Learn effectively from AI

You are reading a paper outside your field and every other paragraph sends you to a reference you have not read. The standard approach is to chase those references until you have built up enough background to understand the thing you actually wanted to read. That takes hours, sometimes days, and half of what you read along the way turns out to be irrelevant to the specific way the paper uses the concept.

There is a faster way. Give the model the material you are trying to understand and an honest description of your background. What are you comfortable with? Where exactly does your understanding stop? It identifies the specific concepts you are missing and explains them in terms of what you already know. If you know quantum mechanics but not quantum information, it explains entanglement entropy using the density matrix formalism, not by starting from scratch.

The important part is you do not have to trust the model to correctly recall and explain these concepts from its training data. The better approach is to give it the actual textbook or paper and have it explain (concisely) the author's words to you. A model that is reading the source material in front of it is dramatically more reliable than one answering from training data alone. Have it find the right references first, then have it walk you through them.

When an explanation does not land, you push back. "Try a different angle." "Give me a concrete example from condensed matter." It adjusts in real time. You can also ask it to test you: not quiz questions, but the kind that reveal whether you could actually use the idea in your own work or whether you are just nodding along. This back-and-forth is what makes it fundamentally different from reading a tutorial. It adapts to what you specifically do not understand, and it does it in minutes instead of days.

How you ask matters more than which model you ask.

Option 1

Web browser

You use ChatGPT or Claude in a browser tab, upload the paper you are stuck on, and ask questions from there.

Option 2

Power user

You use Claude Code or Codex CLI with a running context file so the model remembers what you already know across sessions.

You do not need a research agent for this one

Unlike searching the literature, learning from a paper does not need Deep Research, the Research button, or any other agentic mode. You are not asking the model to go find things. You are giving it the source material directly and asking it to explain part of it. A regular chat window with the paper attached works fine. Open Claude or ChatGPT, drop the PDF in, and start.

What does matter is the prompt.

Let's walk through an example

You are a machine learning researcher. Your collaborator forwards you the LIGO first-detection paper (Abbott et al. 2016, arXiv:1602.03837) because they want to brainstorm whether modern self-supervised methods could improve the detection pipeline. You know probability, signal processing, and Fourier analysis cold. You have never thought about gravitational waves, strain data, or numerical-relativity templates in your life.

You get to the methods section, where the authors describe how the signal was actually identified: by matched-filtering the detector output against a bank of template waveforms, with the signal-to-noise ratio computed against the noise power spectral density. You know what a matched filter is in white noise. You do not know what any of these words mean in this paper. Here is the prompt almost everyone tries first.

Bad prompt

Explain matched filtering in gravitational wave detection.

What comes back is a competent textbook explanation of matched filtering that starts from white-noise detection theory, walks through the standard derivation, and never once touches the actual sentence you were stuck on or the way this particular paper uses the idea. You learn the same things you would have learned from any signal-processing course. You still cannot read the paper.

Good prompt

The same question, asked properly. It uploads the paper, names your background concretely, points at the specific passage that has you stuck, and demands a bridge from what you already know rather than a tutorial from scratch. It also asks for one worked numerical example and a short quiz at the end.

Prompt: bridge from ML to LIGO matched filtering

I have uploaded the LIGO first-detection paper (Abbott et al. 2016, "Observation of Gravitational Waves from a Binary Black Hole Merger", arXiv:1602.03837).

My background: machine learning researcher. Strong on probability, linear algebra, Fourier analysis, and signal processing as it shows up in audio and time-series ML. I have basically zero background in general relativity, and I have never worked with strain data or numerical-relativity waveforms.

I am stuck on the part of the paper that describes how the signal was actually identified: matched filtering against a bank of template waveforms generated from numerical relativity, with the signal-to-noise ratio computed against the detector noise power spectral density.

Build a bridge from what I already know. Walk me through one concrete worked example.

Notice what you did not do. You did not ask the model to teach you gravitational-wave astronomy. You asked it to explain one passage in one paper in terms of something you already believed, with one worked example you could check by hand. That constraint is the entire move.

Same scenario as on the browser tab: you're the ML researcher trying to read the LIGO first-detection paper, and matched filtering against a non-flat detector PSD is the sentence you can't get past. The browser approach works once, but every chat closes and takes half your progress with it. Next week you open the paper, you have forgotten half of what you figured out, and you ask the same questions from scratch.

Claude Code inside VS Code fixes that. Your reading life becomes a folder the agent can navigate: the paper itself, a running notes file about what you already know, a skill that remembers how you like to be tutored. Every unstuck moment compounds into the next. The arc below sets that up once, then turns every future paper into a short loop instead of a fresh start.

Install arxiv-latex-mcp so the agent reads real equations

PDFs are rendered output. The model reads them through OCR, and OCR on dense mathematical typesetting is unreliable: subscripts drift, operators get dropped, integrals turn into the letter J. The community-maintained arxiv-latex-mcp server (at github.com/takashiishida/arxiv-latex-mcp ) pulls the raw LaTeX source of any arXiv paper instead, so the agent reads the equations exactly as the authors typed them. For physics, math, and theoretical CS papers this is the difference between explaining the right equation and confidently explaining a hallucinated one.

Prompt: install arxiv-latex-mcp

I want you to be able to read the raw LaTeX source of arXiv papers, not the OCR'd PDFs, so equations don't get garbled. Set that up for this project and tell me if you need anything from me. When it's working, verify by pulling arXiv:1602.03837 (the LIGO first-detection paper) and reading me back the exact sentence that defines the matched-filter signal-to-noise ratio, with the equation rendered from the source rather than OCR.

