Resources for teachers / supplement

How it’s built

How this site is built, in plain language

This is the supplement to the talk “Fast Tools, Slow Thinking.” The talk is about teaching. This page is about the machinery, for anyone who wants to know what was used, what it cost, and how the passwords are kept. Every number below is either from the talk or from a file in the site’s repository, and the file is named next to it. Where something was never measured, it says not measured. Nothing is estimated.

The pipeline in brief: write, check, fix, install

Every lesson goes through the same four steps, and the check is done by something that did not write the lesson.

  1. Write. A strong AI model drafts the lesson from a packet of source material (twenty years of lectures and, for some courses, research on the scientists involved). The writer is told what is in scope and what must not be invented.
  2. Check, independently. A different AI agent, with web access, is told to assume the draft has at least three errors. It rules on every factual claim: confirmed, imprecise, wrong, or could not be checked. It proposes exact edits. Small scripts check what a script can check better than any model: that a protein structure ID points at the right molecule, and that the answer keys are not all in the same order.
  3. Fix. A script applies the proposed edits, and the lesson checker runs again to make sure the edits did not break anything (for example, by making a section too long).
  4. Install, then a human reads it. The lesson is added to the site. The build refuses to publish a lesson that fails its checks. An expert (here, the course author) reads it, which takes about 1 to 2 hours per lesson. That last step is the bottleneck, not the writing.

What the checking found

For the Cell Signaling course (44 lessons), the checker ruled on 3,468 claims. It found 49 wrong, 453 imprecise, and 287 it could not verify, and 841 edits were made. (In the talk this is rounded to “about 3,300 claims”; the repo’s tally is 3,468.) A separate cross-check of the Genetics course produced 80 corrections, for example a figure for the share of DNA in trisomy 21 corrected from 1.5% to 0.8%.

Earlier, on the first 11 Cell Signaling lessons, the checker found that most wrong claims came from the unpublished book the source packets drew on, not from the AI writer: a reversed anecdote, wrong dates, misattributed credit. No answer key was wrong in any of the 88 questions checked. Source: docs/expansion/bakeoff-1/results.md.

None of this replaces a person. An AI checker can be wrong, and it cannot tell you which of two defensible answers a teacher wants. Expert review is the scarce resource.

Which models and tools are used for what

The site ran a head-to-head test (“bake-off”) before choosing. The same two lessons were each written by three models and scored blind by two judges (source: docs/expansion/bakeoff-1/results.md).

JobModel or toolWhy
Lesson spec and source packetClaude Opus, plus scriptsJudgment about scope and order
Lesson draftClaude Opus, or Claude Sonnet where budget mattersOpus scored 8.7 of 10, Sonnet 8.0, Haiku 5.7. The ranking was the same for both lessons and both judges
Lesson prose from the smallest model (Haiku)Not usedQuestions scored 3 to 6 of 10 and it gave a wrong database ID
Fact-check with web accessClaude SonnetSharper than Haiku and used fewer tokens
Second fact-check on important lessonsClaude Haiku or a different modelCaught two things the first checker missed
Outlines of lecture decks, simple key checksClaude HaikuGood enough and mechanical
Films and simulationsClaude Sonnet, after an Opus pilotThe first Sonnet films made to the revised brief were as good as the Opus pilot

Other tools, from explainers/README.md and explainers/_pipeline/:

What it cost

Here is what the repository actually recorded. Tokens are the units AI models are billed in.

ItemRecordedSource
Writing plus fact-checking one Cell Signaling lessonAbout 440,000 to 480,000 tokensbakeoff-1/results.md, sections F and G
One fact-check pass on one lessonAbout 250,000 to 313,000 tokenssame
One film, written by SonnetAbout 430,000 tokens (Opus pilot film: 411,000)same, section H
One passage-writing test run106,000 to 130,000 tokensbakeoff-0/results.md
Overhead of starting any agent callAbout 60,000 tokens before it does any workPLAN.md, bakeoff-0/results.md
Gemini narration of one filmUnder 20 cents (about 15 to 17 cents)bakeoff-1/results.md; explainers/_pipeline/build_audio.py
Building the whole 44-lesson Cell Signaling course in dollarsNot measuredThe plan lists “cost per lesson” as a goal; the repo has token counts, not dollars
Dollars per lesson, for any courseNot measuredToken prices were not recorded
ElevenLabs narration in dollarsNot measuredIt draws on a monthly quota; no per-film figure was recorded
Image generation in dollarsNot measured
Cloudflare hosting in dollarsNot measured
The author’s own timeNot measured, except expert review at about 1 to 2 hours per lessonPLAN.md

For scale: the first 30 lessons took about 17 weeks (June 12 to October 10, 2026). The next 173 or so took about 8 days (October 11 to 18). The site now has 203 lessons in six courses. A grant request for $10,000 in July 2026 said $5,000 would buy three to five complete lessons. It was not selected, and the site was built anyway. The reason, from the project plan: a student’s preparation should not depend on what the family can pay.

If you want to build with an AI model yourself

Start with the Gemini API. Its documentation walks through getting access and making a first request, and it has a free tier that is enough to try things out. Follow your district’s rules on which AI tools may be used with student work.

Do this on a teacher budget

You do not need any of the above to get the benefit. The part that matters most for a classroom is the pre-work, and it costs close to nothing.

← Back to resources for teachers