| name | read-paper |
| description | Read and analyze research papers using Keshav's three-pass method for efficient comprehension |
Read Paper
Apply Keshav's three-pass method for efficient paper reading. Each pass builds on the previous one.
Procedure
1. Determine depth
If user specifies depth (1, 2, or 3), use that. Otherwise default to pass 2.
- Pass 1: Bird's-eye view. Use when screening papers or outside your specialty.
- Pass 2: Grasp content without details. Use for papers of interest but not core research.
- Pass 3: Full understanding. Use for papers you need to review, reproduce, or build upon.
2. Execute passes sequentially
Always start with pass 1, even if targeting a deeper pass.
Pass 1: Bird's-Eye View (5-10 min equivalent)
Read only:
- Title, abstract, introduction
- Section and sub-section headings (skip body text)
- Conclusions
- References (note which you recognize)
Output: The Five Cs
## Pass 1 Summary
### Category
What type of paper is this?
- [ ] Measurement/empirical study
- [ ] Analysis of existing system
- [ ] Research prototype description
- [ ] Survey/tutorial
- [ ] Theoretical/formal methods
- [ ] Position/vision paper
**Type:** [one sentence]
### Context
- **Related papers:** [list 2-5 key related works mentioned]
- **Theoretical basis:** [frameworks, models, or prior results this builds on]
- **Research area:** [subfield and broader field]
### Correctness
- **Assumptions appear valid?** [yes/no/unclear]
- **Red flags:** [any obvious methodological concerns, or "none noted"]
### Contributions
- **Main contribution:** [one sentence]
- **Secondary contributions:** [bullet list if any]
### Clarity
- **Well written?** [yes/partially/no]
- **Structure quality:** [clear sections, logical flow, or issues noted]
---
### Recommendation
- [ ] **Read deeper** — relevant to my work, assumptions seem sound
- [ ] **Set aside** — interesting but outside current focus
- [ ] **Skip** — invalid assumptions / not relevant / poorly written
- [ ] **Need background first** — unfamiliar terminology or techniques
### References to follow up
- [list any unread references that seem important]
GitHub & Code Links
- [List any GitHub repos, project pages, or dataset URLs mentioned in the paper, or "None found"]
If depth = 1, stop here.
Pass 2: Grasp Content (up to 1 hour equivalent)
Read with greater care, but skip proofs and dense technical details.
Focus on:
- Figures, diagrams, graphs — Are axes labeled? Error bars present? Do results support claims?
- Key arguments — Jot down the logical structure
- Evidence quality — How strong is the support for each claim?
- Unread references — Mark important ones for background reading
Output: Content Analysis
## Pass 2 Analysis
### Main Argument Structure
[Outline the paper's logical flow: problem → approach → evaluation → conclusions]
### Key Figures & Results
| Figure/Table | What it shows | Supports claim? | Notes |
|--------------|---------------|-----------------|-------|
| Fig 1 | ... | Yes/Partially/No| ... |
| Table 2 | ... | ... | ... |
### Key Architectures & Models
For each major component, identify:
- **Model/architecture used** (e.g., BERT variant, LLaMA, GNN)
- **How it was trained/adapted** (pre-trained + fine-tuned? from scratch? distilled?)
- **What it takes as input / produces as output**
- **Why this architecture was chosen over alternatives** (if stated)
If the paper combines multiple components into a pipeline, draw the data flow:
[input] → [component A: model X] → [intermediate] → [component B: model Y] → [output]
### Evidence Assessment
- **Strongest evidence:** [what is most convincing]
- **Weakest evidence:** [what is least convincing or missing]
- **Unstated assumptions:** [implicit assumptions in methodology]
### Technical Gaps (for me)
[List concepts, techniques, or background I'd need to fully understand this]
### Summary
[2-3 sentences: main thrust of paper with supporting evidence, suitable for explaining to someone else]
### Updated Recommendation
- **Worth a third pass?** [yes — if I need to reproduce/review/build on this; no — pass 2 sufficient]
- **Key references to read first:** [if third pass needed, what background is missing]
If depth = 2, stop here.
Pass 3: Deep Understanding (4-5 hours equivalent for beginners)
The goal is to virtually re-implement the paper: make the same assumptions as the authors and mentally recreate the work.
Focus on:
- Challenge every assumption — What if this assumption is wrong?
- Identify hidden assumptions — What's implicit but not stated?
- Compare your approach — How would you present this idea differently?
- Find innovations — What's genuinely new vs. incremental?
- Find weaknesses — Missing citations, flawed experiments, logical gaps
- Generate ideas — What future work does this suggest?
Output: Deep Analysis
## Pass 3 Deep Analysis
### Virtual Re-implementation
If I were to recreate this work:
- **I would keep:** [aspects that are well-designed]
- **I would change:** [aspects I'd approach differently]
- **Key insight I gained:** [what clicked by thinking through the approach]
### Assumption Audit
| Assumption | Stated? | Valid? | Impact if wrong |
|------------|---------|--------|-----------------|
| ... | Yes/No | Yes/No/Unclear | High/Medium/Low |
### Innovations vs. Incremental
- **Genuinely novel:** [what's new]
- **Incremental/expected:** [what follows naturally from prior work]
### Weaknesses & Missing Elements
- **Missing citations:** [relevant work not cited]
- **Methodological issues:** [experimental or analytical problems]
- **Logical gaps:** [arguments that don't follow]
- **Threats to validity:** [internal and external]
### Key Statistical & Mathematical Terms
For each statistical or mathematical concept used in the paper, provide a plain-language definition:
| Term | Definition | Used for |
|------|------------|----------|
| [e.g., F1 score] | [Harmonic mean of precision and recall] | [Evaluating gap detection] |
| ... | ... | ... |
### Proof/Technique Inventory
[Techniques used that I can add to my repertoire]
### Future Work Ideas
[Ideas for follow-on research sparked by this paper]
### Final Assessment
- **Strong points:** [bullet list]
- **Weak points:** [bullet list]
- **Overall quality:** [excellent / good / adequate / poor]
- **Reproducible?** [yes / partially / no — and why]
### One-paragraph summary
[Comprehensive summary suitable for a literature review, covering problem, approach, key results, limitations]
Literature Survey Mode
When the user asks to survey a field (not a single paper):
- Find seed papers: Use WebSearch with well-chosen keywords to find 3-5 recent papers
- Pass 1 each: Get a sense of the work, read their Related Work sections
- Look for surveys: If a recent survey exists, read it and you're done
- Find key papers: Identify shared citations and repeated author names in bibliographies
- Find top venues: Check where key researchers publish to identify top conferences
- Scan proceedings: Look through recent proceedings of top venues for related work
- Iterate: If all papers cite something you missed, obtain and read it
Output: Survey Summary
## Literature Survey: [Topic]
### Key Papers (by influence)
1. [Paper] — [one-line contribution]
2. ...
### Key Researchers
- [Name] — [affiliation, focus area]
### Top Venues
- [Conference/Journal] — [why it's relevant]
### Research Themes
- **Theme 1:** [description, key papers]
- **Theme 2:** ...
### Open Problems
- [Problem 1]
- [Problem 2]
### Recommended Reading Order
1. [Start here — foundational]
2. [Then this — builds on #1]
3. ...