CLAUDE

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

CLI Settings

If Python packages need to be installed, use pip not pip3. If Python files need to be run, use python not python3.

Project Overview

This is a research knowledge repository collecting Stanley Druckenmiller’s investment philosophy and market commentary through curated interview transcripts. It is structured as an Obsidian vault with Git-backed version control.

Directory Structure

• interviews/ — Main interviews directory containing Markdown transcripts and the conversion script
• interviews_translation/ — Markdown transcripts of interviews translated into Chinese
• scripts/ - Relevant python scripts to perform operations on the knowledge base
• indices - Indices guiding the structure of the knowledge base

Content Pipeline

New transcripts are added via:

1. Download YouTube auto-generated subtitles as .en.vtt. The filename must follow the pattern Title [VideoID].en.vtt.
2. Run the conversion script in scripts/:

   && python convert_vtt_to_md.py

This generates .md files directly in with a # Title, **Source:** YouTube link, and ## Transcript section.

Transcript Format

Each generated Markdown file has the structure:

# {Title}

**Source:** [{youtube_url}]({youtube_url})

---

## Transcript

{paragraphs of ~800 chars each, split at sentence boundaries}

Manual Cleanup: VTT → Canonical Format

Auto-generated subtitles often lack punctuation, paragraph breaks, and speaker labels. After running the conversion script, manually clean up each file to match the canonical format used in interviews/:

1. Add YAML Frontmatter

Replace the # Title + **Source:** header block with:

---
title: {Title}
source: {youtube_url}
date: {YYYY-MM-DD | YYYY}
language: en
original_transcript: {yes | no}
tags:
  - Druckenmiller
  - {topic}
  - {interview | speech}
---
# {Title}
 
---
## Transcript

2. Add Speaker Labels

Label each speaker as **Name:** at the start of their turn. Use the person’s last name or role (e.g. **Druckenmiller:**, **Joe:**, **Moderator:**). For solo speeches/TED talks, label every paragraph with the single speaker.

3. Add Punctuation

VTT auto-captions are unpunctuated. Restore:

• Sentence-ending periods, question marks, exclamation marks
• Commas at clause boundaries
• Em dashes (—) for asides and interruptions
• Quotation marks where the speaker quotes someone
• Percent signs, dollar signs, and numeric formatting (e.g. 57%, $780 billion)

4. Break into Paragraphs

Split the wall of text into logical paragraphs — one topic or argument per paragraph. Each speaker turn is a new paragraph. Long monologues should be broken at natural topic transitions.