← Home 03 / The playbook

Craft that scales — because it is encoded, not guarded.

The Claude Learn Playbook is how one person's teaching craft becomes a tutor that runs without them in the room. It splits the work across two time-scales: humans get world-class at exactly two artifacts, and Claude generates every lesson in the moment, on the learner's real work.

● Docent note

This is the JD's line — “encodes its own craft into AI-powered systems rather than guarding it” — as an architecture. Watch where the human is, and is not: humans author the exemplar and the rigor rubric at design time; Claude generates the lesson at the moment of need. The human is removed from the loop at generation, but their judgment is not — it is encoded in the two artifacts the generator is held against.

The core move

Two time-scales

Design-time

Humans + the content team — infrequent, high-craft

The team gets world-class at two artifacts per cell: an exemplar (what a great lesson looks like — the judgment a rubric can't fully pre-state) and a rigor rubric (the checkable standard, with observable indicators and a load-bearing / auto-ship / halt escalation policy). Plus the standards binding, the learning-structure templates, and the locked source-of-truth contract.

Generation-time

Claude — per learner, per moment, on real work

Claude generates one lesson, in situ, grounded in the source-of-truth, produced against the exemplar + rubric, tied to a named standard, centered on the learner's actual current task. No lesson is a replica: swap the learner and the doing changes.

“producing educational content exemplars across products and audiences as well as producing rigor rubrics for the product × audience educational content — that is how we remove the human in the loop at lesson generation in-situ moment.”

— Krystal Martinez, the design ruling that defines the two artifacts
The atomic unit

Say · See · Do

Every lesson is 3–6 Say-See-Do cycles, each teaching one direct-instruction point (credited to Breakthrough New York, 2012).

Say

Explicit instruction

One point, stated plainly and grounded in the locked source-of-truth — never ungrounded model knowledge taught as fact.

See

An anchored visual

An image, data-viz, or example of that same point — ground-truth-verified, because an inaccurate visual anchors a wrong mental model.

Do

On the learner's real work

Implement it on the task actually at hand, with on-demand evaluation. Then independent at-bats, scaffolding faded to the learner's demonstrated level.

Each lesson carries one Bloom's-leveled objective tied to a named AIHC standard, an exit ticket that climbs Bloom's to that objective, and a write to the learner's ledger. See all of it built out across forty-eight lessons.

Quality without a human on every lesson

The rubric is the gate

Because the human is not in the generation loop, quality is enforced by the design-time artifacts. Each generated lesson is produced against, and checked against, the cell's rigor rubric — a list of observable indicators, each marked load-bearing or not. The escalation policy is explicit and versioned:

All load-bearing pass

Auto-ship.

Any load-bearing fail

Halt for human review.

Non-load-bearing miss

Resolve-before-serve queue.

The line moves by policy revision as models improve — never by vibe. The four demo cells each ship a Fable-authored exemplar + rubric pair; read them alongside the lessons in Tracks.

The two generation-time inputs

A locked source-of-truth, and a learner profile

The doc-set (SSOT) contract

Which source documents are authoritative for this cell, frozen with access dates and freshness tracking. Every product claim in a lesson traces to it; where the docs are thin, the lesson says so rather than inventing.

The learner profile

Demonstrated level per standard, the work at hand, elected cultural referents, and declared accommodations. Personalization is read from here — and compounds, lesson over lesson, in the ledger.

Culturally responsive by election, not inference

Cultural referents are elected by the learner from a database — never inferred from identity. Referents are flavor and bridge; every load-bearing claim still stands on the source-of-truth. Anti-essentialism is structural, not a disclaimer.

Memory & measurement

The ledger, and what counts as working

The ledger

The tutor's memory and the learner's mirror: mastery tracked per standard, auto-score and self-score both recorded (their gap is a metacognitive signal), the learner's own journal never displaced. Personalization compounds across lessons.

Measurement

Success is transfer on real work, mastery-by-standard over time, and the calibration gap — not completion, not satisfaction. Content is either current and performing, or it comes down: a freshness beat audits the sources on a cadence.

See the ledger instrument, with its honest defect flags, at the end of each track in the proof.

A shipped pattern, not a hunch

The same shape as Claude Design.

Claude Design points at a brief + a design system and generates the screens; Claude Learn points at a source-of-truth + a learner profile and generates the lesson — the human's craft encoded once into the system that generates, rather than spent per artifact. This very site was built that way: a design system in, pages out.