---
title: "Sequential decision-making"
type: concept
tags:
  - 🧠Concept
  - 🏥ClinicalReasoning
aliases:
  - Sequential diagnosis
related_papers:
  - "[[Synthetic Example: Reader-Study Evaluation of an Assistive Triage Model]]"
  - "[[Synthetic Example: Cost-Aware Diagnostic Agent]]"
  - "[[Synthetic Example: Simulated Clinic Benchmark]]"
status: 🌱Seedling
---

# Sequential decision-making

> **This note is a formatting example.** The papers it links are synthetic. It shows what a
> seedling concept note looks like before the reader has rewritten the definition.

## 📖 Definition, in my own words
> 🌱 *Placeholder — drafted by the skill, awaiting the reader's own wording.*

Deciding what to do next when the picture is still incomplete: order a test, ask a question,
or commit. What separates it from one-shot diagnosis is that each step changes what the next
step should be, so the quality of a path cannot be read off its endpoint.

## 🌐 Why it matters
Benchmarks that score only the final answer cannot distinguish a good path from a lucky one.
Any evaluation meant to say something about clinical use has to score the sequence.

## 📚 How different papers treat it
- **[[Synthetic Example: Reader-Study Evaluation of an Assistive Triage Model]]**: treats the
  path as an outcome in its own right; reports where readers diverged from the model
- **[[Synthetic Example: Cost-Aware Diagnostic Agent]]**: prices each step, so the metric is
  accuracy per unit of cost rather than accuracy alone
- **[[Synthetic Example: Simulated Clinic Benchmark]]**: holds the environment fixed and lets
  the agent choose the order, treating step count as the difficulty axis

## 🔗 Related concepts
- [[LLM co-pilot in medicine]] — the same question from the human side: who chooses the next step
- [[Evaluation beyond endpoint accuracy]] (not yet written) — the measurement problem this creates

## ❓ Open questions
- Is there a scoring rule for a diagnostic path that does not collapse back to its endpoint?
- When a model and a reader take different paths to the same answer, which one should the
  evaluation reward?

## 📝 Update log
- 2026-01-20: drafted from 3 papers

---
title: "Synthetic Example: Reader-Study Evaluation of an Assistive Triage Model"
authors: [Placeholder A, Placeholder B, Placeholder C]
journal: "arXiv:0000.00000"
date_published: 2026-01-15
tags:
  - 📝Paper
  - 🤖AI/LLM
  - 🏥ClinicalReasoning
  - 📊HeadToHead
status: 🟢Completed
aliases:
  - Synthetic-triage-reader-study
---

# Synthetic Example: Reader-Study Evaluation of an Assistive Triage Model

> **This note is a formatting example, not a real paper.** Every value below is invented to
> show the shape of a completed note. Nothing here should be cited.

📎 **Open the PDF inside Obsidian**: ![[synthetic_triage_reader_study.pdf]]

## 📌 One-line summary
In a synthetic 20-case reader study, an assistive model raised median reader accuracy from 7
to 9 of 10 while leaving the rate of confidently wrong calls unchanged.

## 🎯 Background and aim
* Earlier work reported end-point accuracy but not how readers behaved with the tool in hand
* The paper asks whether assistance changes the reasoning path or only the final answer
* Stated hypothesis: assistance improves accuracy without reducing overconfident errors

## 🔑 Methods and results
1. **Design**: cross-sectional reader study, 20 vignettes, 12 readers, blinded to arm
2. **Main results**: median score 9 of 10 assisted versus 7 unassisted (p = 0.004);
   confidently-wrong rate 15% versus 14% (not significant)
3. **Comparison**: against unassisted readers only; no second tool was evaluated
4. **Limitations**: authors note the vignettes were curated, the reader pool was single-site,
   and no downstream patient outcome was measured

## 💡 My reading
* The accuracy gain and the flat error rate point the same way — assistance helps the cases
  readers were already close on, not the ones they were confidently wrong about
* Worth checking whether our own evaluation would even detect that split, or only report the
  headline accuracy

---
## Related notes
* [[MOC]]
* [[Sequential decision-making]]
* [[LLM co-pilot in medicine]]
* [[Reader study design]]

# PAPER_QUEUE.md — example

Progress tracker for batch note-writing. Lives in the vault so a large backlog survives
across sessions. Markers: ✅ done · 🟡 in progress · ⬜ pending · ⏭️ skipped.

