ALTER Framework

Five roles you give AI to assemble a working personal university out of tools you already have. Coined by Sandeep Swadia as the operational payload of How To Become Dangerously Self-Educated With AI (theMITmonk) under the brand name "University in a Box." The roles are chosen so each fills a specific gap the others don't, and — Sandeep argues — collapsing them into one prompt breaks the framework because the roles' goals genuinely conflict.

The mnemonic: Advisor, Librarian, Tutor, Editor, Roommate.

The five roles

Role	What it does	Failure mode it prevents	Concrete prompt shape
Advisor	Builds a personalized curriculum	Generic "teach me X" prompts that surface a Wikipedia dump	"Act as my elite academic advisor. Interview me for baseline, destination, sequencing, cut-list, and milestones. Ask me up to five questions for each, one at a time."
Librarian	Curates sources; defends the curriculum from distraction	Endless algorithmic feed / open-web wandering	Load 3–4 vetted sources into NotebookLM; ask questions grounded to those documents only.
Tutor	One-on-one diagnosis of where you're confused	Fluency illusion — feeling like you understood the material when you didn't	Voice mode. "Ask me one question at a time. Don't lecture me. Find the gap in my understanding. Be precise."
Editor	Real-time feedback on what you produce	Publishing / sharing sloppy work without pressure-testing it	"Challenge my thinking. Find weaknesses in my arguments. Cut repetition. Tighten structure. Make the language more precise or concrete."
Roommate	Cross-domain perspective that expands your range	Getting deeper but narrower over time	"How does a jazz musician think about playing with others, and how does that relate to building teams?"

Why five roles instead of one

Sandeep's argument for the split: the roles have incompatible goals if you collapse them into one prompt:

An Advisor that also acts as a Librarian will over-recommend the sources it already helped you plan around (curriculum-and-source coupling).
A Tutor that also acts as an Editor will smooth over your confusions instead of surfacing them (correction-and-critique coupling).
A Roommate that also acts as any of the above will keep pulling you back to your own field — the whole point is to leave it briefly.

The practical implication: run each role in a separate chat / session / tool, not as roles switched-into-and-out-of inside one conversation. The chat context boundary is what protects each role's job.

The Advisor role in detail (the load-bearing one)

Everything else assumes the Advisor step has been done. Sandeep's five decisions the Advisor should make (interview-style, one question at a time):

Destination — what do you want to know or do at the end?
Baseline — where are you on this topic right now (basic / intermediate / nothing)?
Sequencing — what order should you learn this in?
Cut list — what should you ignore for now?
Milestones — what output from you proves you've mastered the topic and can move on?

This is not a "prompt template"; it's a decision list the AI must extract from you before offering a plan. The point is that the Advisor's plan is only as good as your answers to these five — and the five are things most learners have never explicitly answered for themselves.

Cross-link: this is essentially a PRIME Framework-shaped prompt (Sandeep's earlier prompting rubric — Purpose / Research / Interview / Mechanics / Examples) applied to the specific case of curriculum design.

The Librarian role and NotebookLM

Anchor: NotebookLM as the tool that lets AI operate grounded to a specific document set rather than the open web. Sandeep's specific move:

Build a shelf — 3–4 vetted sources: a textbook, a video course, a podcast, some articles. If you don't know what to load, use deep-research mode (Gemini / Claude / ChatGPT) to find the shortlist first.
Load into NotebookLM.
Use the Studio to transform the shelf — Studio can generate a podcast, a video, or a slideshow only from those sources. Sandeep specifically highlights that NotebookLM's generated podcast supports live call-in — you can interrupt the podcast mid-play with a question, and the AI hosts answer grounded to your source shelf.

The Librarian's real job is a triage function: "which 10 sources deserve your attention, which thousands can you ignore."

