IDVORKIN-AI-TOOLS

Kinds of explainer

Not every good explanation is the same shape. This is the spine we build against: eight kinds, what each one is for, the moves that make it work, and an example of each that we think is good, with the reasons spelled out. Our own sites are mapped onto it; the kinds we have not yet built or found a great example of say so. The kinds are the structure; the craft moves that can go with any of them (metaphor cartoons, receipts, timestamped sources, scrub controls) are on a second axis, the production strategies, with a grid of which pairs have been tried. Back to the index.

  1. 01The field mapA whole subject, laid out unit by unit, sources pinned.
  2. 02The one-question explainerOne question, answered properly, working shown.
  3. 03The post-mortemI believed it, I built it, it died. Here is why.
  4. 04The case studyOne real incident, the evidence, the fix, the lesson.
  5. 05The close readingOne short text everyone quotes, read line by line.
  6. 06The atlasA subject on a time axis you can scrub; on a map when place matters.
  7. 07The illustrated parableA story that carries the model; a page per idea after.
  8. 08The bake-offSame input through N tools, side by side, real costs.

Kind 01 11 of ours

The field map

Take a course, an author or a subject and lay it out page by page, every page linked back to the source it came from. The reader is entering a field and wants the territory, not one trail.

For

Working through a course or a body of writing in order; coming back to one unit months later and finding it where you left it.

Signature moves

  • A spine: arcs of units, each a row, with a progress log.
  • One page per unit, original prose, not a paraphrase of the source.
  • Every claim linked to a timestamp, a page or a line.
  • The cleaned transcript or full text shipped alongside.
  • Links swept before shipping; discrepancies with the source shown, not hidden.

Production strategies its exemplars useS06 Ten-second skeletonS07 Before and afterS08 Honest objectionS09 Best lines, citedS12 Timestamped linksS13 Transcript pageS14 Line-linked code and dataS15 Checked against the original

Ours

CS336, MappedSeventeen lectures, deep dive per lecture, transcripts, assignments.FIELD MAPcourse Technical AI Safety, MappedSix units, twenty-seven chapters, every reading linked back.FIELD MAPcourse Entropy, Mapped3Blue1Brown's compression series in twenty pages, formulas reconstructed.FIELD MAPvideo series Let's build GPT, MappedOne two-hour lecture as eight parts, code annotated line by line.FIELD MAPsingle video micrograd, MappedBackprop from scratch, eight parts, the notebook annotated.FIELD MAPsingle video HarmBench, MappedA benchmark browsed by harm category, 510 behaviors permalinked.FIELD MAPdataset Stevey, MappedOne page per Yegge essay, the source pinned, the best lines quoted.FIELD MAPauthor Arthur Brooks, MappedOne page per idea from the happiness-science podcasts and books.FIELD MAPauthor AI Safety — A Field MapSubfields, who works on what, where to start.FIELD MAPsubject Red Teaming ModelsThreat models, attack families, evals, how the field is organised.FIELD MAPsubject Rails, Ropes & RackEverything that runs on a cable, a rail or a rack.FIELD MAPsubject

In the wild

Exemplar wantedA field map by someone else that we think is done right. None picked yet.

Kind 02 4 of ours

The one-question explainer

One question in the title, one answer, and the working shown so the reader can check it. The reader has this exact question and wants to stop having it.

For

Questions you will ask again: a tax timing choice, why a tool needs a database, how to get a clean alpha mask, why every culture invented a clown.

Signature moves

  • The question is the title; the answer is in the first screen.
  • The evidence is shown: numbers, sources, unit tests, the actual artifact.
  • The assumptions are listed so the reader can swap theirs in.
  • Short. One page, no spine.

