Production strategies

Each abstract point gets one drawing that makes the point as a picture, so the drawing could replace the paragraph.

Why it works

A metaphor is an argument the reader can check at a glance: a developer fishing from a boat labelled “chat” above submerged assumptions says what the section says, faster, and it sticks. Decoration can be deleted without loss; these cannot. Nadeem even borrows a meme format (the midwit bell curve) when the argument is that shape.

What it costs

The expensive part is the brief: one sentence naming the metaphor, the labels and what the reader should conclude. Rendering is cheap. Gutter owns the house raccoon style (gen-image, Gemini 3 Pro, about $0.13 and 40 s an image); Muse Image is about $0.01 and 26 s; GPT Image 2.5 runs $0.10 to $0.26. magick cuts the background to transparent. Expect two or three renders per keeper.

Live examples

Plan mode is dead ↗ (seven cartoons, from the waterfall to the gravestone); Hierarchy of Money, fiat chapter (“claims for black money pile up; the stones do not”).

Pairs with

post-mortemparablefield mapone questioncase studybake-off

The same characters, drawn the same way, in every picture, with a cast page that introduces them once.

Why it works

The reader learns who stands for what once (the raccoon is the bank, the tortoise the central bank) and then every later panel reads instantly. A cast also turns drift into a checkable defect: one red claw, the right shirt.

What it costs

A reference sheet per character and a written canon, then every render is conditioned on it. Gutter and gen-image take the reference images; cheaper models drift (the Playground scores Muse and Recraft as “drifted” on style). Budget a re-render pass and a canon check per image.

Live examples

Hierarchy of Money, the cast; The Den (Larry, Igor and the attic, every week); Gutter Model Playground.

Pairs with

parablebake-offpost-mortemcase studyclose readingatlas

Every image's alt text is a sentence, or a paragraph, that says what the picture argues, not what it depicts.

Why it works

Read the alt texts alone and you should have the argument; that is the test that the pictures are doing work. It serves screen readers, search, and the next model that reads the page. Writing it also exposes cartoons that do not actually argue anything.

What it costs

Minutes per image, no tools. The Den writes panel-by-panel alt with every speech balloon quoted. Our weak spot: the chroma-key page and the Hierarchy of Money cast portraits use short labels.

Live examples

Why task state needs Dolt (six diagrams, paragraph alts); Hierarchy of Money scenes; Gutter Model Playground; Nadeem's seven cartoons.

Pairs with

anything with a picture

A process written as one line of arrows, set against a second chain, so the shape of the contrast is the argument.

Why it works

chat → disambiguate → generate spec → review spec → revise spec → approve → implement → review code against understand → act → inspect → clarify → adjust → act again. One is a waterfall and one is a loop, and you see it before you read it. The prose around it only annotates.

What it costs

Nothing but discipline: finding the fewest steps that are still true. Monospace in the page; optionally redrawn as a cartoon, which Nadeem also does.

Live examples

Plan mode is dead ↗, the two loops. Our sites use arrows only inside prose.

Pairs with

post-mortemcase studyone questionfield mapatlasclose reading

A hand-built diagram of the moving parts: which ref points where, what gets merged, where it breaks.

Why it works

Some explanations are topological. “Two agents commit the same SQLite file and git cannot merge it” is one picture and three confusing paragraphs.

What it costs

Inline SVG written by hand or by an agent, then checked in Playwright at phone width. Free to render, slow to get right; image models are bad at labelled diagrams, so this is not a job for Gutter.

Live examples

Why task state needs Dolt (rollback, dual sync, refs topology).

Pairs with

one questionfield mapcase studypost-mortemclose readingbake-off

The whole argument is readable from the top of the page: a TL;DR, or a list of reasons whose items reappear as the section headings.

Why it works

The skimmer gets the conclusion and the reader who stays gets a map of where each claim is defended. Nadeem lists four reasons he was wrong, then gives each one a section with the same heading; skim the headings and you have the post.

What it costs

An editing pass, no tools. The risk is a TL;DR that summarises the topic instead of stating the claim.

Live examples

Plan mode is dead ↗; the TL;DR box on every page of Stevey, Mapped and The Man in the Arena.

Pairs with

every kindstrongest on post-mortem and field map

Two things that differ in exactly one way, shown next to each other, so the difference is the lesson.

Why it works

A controlled comparison is the cheapest proof there is. Same person, same app, same silence gate: through AirPods only 0.35 s of a spoken “six” survives and the recogniser hears “Sex”; through a USB mic every word comes through whole. The reader draws the conclusion before you state it.

What it costs

Holding everything else fixed, which is the hard part. Then the capture: ffmpeg for audio, Playwright for screenshots, magick to lay two images side by side at the same size.