The Tutor role and Bloom's two-sigma

Anchor: Two-Sigma Problem — Benjamin Bloom's 1984 finding that one-on-one tutoring pushes the average student above the 98th percentile of a classroom control. Bloom's own frustration: tutoring works, and no education system could ever afford to deliver it at scale. Sandeep's claim: AI breaks the economics of tutoring.

Practical form: voice mode (Gemini voice, ChatGPT voice) so the interaction feels like an office hour, not a chat window. The two commands Sandeep highlights:

"Teach me" — start with the concept.
"Test me" — then have the tutor pressure-test whether you actually got it.

A good tutor's job, per Sandeep, is to "make it harder for you to fake understanding" — the opposite of a smoothing / summarizing chat interaction. If your tutor is agreeing with you a lot, you're doing it wrong.

The Editor role and the F1 analogy

"Excellence lives in feedback." Authors have editors, filmmakers have editors, journalists have editors, CEOs have boards. F1 analogy: a race is 90 minutes; the difference between winning and losing is fractions of a second per turn; the team monitors thousands of data points per lap in real time. AI can do a version of that for anything you produce — but only if you ask for critique, not polish.

The prompt discipline: never say "improve this." Say "challenge my thinking, find weaknesses, cut repetition, tighten structure, make it more concrete." The verbs matter — improve/polish/refine are ambiguous, and AI will pick the least demanding interpretation.

The Roommate role and cross-domain perspective

The role that expands your range. Sandeep's dorm anecdote: a roommate who talked about the economy of design and aesthetic tradeoffs taught him more about entrepreneurship / taste / strategy than a business professor would have — precisely because the domain was different.

Institutional analogue: Pixar's in-house school — animators in sculpting class, accountants and security guards in drawing class. The perspective from an unlike world makes your own world newly visible.

The prompt shape is a cross-domain analogy request, not a knowledge request:

"What's the connection between cooking and finance?"
"How does a jazz musician think about playing with others, and how does that relate to building teams?"
"What can a Formula 1 pit crew teach me about code review?"

The point isn't the answer; it's the lens the answer forces you to try.

When ALTER works and when it doesn't

Works well when:

You're self-directed and have a real goal (not open-ended browsing).
You're willing to run the roles in separate sessions instead of collapsing them.
You treat the Advisor step as non-optional — most people skip it and stall in the Librarian phase.

Struggles when:

The topic is fast-moving and the Librarian shelf is stale before the Advisor curriculum finishes.
You have no ability to do something with what you're learning (no application layer → the framework is knowledge-only).
You over-trust the Tutor for anything with adversarial stakes (medicine, law, safety-critical work) — even a good AI tutor's confident-wrong is confident-wrong.

Cross-vault connections

Sandeep Swadia — creator of the framework; 11th named framework in his catalog
theMITmonk — channel
How To Become Dangerously Self-Educated With AI (theMITmonk) — canonical source
Two-Sigma Problem — theoretical anchor for the Tutor role
TRAP Framework — Sandeep's earlier learning framework (Test / Retain / Associate / Perform); ALTER is upstream (curriculum design + application), TRAP is downstream (retention practice on the material)
PRIME Framework — Sandeep's prompting rubric; the ALTER Advisor prompt is a PRIME-shaped prompt applied to curriculum design
DRAG Framework — Sandeep's "what to delegate to AI"; ALTER is a delegation-first framework by construction
GPS Check (for Agents) — Sandeep's Goal/Proof/Steps pre-flight for agents; the Advisor's Destination/Milestones/Sequencing is the analogous pre-flight for learning
Higher-Order Thinking — Justin Sung's concept; ALTER can be read as scaffolding to develop higher-order thinking (advisor decides what connects to what; roommate forces cross-domain synthesis) rather than offloading it
Cognitive Offloading — ALTER is a scaffolding answer to the offloading trap: use AI to organize learning, not to replace it
Fluency Illusion — the Tutor role is Sandeep's direct answer to it
NotebookLM — the concrete tool for the Librarian role
MCP / Zapier — Sandeep's implementation surface for turning the Advisor's plan into calendar events and documents