---
name: reading-papers
description: Use when studying academic papers in the Obsidian vault — user reads a converted markdown paper under private/papers/, discusses it through Keshav's three-pass framework with LLM assistance, and wants annotations added as callouts. Also use when organizing paper notes into PARA topic files.
---

# Reading Papers

## Overview

Three-pass study flow for academic papers, based on [Keshav's framework (2007)](https://doi.org/10.1145/1273445.1273458) with LLM assistance at each stage. Papers live as markdown in `private/papers/` (converted via [Marker](https://github.com/datalab-to/marker)). Each pass has a distinct goal, and the user must engage independently before the LLM deepens understanding.

## Workflow

```dot
digraph reading_flow {
  "Read paper markdown" -> "Pass 1 gate";
  "Pass 1 gate" -> "Suggest skimming first" [label="not skimmed yet"];
  "Suggest skimming first" -> "Pass 1 gate" [label="user returns"];
  "Pass 1 gate" -> "Pass 1: Orient" [label="user has skimmed"];
  "Pass 1: Orient" -> "5 C's evaluation";
  "5 C's evaluation" -> "Worth continuing?";
  "Worth continuing?" -> "Done (file or discard)" [label="no"];
  "Worth continuing?" -> "Pass 2 gate";
  "Pass 2 gate" -> "Suggest careful reading" [label="not read yet"];
  "Suggest careful reading" -> "Pass 2 gate" [label="user returns"];
  "Pass 2 gate" -> "Pass 2: Comprehend" [label="user has read"];
  "Pass 2: Comprehend" -> "Add callouts to paper";
  "Add callouts to paper" -> "Deep dive needed?";
  "Deep dive needed?" -> "Done (organize)" [label="no"];
  "Deep dive needed?" -> "Pass 3 gate";
  "Pass 3 gate" -> "Suggest re-reading with intent to reconstruct" [label="not ready"];
  "Suggest re-reading with intent to reconstruct" -> "Pass 3 gate" [label="user returns"];
  "Pass 3 gate" -> "Pass 3: Reconstruct" [label="user ready"];
  "Pass 3: Reconstruct" -> "Adversarial Q&A";
  "Adversarial Q&A" -> "Add callouts to paper (deep)";
  "Add callouts to paper (deep)" -> "Done (organize)";
  "Done (organize)" -> "Single topic or multiple?";
  "Single topic or multiple?" -> "Publish as blogmark" [label="single"];
  "Single topic or multiple?" -> "PARA topic split" [label="multiple"];
  "Publish as blogmark" -> "Replace callouts with transclusions";
  "Replace callouts with transclusions" -> "Update cross-references";
  "PARA topic split" -> "Delete or slim original";
}
```

## Phase 0: Reading Gate (Before Each Pass)

**The LLM deepens understanding — it must not replace reading.**

Before each pass, confirm the user has done the independent work:

| Pass | Independent work required                    | Gate question                                                                  |
| ---- | -------------------------------------------- | ------------------------------------------------------------------------------ |
| 1    | Skim title, abstract, headings, conclusions  | "Have you skimmed the paper? What's your first impression?"                    |
| 2    | Read with care, study figures and diagrams   | "Have you read through the paper? What terms or sections were unclear?"        |
| 3    | Attempt to mentally reconstruct the argument | "Have you tried reconstructing the paper's argument? Where did you get stuck?" |

**If the user hasn't read yet:** suggest they do so and come back. Don't summarize or spoil. The gate question doubles as a lightweight comprehension check — their answer focuses the upcoming LLM-assisted pass.

## Phase 1: Orient (First Pass)

**Goal:** General idea of the paper. Decide whether to continue.

**Time investment:** User spends 5-10 minutes skimming independently, then LLM session.

After the user passes the gate:

1. **Summarize structure** — what kind of paper is this, how is it organized
2. **Evaluate the 5 C's together:**
   - **Category** — measurement, analysis, prototype, survey, theoretical?
   - **Context** — what other work is this related to? What theoretical bases?
   - **Correctness** — do the assumptions appear valid?
   - **Contributions** — what are the main contributions?
   - **Clarity** — is it well written?
3. **Surface key references as described by the paper** — note which works the authors position themselves against or build upon, based on how the paper cites them
4. **Explicit gate:** "Based on this, is this paper worth a second pass? If it's outside your area or the contributions aren't relevant, it's fine to stop here."