Live examples

When AirPods Eat Your Words (Exhibit A, Exhibit B); Chroma-key hill-climbing (on white against the alpha mask); Let's build GPT, part 6 (the numbers here against the finished script); Nadeem's two loops.

Pairs with

case studypost-mortemone questionfield mapatlas

A section, near the end, that argues against the page: the one caveat, the critiques, the shadow side.

Why it works

It is what makes the rest trustworthy. A page that names its best counter-argument has shown it looked for one. Nadeem's post is the counter-example: it never makes the case for when a plan still pays, and that is its main gap.

What it costs

Research time, and a writer brief that demands it; left to themselves, writers skip it.

Live examples

“The one honest caveat” on every Yegge page; “The one honest objection” on Brooks; six critiques of the Arena; the shadow section of Why We Invented Clowns.

Pairs with

field mapclose readingone questionpost-mortematlasbake-off

The source's own sharpest sentences, quoted verbatim and linked to where they appear, collected so they can be found again.

Why it works

Our prose explains, and the author's lines are what the reader remembers and repeats. A cite link on every quote keeps it honest and sends traffic back.

What it costs

Checking each quote against the original or an archive mirror. Auto-captions mangle names, so quotes from video need a listen.

Live examples

Stevey, Mapped: quotes (176 blockquotes, each cited); “The lines worth keeping”.

Pairs with

field mapclose readingparablebake-off

Every funny or pointed claim links to the real transcript it came from, one click down.

Why it works

The rule is Steve Yegge's and the Den keeps it: “If I can't produce the transcript, I don't get to draw the strip.” The joke lands harder because it happened, and the reader can check.

What it costs

A cleaned, public-safe transcript per claim, usually a gist, and a privacy pass before it ships. The receipts also have to exist before the drawing, which rules out anything made up.

Live examples

The Den (every strip carries a receipts block); Yegge's comic strip ships its receipts. Not yet used on any kind on the spine.

Pairs with

post-mortemcase studyparableclose reading

The evidence itself, in the page, one tap to play: the audio clip, not a description of it.

Why it works

Hearing “six” arrive as a 0.35-second sliver settles an argument a waveform chart cannot. The reader stops taking your word for it.

What it costs

ffmpeg to cut clips, embedded as WAV (Playwright's Chromium cannot decode AAC, so test with WAV). TTS samples cost about $0.02 for five seconds of speech.

Live examples

When AirPods Eat Your Words (six clips); Larry Voice Catalog (twenty-odd voices, each with its own player).

Pairs with

case studybake-offfield mappost-mortem

Every claim about a video links to the second it is said.

Why it works

It makes the page a better way into the source rather than a replacement for it, and every doubtful claim is one click from being checked. CS336 carries 1,618 of them.

What it costs

yt-dlp auto-captions (per video, then deduped), and one writer per chunk who verifies their own timestamps. Captions mangle proper nouns, so names get checked.

Live examples

CS336, lecture 1; Let's build GPT; Entropy, Mapped. Brooks links 166 episodes but without timestamps.

Pairs with

field mapclose reading of a speech on videoatlasbake-off

The cleaned full text ships alongside the explainer, anchored by paragraph, so explanation and source are both one link away.

Why it works

It lets the explainer stay short and original: the source is there for anyone who wants the words. For a close reading, the full text is the object of study.

What it costs

Cleaning the captions into paragraphs with timestamp anchors; Parakeet for audio with no captions. Cheap once the timestamp pipeline exists.

Live examples

CS336 transcripts; Entropy, Mapped; the full 1910 speech.

Pairs with

field mapclose readingcase study (the session log)one question

Code is annotated against line ranges in the real repo, and every data row gets its own permalink back to the line it came from.

Why it works

“Line 69 does this” is checkable; “the model does this” is not. For a dataset, a permalink per row makes the page citable at the level people actually argue about.

What it costs

A pinned commit so the line numbers do not drift, and a script to generate the anchors. The HarmBench table adds a search box and three filters over 510 rows.

Live examples

Let's build GPT, part 5; HarmBench, all 510 behaviors; CS336 assignment 4 (every deliverable linked to its test).

Pairs with

field mapone questioncase study (the diff)parable

The source is checked against itself (paper against repo, handout against code, the famous quote against the first printing) and every disagreement is shown, not smoothed over.

Why it works

It is where a derived page earns more trust than the source. Finding that the paper says 40/20 and the CSV says 41/19 tells the reader someone actually looked.

What it costs

A verification pass with a link sweep (“33 URLs checked, 32 live”). It is slow, and it is the part agents skip unless the brief makes it the acceptance test.

Live examples

HarmBench, where the sources disagree; the Arena against the 1910 printing; micrograd (notebook against video); Robert Moses Atlas (two of Caro's claims marked disputed).

Pairs with

field mapclose readingatlasparableone question

One frozen input through every candidate, outputs at the same size in a grid, with what each one really cost and how long it took.