**Batch size**: 5 subagents × 5–6 papers. Update this file after each batch, before starting
the next.

## Status

| | Papers |
|---|---|
| ✅ Completed | 12 |
| 🟡 In progress | 5 |
| ⬜ Pending | 8 |
| ⏭️ Skipped | 2 |
| **Total** | **27** |

## Batch 1 — Clinical reasoning (✅ done 2026-01-18)

| | Source text | Note |
|---|---|---|
| ✅ | `triage_reader_study.txt` | `Synthetic triage reader study.md` |
| ✅ | `cost_aware_agent.txt` | `Cost-aware diagnostic agent.md` |
| ✅ | `simulated_clinic_bench.txt` | `Simulated clinic benchmark.md` |
| ⏭️ | `triage_reader_study_supplement.txt` | supplement to row 1 — no separate note |

## Batch 2 — Evaluation methodology (🟡 running)

| | Source text | Note |
|---|---|---|
| 🟡 | `rubric_agreement.txt` | `Rubric agreement across raters.md` |
| ⬜ | `judge_calibration.txt` | `LLM-as-judge calibration.md` |
| ⬜ | `benchmark_contamination.txt` | `Benchmark contamination.md` |

## Batch 3 — Deployment (⬜ pending)

| | Source text | Note |
|---|---|---|
| ⬜ | `silent_trial.txt` | `Silent-mode deployment trial.md` |
| ⏭️ | `slides_overview.txt` | slide deck, not a paper |

## Concept extraction log

- 2026-01-20: 12 notes reached → extracted `Sequential decision-making`,
  `LLM co-pilot in medicine`, `Uncertainty and hallucination` (all 🌱, awaiting the reader's edits)
- Next pass at ~20 notes

## Notes

- `benchmark_contamination.pdf` is a scan; extraction is thin. Re-extract with more pages or
  check the source before writing the note.

# Extracting Atomic Concept Notes

Literature notes capture what each paper said. Concept notes hold what the reader now
understands across papers. The second layer is where a vault stops being an archive.

## When

Once **10+ literature notes** exist. Below that there is nothing to cross-reference, and the
notes produced are definitions rather than syntheses.

Natural triggers: the user asks what concepts to pull out; a batch just finished; three or
more notes circle the same unlinked idea; the user is preparing to write.

## What qualifies

A phrase earns a concept note when all five hold:

1. It appears in **3+ literature notes**
2. Someone could **learn something** from the note
3. Different papers **treat it differently** — there is something to compare
4. It is **stable**, not a passing label
5. It needs **synthesis**, not a definition lookup

Entities fail criterion 5 and should not become concept notes: model names (GPT-4, Claude),
datasets (MedQA), journals, organizations, evaluation formats. They are references, and a
note about one is a stub that never grows.

Borderline: technique names. `chain of thought` works as a concept if the vault holds papers
that disagree about it; it is a stub if they merely use it.

## Process

**1. Scan for recurring links and terms**

```bash
grep -ho "\[\[[^]]*\]\]" <vault>/Literature/*.md | sort | uniq -c | sort -rn | head -30
```

Or have a subagent read the literature notes and report term frequency with the notes each
term came from.

**2. Filter** — drop entities, keep what passes all five criteria.

**3. Draft** — one note per candidate, following `templates.md`: the papers it draws on, a
synthesized definition, a comparison of how each paper treats it, and two or three open
questions.

**4. Hand back for the user's own edit.** Present drafts and say plainly that the definition
section is a placeholder until the user rewrites it. A concept note the user has not touched
is still a summary wearing a concept note's frontmatter.

## Growth stages

```
🌱 Seedling   →   🌿 Growing    →   🌳 Mature
  ~3 papers       5–10 papers       10+ papers
  AI-drafted      user-edited       user's own position
```

Every note this skill creates starts at 🌱Seedling. Only the user's edits move it along.