### Callouts for Pass 1

- `[!info]` for the 5 C's evaluation
- `[!warning]` for red flags (questionable assumptions, missing context)

## Phase 2: Comprehend (Second Pass)

**Goal:** Grasp the paper's content, but not every detail. Be able to summarize the main thrust with supporting evidence to someone else.

**Time investment:** User spends up to 1 hour reading independently, then LLM session.

After the user passes the gate:

1. **No spoilers** — do not preemptively summarize sections the user hasn't asked about. Follow Jeremy Howard's approach: answer what's asked, don't reveal what's ahead.
2. **Clarify flagged terms and sections** — the user identified unclear parts at the gate. Explain these with references and citations to related work, not just definitions. This is the key acceleration point.
3. **Walk through figures and diagrams** — trust Marker's conversion. Describe what the LLM can see in the converted output first. Only ask the user for clarification if something appears garbled or missing.
4. **Surface connections as the paper describes them** — when explaining a concept, cite how the paper itself frames its references ("the authors cite X as..."). Add context from training knowledge only when confident, and flag uncertainty explicitly ("I believe this paper is about... but I haven't verified"). Never present inferred knowledge about a reference as fact.
5. **Mark references for future reading** — flag papers from the bibliography that the paper describes as foundational or that appear frequently in the argument. The user should read key references themselves rather than rely on the LLM's knowledge of them.
6. **Check comprehension** — after addressing the user's questions, ask them to summarize the main thrust of the paper with supporting evidence (Keshav's bar for pass 2).

### Callouts for Pass 2

- `[!question]` for Q&A about unclear terms and concepts
- `[!info]` for supplementary context, related work, and reference suggestions
- `[!example]` for concrete illustrations of abstract concepts
- `[!warning]` for common misconceptions about the techniques or results

## Phase 3: Reconstruct (Third Pass)

**Goal:** Virtually re-implement the paper — make the same assumptions as the authors, attempt to re-create the work. Identify innovations, hidden assumptions, and weaknesses.

**Time investment:** User spends 1-3 hours independently, then LLM session.

After the user passes the gate:

1. **Adversarial Q&A** — challenge the paper's claims:
   - "What happens if assumption X is wrong?"
   - "How would you design this experiment differently?"
   - "What's the weakest link in the argument chain?"
2. **Reconstruct the argument** — ask the user to walk through the paper's logic, LLM probes for gaps
3. **Compare with related work** — based on how the paper positions itself against cited work, probe whether the claimed differences hold up. Flag when the LLM lacks knowledge of a referenced paper.
4. **Identify what's novel vs. incremental** — what's the actual contribution beyond prior work?

### Callouts for Pass 3

- `[!question]` for adversarial Q&A exchanges
- `[!warning]` for identified weaknesses, hidden assumptions
- `[!example]` for alternative approaches or experimental designs discussed
- `[!info]` for connections to other work

## Phase 4: Organize

Follow the same publish/organize conventions as `studying-articles`:

### Single-topic paper → Blogmark

1. Create public blogmark at `content/notes/blogmarks/<filename>.md`
2. Frontmatter: `tags: [Blogmarks, Papers]`
3. AI disclosure callout
4. Replace callouts in private paper with transclusions using folder paths

### Multi-topic paper → PARA Split

1. Ask user whether to publish as single blogmark or split by topic
2. Follow reviewing-notes Phase 4 conventions for PARA placement
3. Each file gets frontmatter, AI disclosure, source link
4. Handle original per user preference

### Cross-references

- Public-to-public: wikilinks
- Public-to-private: external URLs (e.g., DOI links), never wikilinks

## Common Mistakes

| Mistake                                             | Fix                                                                    |
| --------------------------------------------------- | ---------------------------------------------------------------------- |
| Summarizing paper before user has read it           | Always check reading gate first — each pass requires independent work  |
| Treating all passes the same (just Q&A)             | Each pass has a distinct goal: orient → comprehend → reconstruct       |
| Skipping the "worth continuing?" gate after Pass 1  | Explicitly ask — not every paper deserves three passes                 |
| LLM doing the reconstruction for the user in Pass 3 | User walks through the argument; LLM probes and challenges             |
| Forgetting the 5 C's in Pass 1                      | Always evaluate Category, Context, Correctness, Contributions, Clarity |
| Not marking unread references                       | Surface key references in Pass 1, mark relevant ones in Pass 2         |
| Block ID on own line after callout                  | Put `> ^id` on last line inside blockquote                             |
| Wikilinks from public to private content            | Use DOI or source URLs for private content                             |