Write a learning.md so the agent knows what you already know

The single biggest lever in agentic learning is a file the agent reads at the start of every session. Three sections: what you are already comfortable with, what you are currently learning, and a running log of questions you got stuck on plus the explanations that finally unstuck them. Hand the prompt below off and the agent will write the file for you, scoped to the paper you are about to read.

Prompt: write learning.md for this project

Create a file called learning.md at the root of this project. Three sections.

1. Background I am comfortable with. Fill this in with: machine learning researcher, strong on probability, linear algebra, Fourier analysis, signal processing as it shows up in audio and time-series ML, comfortable with matched filters from a detection-theory standpoint. Zero background in general relativity, never worked with strain data or numerical-relativity waveforms.

2. Currently learning. Fill in: enough of the LIGO detection pipeline (arXiv:1602.03837) to read the methods section without getting stuck.

3. Questions I got stuck on. Leave this empty for now. Future sessions will append here.

Keep the file short and skimmable. This file is the persistent context for every learning session in this project.

The file is short on day one and grows by a few lines per paper. After a few weeks it becomes a personal textbook sitting in your file explorer, one click away whenever you want to edit it by hand.

Run paper-reviewer in confused-grad-student mode

The Harvard paper-reviewer skill was built for critiquing a paper you are writing. One of its reviewer personalities, confused-grad-student, is a gift when flipped: pointed at a paper you are trying to read, it surfaces every place the paper introduces notation or terminology without explaining it. You get the inventory of gaps in 30 seconds, before wasting an hour discovering them the hard way.

Prompt: surface notation gaps with paper-reviewer

Run the paper-reviewer skill from github.com/Harvard-Agentic-Science/Skills in confused-grad-student mode over arXiv:1602.03837 (the LIGO first-detection paper). Use the LaTeX source via arxiv-latex-mcp, not the PDF.

I want a structured list of every place the paper introduces notation, terminology, or a result without explaining it for a reader who is not a gravitational-wave physicist. For each item: the exact passage, page or section number, and one line on what a reader without a GR background needs to know to parse it. Do not try to teach any of it yet. I just want the inventory of gaps so I can decide which ones are actually blocking me.

The output is a structured list. Most items will be things you can skip. One or two will be the actual blockers. Those go straight into the next prompt.

Ask /teach to bridge from your background to the blocker

Now the loop. Pick the blocker, hand it to a custom /teach skill that reads learning.md for your current mental model and the paper for the source material, and ask for a bridging explanation with one worked numerical example. The skill itself is a short markdown file at .claude/skills/teach/SKILL.md in your project (you can write it once with a single prompt, or copy it from the skills library linked at the bottom of this page).

Prompt: ask /teach to bridge from what you know

Read learning.md so you know what I already understand and what I am working on. Then read arXiv:1602.03837 via arxiv-latex-mcp.

I am stuck on the matched-filtering description in the methods section. Specifically, I do not yet have an intuition for why the detector noise power spectral density appears in the denominator of the SNR integrand, and what that means in practice when the LIGO noise floor is so far from white.

Bridge from what learning.md says I already know. Quote the relevant equation from the LaTeX source verbatim before explaining it. Then work one concrete numerical example: a short chirp template, a synthetic noise time series with a non-flat PSD shaped roughly like LIGO's, and the matched-filter SNR computed step by step in the frequency domain. Eight or so numbers, enough that I can see why a noisy frequency band gets down-weighted.

Do not give me a GR tutorial. Do not summarise the paper. Stop after the worked example.

A good response quotes the equation verbatim from the LaTeX source, then maps each piece onto the reader's existing language. For the LIGO matched-filter integrand, that means recognising the 1/S_n(f) factor as the standard whitening filter from coloured-noise detection theory, and showing on a short worked example that frequency bands where the detector is noisy contribute almost nothing to the SNR. Two short paragraphs and a few lines of arithmetic, not a derivation. The reader can check it by hand.

Ask to be quizzed

The part most researchers skip. A two-minute quiz at the end of each session is what separates "I followed the explanation" from "I could use this idea in my own work." Three questions is enough.

Prompt: quiz yourself on what just landed

Quiz me on what we just covered. Three short questions, increasing in difficulty. The first should check that I can recover the matched-filter SNR formula from memory. The second should test whether I understand why the PSD appears where it does. The third should be applied: give me a one-sentence scenario (a different noise spectrum, or a template with most of its energy in a noisy band) and ask me to predict qualitatively what happens to the SNR.

Wait for my answers before grading. When you grade, point out any conceptual gap, not just whether I got the algebra right.

Append what you learned back into learning.md

Last step in the loop. The thing you just figured out gets written back into learning.md in your own voice, so next week's session starts where this week's left off. Auto-memory in Claude Code's settings (~/.claude/settings.json, set "autoMemoryEnabled": true) handles a parallel layer of agent-curated notes, but the file you maintain by hand is the one you will actually read.

Prompt: append what you learned to learning.md

Append a new entry to the "Questions I got stuck on" section of learning.md based on what we just worked through.

Format:
- "<one-line version of the question I was stuck on>"
  -> <two or three lines, in my voice, of the explanation that finally landed>

Be concrete. The point is that when I open this file in three weeks I should be able to recover the intuition without rereading the paper.

Run that loop once per blocker. By the third or fourth paper, your learning.md is the textbook you wish you had been handed, your /teach skill has stabilised into your preferred way of being tutored, and the time from "I do not understand this sentence" to "I can use this idea" has dropped from an afternoon to a coffee break.