## Guarding against sprawl

Roughly **one concept note per 5–7 literature notes** is healthy. Warning signs:

- Dozens of concept notes and none the user has edited — slow down
- Many concepts linked from a single paper — wait for corroboration
- Near-duplicates (`CoT`, `CoT reasoning`, `Chain-of-Thought`) — merge them

## Worked example

Ten medical-AI papers, with the concepts each literature note mentions:

| Literature note | Concepts mentioned |
|---|---|
| AgentClinic | sequential decision-making, multi-agent, bias simulation |
| GPT-4 RCT | management reasoning, LLM co-pilot, clinical trial |
| Med-Gemini | uncertainty-guided search, multimodal, RAG |
| AMIE DDx | differential diagnosis, OSCE, self-play |
| MAI-DxO | sequential diagnosis, cost-aware, multi-agent |
| R-IDEA | clinical reasoning, physician comparison |
| Medprompt | prompt engineering, few-shot, ensemble |
| MAIRA-2 | radiology report, grounded generation |
| HealthBench | physician rubric, LLM-as-judge |
| Reasoning-model CPC | reasoning models, case conference |

**Extracted** (3+ appearances, all five criteria): sequential decision-making; management
reasoning; LLM co-pilot in medicine; uncertainty and hallucination; healthcare benchmarking;
multi-agent systems.

**Not extracted** (entities): GPT-4, o1, Med-PaLM 2; MedQA, NEJM CPC; OSCE; Google, Microsoft.

# Subagent Prompt for Batch Note Writing

Reusable prompt for converting extracted PDF text into literature notes. Five subagents ×
5–6 papers each is the working batch; launch them in one message so they run concurrently.

## Filling it in

| Placeholder | Value |
|---|---|
| `{DOMAIN}` | the field, e.g. "medical AI", "interventional radiology" |
| `{TEXT_DIR}` | the extracted-text cache from `scripts/extract_pdfs.py` |
| `{LIT_DIR}` | `<vault>/Literature/` or the vault's existing equivalent |
| `{ASSIGNMENTS}` | this agent's table of text file → destination note filename |
| `{EXISTING_CONCEPTS}` | concept notes already in the vault, so links point somewhere real |

## Prompt

```
You are writing Obsidian literature notes for {DOMAIN} papers. Each note is written from a
text file extracted from the source PDF, which you read with the Read tool.

INPUT   {TEXT_DIR}   (one .txt per paper, first 12 pages)
OUTPUT  {LIT_DIR}    (one new .md per paper, created with the Write tool)

Assigned papers:
{ASSIGNMENTS}

Concept notes that already exist — link to these by exact name, and propose new ones only
when a paper genuinely introduces something absent from this list:
{EXISTING_CONCEPTS}

For each paper:

1. Read the whole text file.

2. Take from the text, and only from the text: the exact title including any subtitle; up to
   five authors as written; the journal or arXiv id; the publication date as YYYY-MM-DD; the
   design, results, comparisons, and stated limitations.

   Numbers, authors, and dates that are not in the text file do not go in the note. Do not
   supply them from memory, even for a paper you recognise. If the text does not state
   something, write "not stated in the extracted text" rather than filling the gap.

3. Write the note to {LIT_DIR} using the template below. The PDF embed link must reproduce
   the source filename exactly — take the .txt filename and change the extension to .pdf,
   character for character, preserving case, spaces, and punctuation.

4. If a note of that name already exists, skip it and report it. Do not overwrite.

Template — follow it exactly, including frontmatter field names:

---
title: "exact title from the text"
authors: [Family1, Family2]
journal: "journal or arXiv id"
date_published: YYYY-MM-DD
tags:
  - 📝Paper
  - 🤖AI/LLM
  - {domain tag}
status: 🟢Completed
aliases:
  - ShortAlias
---

# Paper title

📎 **Open the PDF inside Obsidian**: ![[exact_source_filename.pdf]]

## 📌 One-line summary

## 🎯 Background and aim
*
*

## 🔑 Methods and results
1. **Design**:
2. **Main results** (figures exactly as the text gives them):
3. **Comparison**:
4. **Limitations**:

## 💡 My reading
*
*

---
## Related notes
* [[existing concept note]]
* [[proposed new concept]]

Report back: notes written, notes skipped as existing, and any paper whose text file was too
short or too garbled to write from.
```

## Why the text-file constraint is absolute

A subagent handed a PDF path it cannot open does not stop and report the failure. It writes
the note anyway, from whatever it knows about a paper with that name — and the output is
fluent, correctly formatted, and wrong in the numbers. Sample counts, IQRs, author lists, and
p-values have all been produced this way. The `.txt` indirection is the only reliable guard.

# Tag Vocabulary

Tags exist so the graph view and Dataview queries stay usable. They are a filtering surface,
not a metadata dump. Three well-chosen tags beat ten exhaustive ones.

