The work behind the lessons
How we build Quarry School
Read a lesson, work through an example and hear it spoken. Here is how we turn a course outline into those pages, and what we check before we publish.
We build Quarry School at Manic Agency. Quarry School is part of Quarry, made by Manic Agency LLC.
Files a model can write and a script can check
We use OpenQuarry, an open format for knowledge bases and courses kept as Markdown. A fabric holds the collection. A weave represents a course, a loom a unit, and a thread a lesson.
Quarry reads the OpenQuarry format. Our School exporter writes course and unit metadata, with lesson threads that link back to the School. Those threads contain the lesson address, rather than a copy of its text. The export is kept as a CI artifact; it is not a published lesson download.
One real source file is content/college-algebra/quadratic-functions.json. It holds the lessons, their examples and practice. Our exporter maps that unit to weaves/college-algebra/quadratic-functions/loom.yaml.
From an outline to a lesson you can use
Choose the scope Person’s decision
We choose what to teach, which sources may be used and what a lesson must contain. The outline names each lesson and what it teaches.
Draft the lesson AI agent + script
A writing agent turns the outline into plain explanations, worked examples and exercises. Scripts check the file’s structure. Source records and citations travel with the content.
Recompute and check AI agent + scripts
A separate engine checks results in the evidence pass. Each result is tied to the item’s content hash. Answer checkers test accepted and rejected answers. A failed evidence check stays in the report until the item is fixed or replaced.
Read the whole unit Independent AI reader
A reader from a vendor that did not write the unit reads it with its evidence and report. Its sign-off records the exact content hash. This is an AI review, not a claim that a person has read every lesson. The owner decides cases reserved for owner clearance.
Gate publication Scripts + release decision
The publication gate checks the candidate edition, evidence and sign-off. Release checks include the built pages, accessibility and browser behavior. We publish the edition that passes the required checks.
Render and hear the narration Speech models + scripts
We turn lesson cues into speech. The clip records bind recordings to the displayed words. Scripts check files, durations and hashes. Speech-to-text checks help find skipped symbols and cut sentences; flagged clips get a plainer spoken reading and another render.
Correct and release again Team decision + checks
We correct the source, repeat the affected checks and obtain a fresh sign-off when the content changes. A unit can be withdrawn from publication while a correction is made. Changed narration cues need matching recordings.
For a learner’s guide to each check and what it cannot establish, read How lessons are made.
What the records show about cost
These are samples of specific jobs, in US dollars. They are not an all-in price for a lesson.
| Unit | Characters | Speech cost |
|---|---|---|
| Complex numbers | 66,305 | $1.99 |
| Dividing polynomials | 79,003 | $2.37 |
| Graphs of polynomial functions | 103,626 | $3.11 |
| Power and polynomial functions | 106,110 | $3.18 |
| Zeros of polynomial functions | 114,947 | $3.45 |
| Quadratic functions | 119,957 | $3.60 |
That is $1.99–$3.60 per unit for the initial spoken text, at the recorded Aura-2 rate of $0.030 per 1,000 characters. Repeated text can reuse a clip. These estimates exclude later additions, pronunciation repairs, speech-to-text checks and hosting. They describe those versions of the units.
The content ledger from 27 October to 4 November 2026 records $1.50–$4.17 per research run, $1.32–$7.57 per writing run and $1.23–$5.30 per checking run. These are separate API jobs, including repeat runs. They do not include narration, delivery, subscriptions or a full unit’s development.
Our coding and writing agents also run on subscriptions. We do not split that subscription cost per lesson. API credits offset eligible usage; a recorded usage charge is not necessarily cash paid.
Time includes review and release
We measured additions to four existing units, from the first content commit to the recorded live release. This includes waiting for review, checks and deployment. It excludes planning before that commit and the original unit’s creation.
| Unit | First commit to live |
|---|---|
| Quadratic functions | 2 h 20 min |
| Dividing polynomials | 9 h 34 min |
| Zeros of polynomial functions | 11 h 1 min |
| Graphs of polynomial functions | 5 h 55 min |
The range is 2 hours 20 minutes to 11 hours 1 minute. These are elapsed delivery times for those changes, not a promise about how long a new unit takes.
Why we can offer the lessons free
We fund the build as development work. Generative AI drafts explanations, makes exercises and helps check results; speech models narrate the lessons. API credits and Deepgram startup credits support that work. Hosting is an ongoing running cost.
We plan to use non-intrusive sponsorship placements to help pay for hosting. That is a funding plan, not a claim that sponsorship pays the bills today. AI tutoring is a planned part of the product; the process described here covers the lesson files and checks we have built.
Questions about the build
Is it free, and why?
Yes. We pay for the build as development work, supported by API and speech startup credits. We cover hosting while we develop the sponsorship plan.
Who builds it?
We are the team at Manic Agency. Our Terms state: Quarry School is part of Quarry, made by Manic Agency LLC.
How much does a lesson cost to make?
We do not have an all-in per-lesson figure. The recorded narration estimates are $1.99–$3.60 per unit. Research, writing and checking have separate per-run costs, and subscriptions are not allocated per lesson.
How long does a unit take?
The table above measures changes to existing units from their first content commit to live release. A new unit also needs planning, sources, drafting and a full independent read.
Which AI models and voices do we use?
Our content configuration names Claude Code with Opus for writing and Gemini 3.1 Pro High for recomputation and paraphrasing. Individual reviews can name a different model: the Quadratic Functions sign-off records Gemini 4 Argon High. The narration record names Deepgram Aura-2 Athena for speech and Nova-3 for speech-to-text checks. These are recorded configurations, not a promise that every run uses the same model.
What does a person decide?
We choose the scope, source policy, product design and funding. The owner rules on clearance cases reserved for them. The independent content reader can be an AI agent; its recorded sign-off tells us what was checked.
What is planned for money?
We plan non-intrusive sponsorship placements to help cover hosting, and later a paid plan for the AI tutor. Lessons, practice, narration and your progress stay free.