The Patient Robot in My Reading Room

Artificial intelligence can summarize an academic paper in seconds. Upload a PDF, ask for the main contributions, and receive a neat list before you have finished making coffee.

That is useful—but it is not the same as reading the paper.

As I began preparing for my PhD research in AI and robotics, I found myself using ChatGPT frequently while reading. Whenever I encountered an unfamiliar definition, a compact mathematical formulation, or an assumption that seemed obvious to the authors but not to me, I stopped and asked questions.

The explanations were often very good. The overall experience was not.

A normal conversation with an AI tends to drift toward summarization. It jumps ahead, combines several ideas, and explains what the paper will eventually say before I have had the opportunity to discover it. If I return the next day, the conversation may remember the general subject but not the exact sentence I reached, the concepts I understood, or the question I deliberately postponed.

Sometimes the reading unit is too small to express a complete idea. Sometimes I am told to go back to the PDF, locate half a paragraph, read it, and then return to the conversation. Figures present another problem: a cropped image without its caption may remove exactly the information required to interpret it.

I did not want an AI to read the paper for me. I wanted an AI to read it with me.

So I created ReCo.

What is ReCo?

ReCo is a custom AI skill designed to act as a research-paper reading companion.

Its purpose is not to produce a faster summary. Its purpose is to guide the reader through the original paper in meaningful units while preserving the process of learning.

ReCo presents one complete idea at a time. It gives the reader the relevant text, figure, table, or equation; waits while the reader examines it; answers questions at the reader's current position; checks understanding when useful; and continues only when the reader is ready.

It also maintains a learning state that is richer than a page number. That state may include:

• Which reading units have been completed.
• Which concepts the reader has already demonstrated understanding of.
• Which questions remain open.
• Which issues were intentionally deferred because the paper addresses them later.
• The precise point from which the next session should resume.

The distinction is important. A paper summary records what the paper says. ReCo tries to preserve what the reader has learned.

A small example

Imagine that we are beginning a fictional paper titled Robots That Anticipate Human Goals.

The ordinary approach would be to upload the paper and ask:

Summarize this paper and explain its main contribution.

Within seconds, the AI might tell us the problem, method, experiments, and conclusion. We would know what the paper contains—but we would not necessarily understand how the authors arrived there.

With ReCo, the interaction begins differently.

Reader: Help me study this paper.

ReCo: Let us begin with the first complete idea in the abstract:

Collaborative robots must infer a human partner's intended goal from partial observations if they are to provide useful assistance before being explicitly instructed.

Read this unit and tell me when you are done. You can stop and ask about any term or assumption.

Reader: What exactly counts as an observation? Is it an observed action, a sampled state, or sensor data?

ReCo: Any of those can serve as an observation, depending on the formal model. This sentence does not define the paper's observation model yet, so we should not assume which one the authors use. I will preserve this as an open question and return to it when the paper introduces its formal definition.

Reader: Okay, done.

ReCo: Before continuing, explain the central problem in your own words.

Reader: The robot needs to determine what the human is trying to achieve even though it has observed only part of the human's behavior.

ReCo: Exactly. The important relationship is: partial observations lead to an inference about the hidden goal. We can continue.

This tiny exchange demonstrates most of the idea behind ReCo.

The AI did not reveal the entire paper. It did not invent a missing definition. It recognized that the question was valid but premature, preserved it, and allowed the reader to continue. It then checked whether the central relationship had been understood.

The reader remains responsible for the intellectual work. ReCo provides the scaffolding.

ReCo was shaped by actual use

I did not design the entire workflow in advance. I created an initial version and began using it to study Goal Recognition as Planning: A Survey, a paper directly related to my research interests.

The first sessions quickly revealed weaknesses in my original design.

At first, ReCo divided the paper into units that were technically manageable but often too short. A single sentence is not necessarily a useful learning unit. Sometimes the following sentence completes the argument, supplies the contrast, or explains why the first sentence matters. I therefore changed the skill to prefer complete conceptual units, even when their lengths differ.

In another version, ReCo told me exactly which passage to read in the PDF and waited for me to return. That sounded reasonable, but in practice it created constant context switching. I wanted to concentrate on the concept, not alternate between windows every few sentences. ReCo was changed to present the relevant passage directly whenever possible.