Emoji prefixes keep Obsidian's tag panel scannable by group.

## Source type — always exactly one

- `📝Paper` — peer-reviewed or preprint research paper
- `📰Article` — blog post, news piece, whitepaper
- `📘Guide` — tutorial, how-to, reference guide
- `🧪TechReport` — model technical report or system card

## Note type

- `🧠Concept` — atomic concept note
- `💭Thought` — the reader's own argument, drawing on several concepts
- `🗺️MOC` — map of content
- `📚Reference` — glossary or lookup note

## Technology

- `🤖AI/LLM` — general large-language-model work
- `🤖AI/{ModelFamily}` — a specific family when the paper is about it (`🤖AI/Claude`,
  `🤖AI/Gemini`, `🤖AI/Llama`, `🤖AI/DeepSeek`)

## Clinical domain

- `🏥ClinicalReasoning` — diagnostic or management reasoning
- `🏥ClinicalSimulation` — simulated clinical environments
- `🏥ClinicalTrial` — randomized or controlled trials
- `🏥ClinicalApplication` — real-world deployment studies
- `🏥DifferentialDiagnosis` — differential-diagnosis focused

## Technical approach

- `🧬RAG` — retrieval-augmented generation
- `🤝MultiAgent` / `🤝Agent` — multi-agent or single-agent frameworks
- `🔤Prompting` — prompt engineering
- `🎓Training` — pre-training or post-training methods
- `🧠Reasoning` — reasoning models and chain-of-thought work
- `👁️Multimodal` — vision-language or audio

## Evaluation and safety

- `📊Benchmark` — benchmark construction or results
- `📊Evaluation` — evaluation methodology
- `📊HeadToHead` — one system directly against another
- `⚠️Safety` — safety-focused
- `🛡️RedTeam` — adversarial testing
- `⚖️Regulation` — regulatory frameworks

## Adding a tag

Check first whether an existing tag covers it. If it genuinely does not, keep the
emoji + PascalCase form (`📊NewConcept`). Avoid tags implied by the vault itself — `🏥Medicine`
on every note in a medical vault filters nothing. Cap a note at four or five tags.

Korean vaults using Korean tag names: `locale/ko/note_templates.md`.

# Note Templates

Two templates: literature notes (one per paper) and atomic concept notes (synthesized across
papers). Match the frontmatter field names exactly — Dataview fails silently on a renamed
field, returning an empty table rather than an error.

Korean-structured vaults: see `locale/ko/note_templates.md`.

## 1. Literature note

Save to `<vault>/Literature/{short descriptive title}.md`.

```markdown
---
title: "Exact paper title, including any subtitle"
authors: [Family1, Family2, Family3]   # max 5, then "et al."
journal: "Journal name, or arXiv:XXXX.XXXXX"
date_published: YYYY-MM-DD
tags:
  - 📝Paper
  - 🤖AI/LLM
  - 🏥ClinicalReasoning
status: 🟢Completed
aliases:
  - ShortAlias
---

# Paper title

📎 **Open the PDF inside Obsidian**: ![[exact_source_filename.pdf]]

## 📌 One-line summary
Who did what, how, and how much — one sentence, from the text.

## 🎯 Background and aim
* What earlier work left unresolved
* The problem this paper attacks
* The hypothesis, if stated

## 🔑 Methods and results
1. **Design**: n, dataset, evaluation protocol — as written in the text
2. **Main results**: exact figures only (e.g. "accuracy 91.1%", "p < 0.001",
   "median 10, IQR 9–10")
3. **Comparison**: against baselines or competing models
4. **Limitations**: the ones the authors state

## 💡 My reading
* What this implies for my own work — the reader's judgement, not the authors' claims
* Second implication

---
## Related notes
* [[domain hub note]]
* [[existing concept note]]
* [[proposed new concept]]
```

### Rules

1. **Field names are fixed**: `title`, `authors`, `journal`, `date_published`, `tags`,
   `status`, `aliases`.
2. **`authors` is a YAML list** of at most five entries; append `et al.` beyond that.
3. **`date_published` is `YYYY-MM-DD`.** Year-only sources use `YYYY-01-01` with a note in
   the body.
4. **Tags carry emoji prefixes** — see `tag-vocabulary.md`.
5. **`status: 🟢Completed`** only once the note is written from verified extracted text.
6. **The PDF embed is mandatory** and goes directly under the H1, reproducing the source
   filename exactly.
7. **Every number comes from the extracted text.** Never from model knowledge.
8. **The related-notes section carries at least four wikilinks** — a domain hub, at least one
   cross-folder link, and the concept notes this paper touches. A note with no outbound links
   is an island, and islands are never revisited.