Production strategies its exemplars useS03 Full-sentence alt textS05 Mechanism diagramsS07 Before and afterS08 Honest objectionS16 Bake-off sheetS17 Score you can climbS19 Tested calculator

Ours

Sell now or sell laterThe capital-gains timing question, sources cited, unit tested.EXPLAINER2026-08 Why We Invented ClownsThe licensed pressure valve, the lineage, the evidence, the shadow.EXPLAINER2026-07 Why task state needs DoltHow a task database and a code repo share one fork.EXPLAINER2026-04 Chroma-key hill-climbingTreating the chroma key as an eval: score the mask, then climb.EXPLAINER2026-04

In the wild

Exemplar wantedA one-question explainer by someone else that we think is done right. None picked yet.

Kind 03 1 in the wild

The post-mortem

The writer believed something hard enough to build on it. It failed. The essay autopsies the specific failure to earn a general claim, and is honest about the part that is still unsolved. It is an argument disguised as an obituary.

For

Learning from someone else's expensive mistake before making it yourself. The reader gets the conclusion and, more valuable, the reasoning that was wrong.

Signature moves

  • Open on the belief, with the stakes attached, not on the topic.
  • Split the thesis: what died and what still matters.
  • List the reasons up front, then one section per reason with the same heading.
  • Before-and-after in the fewest possible symbols: an arrow chain, a diagram.
  • Name the mistake in one sentence. End on the open problem, not a pitch.

Production strategies its exemplars useS01 Metaphor cartoonsS03 Full-sentence alt textS04 Arrow chainsS06 Ten-second skeletonS07 Before and after

In the wild

Plan mode is dead ↗Ayman Nadeem, September 24, 2026, about 1,900 words

Kind 03 · The post-mortem

Nadeem built Nuanced, a desktop coding app whose whole premise was a persistent plan produced before implementation. Users did not want the spec, models stopped needing it, and the plan-then-build split broke the way people actually think. The post says why, and what is left.

What is good about it

  • It opens on a belief with stakes, not a topic. First sentence: Earlier this year, I believed planning was going to become the most important part of building software with AI. Second paragraph: he built and launched a whole app on it. The title already gave away the ending, so the reader reads for the why.
  • The thesis is a split, which keeps “dead” from being clickbait. Plan modes did two jobs, precise instructions for the agent and understanding for the human; #1 is rapidly becoming obsolete as models get better. I think #2 matters more than ever, but plan modes are the wrong abstraction for it. Every later section serves one side of that split.
  • The skeleton is readable in ten seconds. “Why I was wrong” lists four reasons as bullets (conflated planning with a plan; models got really good; no one wants to read AI-generated text; separated planning from building), and then each bullet becomes a section with that heading. Skim the headings and you have the argument.
  • Arrows do the work of paragraphs. The old loop chat → disambiguate → generate spec → review spec → revise spec → approve → implement → review code against the new one understand → act → inspect → clarify → adjust → act again. The contrast is the argument; the prose around it only annotates.
  • Every abstract point gets a drawing that carries the metaphor. Seven hand-drawn cartoons: fishing from a boat labelled “chat” above submerged code and assumptions, tumbling down a waterfall of sequential steps, mourners at a gravestone reading “Plan Mode, 2025–2026”. They argue rather than decorate, and each one has a full-sentence alt text.
  • It names the mistake in one line and ends on what is unsolved. The biggest mistake I made was turning the plan into an artifact instead of designing a process for improved human understanding. The last section is titled “We still haven’t solved the understanding problem”, and it means it: no pitch, no five tips, a hard question left standing.

What it leaves out, on purpose. No usage numbers, no screenshots of Nuanced, no case for when a plan still pays (large irreversible changes, an unfamiliar codebase), and Claude Code's plan mode is a name-drop, not a comparison. The kind trades breadth for one confessed mistake; that is the trade.

Ours

Exemplar wantedWe have not written one of these yet. When a thing we built dies, this is the shape to write it in.

Kind 04 1 of ours

The case study

One real incident, told with the evidence: the symptom, the wrong guesses, the measurement that settled it, the fix, and the one general lesson. Where the post-mortem is about a belief failing, the case study is about a system failing.

