Study Papers is an installable AI Agent Skill that turns a research paper into an interactive, prediction-first study session. Instead of summarizing the paper all at once, it progressively reveals questions, experimental designs, figures, results, and interpretations so the learner reasons before seeing each answer.
It is designed for scientific paper reading, academic research study, figure analysis, active recall, retrieval practice, spaced repetition, and critical evaluation.
Give this instruction to Codex or another AI agent that can install Agent Skills:
Install the
study-papersskill from https://github.com/Loffee5422/study-papers-skill/tree/main/skills/study-papers
For Codex's bundled skill installer:
python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py \
--url https://github.com/Loffee5422/study-papers-skill/tree/main/skills/study-papersThe skill becomes available on the next turn. Invoke it explicitly with $study-papers, or let Codex select it when you ask to study a paper interactively.
Most AI paper tools answer, “What does this paper say?” Study Papers instead asks, “What would you predict before seeing what the authors did?”
- Cascading reveal: question → design → prediction → data → interpretation → critique → next experiment
- Figure-first reasoning: treats each major figure, experiment, or argument as a learning unit
- Time-based depth: Snapshot, Core, Deep, or a custom time budget
- Visible progress: displays the current checkpoint, reveal stage, and remaining work
- Memory-first retrieval: asks for reconstruction before reopening the paper
- Scientific critique: separates observed evidence, author claims, learner predictions, and alternative explanations
- Paper compression: produces a compact contribution, evidence-chain, limitation, and connection note
- Spaced recall: offers later retrieval sessions for durable understanding
| Level | Typical time | Coverage |
|---|---|---|
| Snapshot | 10–15 min | Orientation, 1–2 decisive evidence units, main limitation, final reconstruction |
| Core | 30–45 min | Central evidence chain, standard reveal loop, retrieval, compression note |
| Deep | 60–90+ min | All major evidence, methods and assumptions, alternatives, repeated retrieval, spaced recall |
| Custom | Your budget | Highest-value evidence fitted to the available time, with omissions stated |
Time estimates vary with paper length, complexity, and prior knowledge.
Depth: Core (~30–45 min)
Progress: [████░░░░░░] 4/9 checkpoints
Now: Figure 2 of 4 — Data (stage 3/6)
Remaining: 2 evidence units · final reconstruction · compression
The route only advances when the checkpoint's learning criterion is met. Skipped material is labeled separately from learned material, and changing depth recalculates the plan transparently.
Use $study-papers to guide me through this attached paper. I have 15 minutes.
Use $study-papers in Core mode. Hide each result until I make a prediction.
Use $study-papers to quiz me on this paper from memory, then repair only my gaps.
Use $study-papers for a Deep study of this theoretical paper. Focus on assumptions and counterexamples.
Survey → Question → Predict → Reveal → Compare → Critique → Retrieve later
For empirical papers, the loop centers on experiments and figures. For theoretical, mathematical, or review papers, it adapts to claims, assumptions, derivations, evidence, implications, and counterexamples.
- An AI agent that supports the
SKILL.mdAgent Skills convention - Access to the paper as a PDF, attachment, URL, or extracted text
- No runtime dependencies, API keys, accounts, or external services required by the skill itself
The repository uses a path-addressable layout so an AI agent can install the skill directly from the GitHub tree URL.
The workflow draws on:
- The Questioning–Predicting/Comparing approach to active scientific-paper reading described in “Active versus passive reading: how to read scientific papers?”
- Research reviewed in “The science of effective learning with spacing and retrieval practice”
- Experimental work on the interaction between retrieval and spacing in “The Critical Importance of Retrieval—and Spacing—for Learning”
These sources motivate the learning mechanics; they do not imply that this particular skill has been independently validated as an educational intervention.
study-papers-skill/
├── README.md
├── README.zh-CN.md
├── llms.txt
└── skills/
└── study-papers/
├── SKILL.md
└── agents/
└── openai.yaml
No. It can end with a compact paper note, but its primary job is to make the learner predict, interpret, critique, and retrieve the paper's argument rather than passively receive a summary.
Yes. Snapshot mode prioritizes the paper's central question, decisive evidence, main limitation, and one final reconstruction. It explicitly leaves lower-priority material out.
Yes, when the host AI agent can read the attached or linked PDF. The skill itself remains independent of a specific PDF tool.
Yes. The empirical workflow fits experimental papers, while the alternate reasoning loop supports theoretical, mathematical, and review papers.
Progress is checkpoint-based, not an unreliable estimate of pages read. Orientation, selected evidence units, retrieval, and compression each have explicit completion criteria.
Focused improvements are welcome. See CONTRIBUTING.md.
MIT License. See LICENSE.