringness of a textbook, but still carries a lot of the important information. Note that an immediate solution definitely does not come, as there's lots to unpack — but quite worthwhile.
Pre-release readerRead a free pre-release copy

The disclaimer that matters

This teaches a process. It guarantees no outcome.

Every performance figure here is from one channel, one audience, one format family, and a corpus in the dozens rather than the thousands. Several of the findings are explicitly labelled as association rather than causation, and one whole section is about a rule I believed, published, and then falsified with my own experiment.

Nothing here promises views, subscribers, revenue, or a working business. If you apply all of it perfectly, you’ll have a more rigorous process and a better argument for why your video is the way it is. What the algorithm does with that is not something I can sell you.

The receipts

Fourteen articles that used the method before the book wrote it down

Fourteen published articles, each attached to a shipped video and a public dataset. All free, none behind an email.

Objections

The questions people ask first

Every answer below is drawn from something already written in the book, not from a copywriting exercise.

Before you decide

Chapter 7 is free,
and it’s the honest sample.

It’s the chapter about the things that cost me a day each. If that one isn’t useful to you, nothing behind the price will be either — which is the most reliable way I know to stop someone buying something they’ll regret.

  • $39.99 once, no subscription
  • $79.99 from December 15 — you are buying at the founding price
  • 30 days to change your mind, no questions
  • Lifetime updates at the same address
  • The free chapter stays free either way