Figures exposed a similar problem. During one session, the captured image of a figure was cut off. Part of the diagram—and its caption—was missing. Once the full figure was captured, the meaning became much clearer. This led to another explicit rule: when a visual element matters, ReCo must inspect the complete figure, including its caption, before presenting it.

The conversations also taught me that a good reading companion must know when not to answer.

While reading about goal recognition, I asked whether observations were actions, sampled states, or something determined by the system. The correct response at that point was not to force a universal definition. Different formulations represent observations differently, and the paper had not yet resolved the issue. ReCo needed to distinguish between:

• A prerequisite that should be explained immediately.
• A misunderstanding that should be corrected.
• A legitimate open question.
• A topic that the authors intentionally introduce later.

That distinction prevents the AI from replacing the paper's argument with its own convenient explanation.

Why not just use a good prompt?

A carefully written prompt can reproduce part of this behavior. You can ask an AI to move slowly, avoid spoilers, quiz you, and wait after each section.

The difficulty is consistency across a long reading process.

Research papers contain prose, definitions, notation, equations, algorithms, diagrams, tables, citations, and assumptions from earlier literature. The appropriate reading behavior changes with the material. A mathematical definition may require symbol-by-symbol reconstruction. A figure must be seen in full. A related-work paragraph may need a larger unit so that the comparison remains intact. A comprehension check that is valuable after a difficult formalization may be annoying after a straightforward transitional paragraph.

A reusable skill makes these decisions part of the workflow rather than forcing the reader to renegotiate them in every conversation.

ReCo is therefore not simply a long prompt. It is a set of reading policies: how to divide material, when to explain, when to wait, what to remember, how to handle visuals, how to avoid spoilers, and how to produce a useful learning record at the end.

The slower path may be the faster one

Reading with ReCo is slower than requesting a summary. That is intentional.

The goal is not to finish the PDF as quickly as possible. The goal is to build a mental model that remains useful after the conversation ends: to understand the terminology, see how the claims depend on one another, identify assumptions, and recognize where questions remain unresolved.

This does not mean that every reader needs the same pace. I often prefer to stop at each important idea, take it apart, question it, and connect it to what I already know. Another reader may prefer larger units and fewer comprehension checks. A proper reading companion should adapt to both while protecting the same basic principle: the AI should support the reader's understanding, not substitute for it.

There is also a broader lesson here about AI-assisted learning. The most impressive output is not always the most helpful one. An instant, polished explanation can create the feeling of understanding without the structure that makes understanding durable. Sometimes the better AI system is the one that says less, waits longer, and asks the right question at the right moment.

Try ReCo yourself

I am sharing ReCo so that other researchers and students can experiment with it, inspect how it works, and suggest improvements.

The skill is published as source code rather than as a closed service. The repository contains the main SKILL.md instructions together with the supporting resources required by the reading workflow. This makes the behavior visible and modifiable: readers are free to examine the decisions behind ReCo rather than treating it as a black box.

To install it in a Codex environment that supports skills, provide the public GitHub repository path and ask:

Install the ReCo skill from this GitHub repository: https://github.com/Turgibot/reco-skill

Alternatively, users who manage their local Codex skills directly can place the ReCo skill directory under their Codex skills folder. After installation, ReCo becomes available from the next turn.

Then attach an academic-paper PDF and begin with a request such as:

Use ReCo to help me study this paper. I want to understand it carefully, one meaningful unit at a time.

The repository should be considered an evolving project. ReCo improved because real reading sessions exposed problems that seemed minor in the abstract but mattered during learning. I expect the same process to continue as it is used with different papers, disciplines, mathematical styles, and reader preferences.

A companion, not a replacement

AI makes it easier than ever to avoid reading. It can extract conclusions, generate notes, answer questions, and create the convincing impression that we have absorbed a paper we barely examined.

ReCo began with a different question:

Can AI help us read more carefully rather than helping us skip the reading?

My experience so far suggests that it can—but only if the interaction is deliberately designed around the reader's learning process.

That is what I am trying to build with ReCo: not an automatic paper reader, but a patient companion that moves at your pace, remembers the path you took, and leaves the understanding with you.

ReCo is an ongoing personal project. If you try it, I would be interested to hear which parts help, where the workflow becomes frustrating, and what kinds of papers expose its next weaknesses.