Why it works

Only the tool varies, so the grid is the verdict. Measured cost per item ($0.010 to $0.255 an image on the Playground) beats a list-price table, and latency is often the deciding column.

What it costs

The renders themselves (the Playground cost about $0.96 for ten images), OpenRouter's billed cost field, and a stopwatch per call. Playwright to check the grid at phone width.

Live examples

Gutter Model Playground; Larry Voice Catalog; Chroma-key (each keying approach side by side); The Den, same script, three models.

Pairs with

bake-offone questionparableclose readingatlas

A written rubric or formula applied to every attempt, shown per attempt, so “better” means a number going down or a checkbox turning green.

Why it works

It turns taste into an eval. Chroma-key's score = residual × 5 + holes goes from 17,385 to 269; the Playground's five-point canon check (style, claw, beard, shirt, Crocs) makes model drift visible.

What it costs

Designing the metric, then scoring every output: a script for masks (magick), a zoomed-in look for canon checks. The metric can be gamed, so show the images next to the score.

Live examples

Chroma-key hill-climbing; Gutter Model Playground.

Pairs with

bake-offone questioncase studyclose readingparable

One slider, playhead or step control that drives everything on screen, so the reader moves through time or through the steps themselves.

Why it works

Sequence becomes something you feel. Dragging from 1924 to 1968 and watching the expressways appear is the argument about accumulated power. On the parable, stepping through a chapter one move at a time shows the balance sheet change with each transaction.

What it costs

Real engineering: a dated, sourced event list, Leaflet or inline SVG, requestAnimationFrame, keyboard controls. Days, not hours, and Playwright tests for every breakpoint.

Live examples

Robert Moses Atlas (drag, play, arrow keys); Hierarchy of Money walkthrough (a 0 to 10 step slider). In the wild: The Mongol Conquests ↗ by Yanqing Cheng (drag, play, a month or a year per keypress, event to event).

Pairs with

atlasparablecase studypost-mortemclose reading

The answer as a live calculator with every assumption an editable field and every constant cited, and the arithmetic unit-tested.

Why it works

The reader's situation is never the example's. Letting them swap in their own numbers, with charts that redraw, turns a one-off answer into a tool, and the tests are why they can trust it.

What it costs

Plain JS and canvas charts, a Node test file per module (35 tests on cap-gains), and a citation for every constant. A day or two.

Live examples

Sell now or sell later (three sliders, eleven inputs, two charts).

Pairs with

one questionfield map (the napkin math)bake-off (cost at your volume)close reading

The scrubber’s track is itself a chart: chapter bands on top, one tick per event beneath, so the axis is the table of contents.

Why it works

The reader sees the shape of the story before playing it: where events bunch, where nothing happens, which chapter is long. On the Mongol atlas, ticks are coloured by kind and sized by importance, and the late-1240s lull shows at a glance.

What it costs

A dated event list with a kind and an importance per row, which you need for the cards anyway. The render is a row of divs or an SVG. On a phone the strip needs to scroll sideways inside itself rather than squeezing ninety years into 350 px.

Live examples

The Mongol Conquests ↗ (nine chapters, 103 ticks); Robert Moses Atlas (six chapters over its timeline).

Pairs with

atlascase studypost-mortemclose reading (the afterlife)bake-off

How sure each mark is shows up in the mark itself: a point when the place is known, a dashed region when it is not, a certainty grade and a one-line reason on every card, and a stated list of what was deliberately not drawn.

Why it works

A map or a timeline looks equally confident about everything unless you make it do otherwise. The Mongol atlas pins 35 events to a city and draws 63 as regions, grades each one high, medium or low, and leaves out army routes because a confident arrow across a steppe would claim more than the chronicles do. The reader trusts the sure marks more because the unsure ones admit it.

What it costs

Two extra fields per event (location precision, certainty) and a sentence on why, filled during research, not after. The honest part is resisting the arrow you would like to draw.

Live examples

The Mongol Conquests ↗. Ours, partly: the Robert Moses Atlas marks two of Caro’s claims as disputed and dashes unbuilt roads, but does not grade every card.

Pairs with

atlasclose readingcase studyfield mapparable

A small panel that states what the playhead means right now, as a number or a name that changes while you scrub.

Why it works

The map shows the change; the readout gives it a unit. “About 2 million km² in 1206, about 16 million by 1241” and the reigning Great Qan turn a spreading colour into a claim you can quote. The Moses Atlas does the same with the count of offices he held at once, which is its argument about power.

What it costs

One derived series computed from the event data (area, count, balance), and the discipline to show only one or two numbers.

Live examples

The Mongol Conquests ↗ (month, Great Qan, km²); Robert Moses Atlas (offices held at once); Hierarchy of Money walkthrough (the balance sheet at each step).

Pairs with

atlasparablecase studyclose readingbake-off