For

The reader who will hit the same bug next month and search for the symptom.

Signature moves

  • The symptom first, in the words you would search for.
  • The evidence shown, not described: the audio, the log, the diff, the plot.
  • The hypotheses that were wrong, and what ruled them out.
  • One lesson that survives the specifics.

Production strategies its exemplars useS07 Before and afterS11 Playable exhibits

Ours

When AirPods Eat Your WordsHow a Bluetooth AirPods microphone destroyed on-device ASR, shown with the actual audio.EXPLAINER2026-07

In the wild

Exemplar wantedA debugging case study by someone else that we think is done right. None picked yet.

Kind 05 1 of ours

The close reading

One short text that everyone quotes and nobody has read, put back in its context and read clause by clause, with its afterlife and its misattributions checked.

For

A speech, an essay, a poem, a paragraph that has become a slogan; the reader wants to know what it actually says.

Signature moves

  • The full text shipped, with the famous part marked.
  • Clause-by-clause annotation of the quoted passage.
  • The afterlife: who quoted it, for what, and whether they read it.
  • Every quotation checked against the original printing; the folklore tested and, where it fails, said so.

Production strategies its exemplars useS06 Ten-second skeletonS08 Honest objectionS09 Best lines, citedS13 Transcript pageS15 Checked against the original

Ours

The Man in the Arena, MappedRoosevelt's 1910 speech around its most-quoted paragraph; every quote checked against the 1910 printing; the Mandela story found wanting.FIELD MAPfiled as a map; reads as a close reading

In the wild

Exemplar wantedA close reading by someone else that we think is done right. None picked yet.

Kind 06 1 of ours · 1 in the wild

The atlas

A subject laid on a time axis the reader can scrub, and on a map when place matters too. Where and when at once, with a card behind every mark. A timeline is this kind with the map left out; it does not get its own slot, because the structure is the same with or without the geography: a dated axis, chapters that set it for you, and a sourced card behind every event.

For

Anything whose story is really a story of sequence, and often of place: a career, a city, a network, a war, an author's body of work, a quotation's afterlife.

Signature moves

  • One scrub control that drives everything on screen.
  • The axis carries the data: chapter bands above it, one tick per event, so the busy and quiet years show before you press play.
  • A readout that says what the playhead means: who rules, how much, what it cost.
  • Every mark opens a sourced card that says how sure it is.
  • Chapters that set the scrubber for you, so it still reads as an essay.

Production strategies its exemplars useS15 Checked against the originalS18 Scrub controlS20 The event stripS21 Uncertainty drawnS22 A running readout

Ours

Robert Moses AtlasScrub from 1924 to 1968 and watch the parkways, bridges and expressways appear, with the offices he stacked and the neighborhoods it cost. Six chapters, every card sourced.ATLAS2026-09

In the wild

The Mongol Conquests, 1206–1294 ↗Yanqing Cheng, Atlas of Wars in Qing’s Workshop, September 2026; 103 events in nine chapters

Kind 06 · The atlas · announced as the Opus 5.5 edition

A globe with a timeline strip along the bottom. Drag or play eighty-eight years, from Temüjin’s enthronement on the Onon to Qubilai’s death, and watch Mongol rule spread and then split into khanates, with a researched card behind each of 103 events.