## 2. Atomic concept note

Save to `<vault>/Concepts/{concept name}.md`.

```markdown
---
title: "Concept name"
type: concept
tags:
  - 🧠Concept
  - (domain tag)
aliases:
  - AlternativeName
related_papers:
  - "[[literature note 1]]"
  - "[[literature note 2]]"
  - "[[literature note 3]]"
status: 🌱Seedling      # 🌱Seedling | 🌿Growing | 🌳Mature
---

# Concept name

## 📖 Definition, in my own words
The reader's understanding, not the textbook phrasing and not a sentence copied from any
paper. This section is the whole point of the layer.

## 🌐 Why it matters
Why this concept earns attention in this domain — the reader's judgement.

## 📚 How different papers treat it
- **[[Paper A]]**: its treatment, with a specific quoted claim
- **[[Paper B]]**: a different angle or method
- **[[Paper C]]**: a contrary or complementary position

## 🔗 Related concepts
- [[another atomic concept]] — how they connect
- [[a concept not yet written]] — related, still a stub

## ❓ Open questions
- What remains unanswered about this concept
- What a future paper would have to show

## 📝 Update log
- YYYY-MM-DD: drafted from {N} papers
```

### Rules

1. **`type: concept`** separates these from literature notes in every query.
2. **Status lifecycle** — 🌱Seedling: drafted, ~3 papers. 🌿Growing: 5–10 papers, the reader
   has edited it. 🌳Mature: 10+ papers, the reader's own position is established.
3. **The definition must be rewritten by the reader.** A concept note that is still entirely
   AI-drafted is a summary, not a second layer. Create it as 🌱Seedling and say so.
4. **`related_papers` entries are quoted wikilinks**: `"[[note title]]"`.
5. **Open questions are never left empty** — one or two minimum. They are the seed of the
   next layer.

## 3. Filename conventions

**Literature notes** — 3–5 keyword concepts, space-separated:

- ✅ `Clinical reasoning GenAI vs physicians.md`
- ✅ `MAIRA-2 grounded radiology report.md`
- ❌ `towards accurate differential diagnosis with LLM clinicopathological NEJM.md` (the PDF's full title)
- ❌ `paper_001.md` (carries no meaning)

**Concept notes** — the concept, with an English alias in parentheses when the field uses one:

- ✅ `Sequential decision-making.md`
- ✅ `Management reasoning.md`
- ❌ `Decision making.md` (too general to synthesize)
- ❌ `SDM.md` (an abbreviation alone)

# The Three-Layer Vault

Read this when the user asks how the pieces fit together, or why the skill refuses to write
certain notes.

## The layers

```
┌──────────────────────────────────────────────────────┐
│  Layer 3 — Thought      Thoughts/                    │
│  The reader's own questions and hypotheses            │
│  USER-AUTHORED. This skill never writes here.         │
├──────────────────────────────────────────────────────┤
│  Layer 2 — Atomic       Concepts/                    │
│  Concepts synthesized across papers                   │
│  SKILL DRAFTS, USER REWRITES.                         │
├──────────────────────────────────────────────────────┤
│  Layer 1 — Capture      Literature/  +  PDFs/        │
│  One templated note per paper, plus the sources       │
│  SKILL GENERATES from extracted text.                 │
└──────────────────────────────────────────────────────┘
```

Links alone do not make a second brain — they give topology without meaning. What makes the
vault usable is that captured knowledge, the reader's understanding of recurring ideas, and
the reader's own open questions live in separate places and are maintained differently.

Only layer 1 scales through automation. Layer 2 is a collaboration. Layer 3 is not the
skill's to write.

## What this skill does at each layer

**Layer 1 — automated.** Extract PDF text, write one templated note per paper, keep
frontmatter consistent so Dataview resolves, embed the PDF for split-view reading, track
progress in the queue file.

**Layer 2 — drafted, then handed over.** Identify ideas recurring across 3+ notes, draft
concept notes with the definition section explicitly marked as a placeholder, propose which
literature notes should link to each, and say when enough papers have accumulated to extract
more.

**Layer 3 — untouched.** The skill may point out tensions between papers that look worth
thinking about. It does not write thought notes. What is worth thinking about is the reader's
call.

