Alvarmethod: one teacher, one mind — how the loop actually helps — Vas

I watched Eero Alvar’s How I Use AI to Learn Things and wanted the same teacher on the agents I actually use — Claude Code, Codex, Grok, Pi, OpenCode.

The video’s claim is simple. A course built for a crowd cannot sit on your edge. A pile of tabs makes the brain hedge. One interface, many sources, is the fix. Struggle stays in the material. The system eats logistics.

I shipped that as Alvarmethod. This note is how it actually helps, not just what the folder names are.

One lamp, one notebook, arcade cabinets in the dark — One lamp. One notebook. The cabinets stay in the dark.

What was leaking

Two directions of waste:

One outlet → many students. The path is never fitted to what you already hold, so you sit through stuff you know and bounce off stuff you cannot take yet.
One student → many outlets. Switching notation, trust, and UI is cognitive work that is not the subject.

Alvarmethod is the one teacher. It does not mean one perspective. The teacher is an interface that is allowed to pull sources — and it has to show its work before you accept a fact.

The loop, as it actually runs

1. Probe — find the edge, don’t guess it

The agent is not allowed to dump A/B/C/D in chat. It has to open the harness quiz UI (ask_user_question on Grok/Codex, AskUserQuestion on Claude Code, question on OpenCode, Pi’s quiz / ask-user tool).

It starts wide, then binary-searches every strand the lesson will need. “I don’t know” is a real answer. A right letter with no reason is only edge, not known.

A topographic map with a gold tear at the cliff — the edge of understanding

I ran this on synthetic biology. I held the frame — DNA → RNA → protein, promoter, chassis, DBTL. I blanked on Gibson, RBS, and inducible promoters. That map is the whole point. A normal chatbot would have started at “what is DNA.”

2. Plan — force the model to finish the reasoning

Before any teaching, it has to draw a mermaid DAG for this mind. Showing the graph is not decoration. It is how you stop the model winging a survey.

Brass nodes with one path lit through the graph

On synbio the plan started at the cassette (promoter–RBS–CDS–terminator), not at a textbook chapter 1. The graph is written into .alvar/sessions/ so another agent can resume.

3. Teach — one stone, then a quiz

This is the ChatGPT failure mode the video names: the model gets excited and rushes the whole course. The skill forbids that. One node. Stop. Quiz that node. Advance only on lock-in. Fail → stay, or insert a prerequisite.

Stepping stones, only the nearest lamp lit

Questions mid-step are allowed. The agent does not “finish the slide” over you.

How the other skills earn their keep

learn-visual is why pictures don’t stay slop. The agent writes an SVG, looks at it, and edits until the picture makes the claim. Same loop Eero used with subagents — portable, no Pi-only extension.

A drawing locking into a single geometric idea

learn-verify is how trust gets engineered. Empirical or named facts get a verdict (confirmed / qualified / contradicted / unknown) before they are taught as fact. No invented papers.

learn-profile writes .alvar/LEARNER.md so the next session does not re-interview you. Pace, voice, solid ground.

What you feel as the learner

You stop spending tokens on “where do I start.” You answer a few picker questions. You see the graph. You take one step you can actually hold. The files persist:

.alvar/LEARNER.md
.alvar/maps/<topic>.md
.alvar/sessions/<date>-<topic>.md
.alvar/visuals/…

Point Obsidian at .alvar/ if you want LaTeX. The agent is the teacher. The folder is the notebook.

Install

npx skills add vasanthsreeram/Alvarmethod -g --all

Or:

npx github:vasanthsreeram/Alvarmethod

Then, in a learning folder (not the repo):

/teach I want a solid introduction to <topic>

If it pastes letters, say use the quiz tool. If it dumps a textbook, say one node only.

Pack: teach, probe, learn-profile, learn-visual, learn-verify. Method: Eero Alvar. Code: vasanthsreeram/Alvarmethod.