What makes the timeline work

  • The time axis is the table of contents. Nine chapter bands sit over the years, with one tick per event under them, coloured by kind and sized by importance. Before you press anything you can see where the fighting was dense and where it thinned out in the late 1240s. You can drag, play, step a month with the arrow keys or a year with Shift, or jump event to event with the bracket keys.
  • One playhead drives everything, and a readout says what it means. Territory fills in and the globe turns to follow the action (a Follow toggle turns that off). A panel states the month, the reigning Great Qan and the area under Mongol rule or tribute, about 2 million km² in 1206 and about 16 million by 1241. Crossing a band swaps in that chapter’s paragraph, so scrubbing still reads as an essay.
  • Every card says how sure it is, and the map draws it. A card gives the kind, date, place, both sides with their commanders, a summary, a longer account, the significance, a high, medium or low certainty grade with a note on why, and sources. The map is just as honest: 35 events pin to a named city, 5 to a battle locality, and 63 are dashed regions because the sources support a place but not a point.
  • It says what it refused to draw. There are no army routes, and the About page explains why: a confident arrow across a steppe would claim more than the chronicles do. Territory is shaded as ruled, vassal or contested, with the warning that frontiers were zones of tribute and raiding, not lines.
  • It holds up on a phone. At 390 px the panel shrinks to the year and chapter, the strip scrolls sideways inside itself, the card becomes a bottom sheet with previous and next links, and nothing overflows the page.

What it leaves out, on purpose. No routes, no quoted translations (the prose is the author’s own synthesis over a fixed research pack of 109 candidate events, 103 kept), and no argument: it is an exhibit, not a thesis about why the Mongols won. A second take on the same subject, war atlas B, sits beside it. Our page links to both rather than copying anything from them.

Where a timeline view would pay next

Stevey, Mapped: best candidateTwenty years of essays filed as a list. Lay them on a strip with a chapter per job, from the blog rants to the 2025–26 agent essays, and the list becomes a career arc, long quiet stretch included. S20 and S22 (what he was building at the time).FIELD MAP → ATLAS VIEWno map needed The Man in the Arena: the afterlifeA hundred years of people quoting one paragraph. A strip of who quoted it and when, each grade saying how well the attribution holds up, turns the afterlife page into something you can scrub. S20 and S21.CLOSE READINGsmall data, one day Robert Moses Atlas: upgradeIt already has the strip, the chapters and a readout of offices held. What it lacks is per-card certainty and approximate places drawn as regions, not lines. S21.ATLASalready the kind

Kind 07 1 of ours

The illustrated parable

A story that smuggles the model in: characters, a village, a problem, and each turn of the plot introduces one concept. The concept pages come after and point back into the story.

For

Systems with several interlocking ideas that are dull as definitions and obvious once you have watched them happen to someone.

Signature moves

  • The story first, illustrated, in the source's own wording where it has one.
  • One page per idea the story introduced, in the order the story introduced it.
  • Every concept page links back to the panel where it happened.
  • A cast that stays consistent across every drawing.

Production strategies its exemplars useS01 Metaphor cartoonsS02 Recurring castS03 Full-sentence alt textS18 Scrub control

Ours

The Hierarchy of Money, MappedGregory Gundersen's village parable, his wording, our animals, then a page for each idea: stones, bank money, reserves, reserve currency, fiat.FIELD MAPfiled as a map; the story is the spine

In the wild

Exemplar wantedAn illustrated parable by someone else that we think is done right. None picked yet.

Kind 08 2 of ours

The bake-off

The same fixed input through N tools or models, side by side, with what each one cost and how long it took. The reader is choosing.

For

Picking a model, a voice, a renderer, a library, without trusting anyone's benchmark table.

Signature moves

  • One input, frozen, so the only variable is the tool.
  • Outputs side by side at the same size.
  • A canon check: the details the output must get right, scored.
  • Real cost and latency per item, not list prices.

Production strategies its exemplars useS02 Recurring castS03 Full-sentence alt textS11 Playable exhibitsS16 Bake-off sheetS17 Score you can climb

Ours

Gutter Model PlaygroundNine image models draw one scene; per-image cost, latency and a canon check.PLAYGROUND2026-09 Larry Voice CatalogCandidate TTS voices for Larry, side by side.SAMPLES2026-04

In the wild

Exemplar wantedA bake-off by someone else that we think is done right. None picked yet.

The spine is a working taxonomy, not a law. A site can be filed under one kind on the index and belong to another here; two of ours are. When a new kind earns a slot, or an exemplar in the wild is found for an empty one, it goes here. Suggest one by opening an issue on this index's repo.