## Folder layout

```
vault/
├── MOC.md                    # dashboard, usually Dataview-driven
├── Literature/               # layer 1 — one note per paper
├── Concepts/                 # layer 2 — atomic concepts
├── Thoughts/                 # layer 3 — user only
└── PDFs/                     # the source files the notes embed
```

Text extraction caches outside the vault (`~/.local/cache/paper-vault-texts/` by default) so
intermediate files never sync to other devices.

**Honor the layout that exists.** Vaults in the wild use `02_research/논문/`, `papers/`,
`Reading/`, and many others. Match what is there and never rename a folder the user already
organizes around. Korean-structured vaults: `locale/ko/note_templates.md`.

## End to end

**First run** — confirm PyMuPDF; agree the vault paths; create the literature folder and the
text cache; extract; build `PAPER_QUEUE.md` from the PDF list; run the first batch of five
subagents.

**Each subsequent batch** — extract any new PDFs, group the batch thematically, run five
subagents, mark the queue, and report which concepts are approaching the 3-paper threshold.

**Concept pass, every 10+ papers** — scan for recurring terms, filter against the criteria in
`concept-extraction.md`, draft three to five notes, hand them to the user to rewrite the
definitions.

## When the vault is working

The user opens a note and reads the PDF beside it. Dataview tables populate. Concept notes
carry three or more backlinks. The user has started writing in layer 3 unprompted.

And the one that decides whether any of the rest matters: **the user trusts the notes enough
to cite them.** That trust is broken by a single fabricated number, and it does not come
back — which is why extraction to text, and never writing from model knowledge, is the rule
the others are built around.

"""
Extract text from a folder of PDFs using PyMuPDF.

Usage:
    python extract_pdfs.py <pdf_folder> <output_folder> [max_pages]

Default max_pages: 12 (covers abstract + intro + methods + results + discussion
for most research papers).

Output: One .txt file per PDF in the output folder, with page breaks marked by
'===PAGE BREAK==='.
"""

import fitz  # PyMuPDF
import sys
import os


def extract_pdf_pages(pdf_path, output_dir, max_pages=12):
    """Extract text from the first max_pages of a PDF to a .txt file."""
    if not os.path.exists(pdf_path):
        print(f"ERROR: PDF not found at {pdf_path}")
        return False

    filename = os.path.basename(pdf_path)
    base_name = os.path.splitext(filename)[0]
    out_path = os.path.join(output_dir, f"{base_name}.txt")

    try:
        doc = fitz.open(pdf_path)
        num_pages = min(len(doc), max_pages)
        texts = []
        for page_num in range(num_pages):
            page = doc.load_page(page_num)
            texts.append(page.get_text("text"))

        full_text = "\n\n===PAGE BREAK===\n\n".join(texts)

        with open(out_path, "w", encoding="utf-8") as f:
            f.write(full_text)

        return True
    except Exception as e:
        print(f"FAILED {filename}: {e}")
        return False


def extract_folder(pdf_folder, output_folder, max_pages=12):
    """Extract text from all PDFs in a folder."""
    os.makedirs(output_folder, exist_ok=True)
    count = 0
    failed = 0

    for filename in os.listdir(pdf_folder):
        if filename.lower().endswith(".pdf"):
            pdf_path = os.path.join(pdf_folder, filename)
            if extract_pdf_pages(pdf_path, output_folder, max_pages):
                count += 1
            else:
                failed += 1

    print(f"Extracted {count} PDFs (up to {max_pages} pages each)")
    if failed:
        print(f"Failed: {failed}")
    return count, failed


if __name__ == "__main__":
    if len(sys.argv) < 3:
        print("Usage: python extract_pdfs.py <pdf_folder> <output_folder> [max_pages]")
        sys.exit(1)

    pdf_folder = sys.argv[1]
    output_folder = sys.argv[2]
    max_pages = int(sys.argv[3]) if len(sys.argv) > 3 else 12

    if os.path.isdir(pdf_folder):
        extract_folder(pdf_folder, output_folder, max_pages)
    elif os.path.isfile(pdf_folder):
        # Single PDF
        os.makedirs(output_folder, exist_ok=True)
        extract_pdf_pages(pdf_folder, output_folder, max_pages)
    else:
        print(f"ERROR: {pdf_folder} is not a file or directory")
        sys.exit(1)