CLAUDE.md

Schema for LLM wiki maintenance. Obsidian is the IDE; Claude is the programmer; this vault is the codebase.

What This Repo Is

Obsidian personal knowledge base ("second brain") backed by Git. Not software project — no build steps, tests, package managers. Content in Markdown, synced to GitHub via Obsidian Git plugin (auto timestamped backup commits).

Remote: git@github.com:vatsal30/Obsidian_backup.git

Content Organization

Rebuilt 2026-07-28 into 5 top-level life-area folders (was: technical-domain folders + a catch-all Private/). Private/ no longer exists — everything it held now lives in one of the 5 folders below, or in Archive/. Each folder has its own index.md with a recommended reading order; read that before working inside it.

Interview Prep/ — 6 folders, 2 root files, no incidental nesting

Rule: nothing gets a wrapper subfolder just for taxonomy. Reorganized 2026-07-29 around what the content actually is, not a technical-domain split. Content/DSA/ and Content/System Design/ are flat internally too (numbered technique files, concept files — no Techniques for Interviews/ or Concepts/ sub-subfolders). Only Content/ itself stays deeply nested by category (Backend/Frontend/System Design — ~150 files, genuinely needs it, has its own index.md).

File Conventions

Frontmatter — all notes use YAML frontmatter:

---
title: "Title"
category: "Category"
date: "YYYY-MM-DD"
tags: ["tag1", "tag2"]
---

Naming — Title Case or PascalCase (e.g., Dynamic Programming.md, JavaScript.md). Diagrams use Excalidraw format (.excalidraw.md).

.gitignore excludes .obsidian/workspace.json, .obsidian/workspace-mobile.json, .trash/, .DS_Store.

Git Workflow

• Obsidian Git auto-commits with messages like vault backup: YYYY-MM-DD HH:MM:SS
• Manual commits: chore: ..., feat: ...
• Don't amend automated backup commits

Mermaid Diagrams

• Use <br/> for line breaks inside node labels — never \n
• Example: Node[Line 1<br/>Line 2] ✅ — Node[Line 1\nLine 2] ❌

Installed Community Plugins

• obsidian-tasks-plugin — tasks use - [ ] syntax with due dates (📅 YYYY-MM-DD)
• obsidian-excalidraw-plugin — diagrams as .excalidraw.md files
• obsidian-local-rest-api — vault via local HTTP API (port 27123 default)
• calendar — daily/weekly notes integration
• table-editor-obsidian — GFM table editing
• obsidian-importer — import from other formats; no ongoing structural impact
• vscode-editor — VS Code keybindings in Obsidian
• obsidian-style-settings — theme customization; no content impact

---

Three-Layer Wiki Architecture

Layer 1 — Raw Sources (raw/): Immutable. Web clips, articles, papers, PDFs. LLM reads, never modifies. Frontmatter: source_url, date_clipped, ingested: true/false.

Layer 2 — Wiki (Interview Prep/ + Learning/): LLM-owned. Domain notes + cross-domain synthesis pages (now filed inside Learning/ and Interview Prep/ rather than a separate synthesis/ folder). LLM creates and updates. Human reads.

Layer 3 — Schema (this file): Conventions + workflows. Human + LLM co-evolve over time.

Navigation files (always update):

• index.md — catalog of all pages, one-line summary per page. Read first when querying.
• log.md — append-only operation history. Format: ## [YYYY-MM-DD] action | description.

---

Operations

Ingest

When user drops a source in raw/ and says "process this":

1. Read source file
2. Discuss key takeaways with user
3. Write summary page in Interview Prep/ or Learning/ as appropriate (cross-domain → a synthesis page, e.g. Learning/Tech Stack Overview.md)
4. Update relevant existing pages (add cross-refs, update synthesis pages)
5. Set ingested: true in source frontmatter
6. Update index.md — add new page, increment page_count
7. Append to log.md: ## [date] ingest | Source Title

One source may touch 5–15 wiki pages. That's expected.

Query

When user asks a question against the wiki:

1. Read index.md first — find relevant pages
2. Read those pages
3. Synthesize answer with [[page]] citations
4. If answer is valuable (comparison, analysis, new connection) → file it as new wiki page
5. Update index.md if new page created
6. Append to log.md: ## [date] query | Question summary

Lint

When user says "lint the wiki" or "health check": Check for:

• Orphan pages (no inbound [[links]])
• Stale claims superseded by newer content
• Concepts mentioned across pages but lacking own page
• Missing cross-references between related pages
• DSA/SD gaps vs job switch plan milestones
• Synthesis pages that need updating from new domain content

Append to log.md: ## [date] lint | Issues found: N

---

Wiki Page Frontmatter

Domain pages (anything inside Interview Prep/ or Learning/):

---
title: "Title"
category: "Category"
date: "YYYY-MM-DD"
tags: ["tag1", "tag2"]
---

Synthesis pages (filed inside Learning/ or Interview Prep/):

---
title: "Title"
type: "synthesis"
category: "synthesis"
date: "YYYY-MM-DD"
last_updated: "YYYY-MM-DD"
tags: ["synthesis", "topic1", "topic2"]
source_count: N
---

Raw sources (raw/):

---
title: "Article Title"
type: "source"
source_url: "https://..."
source_type: "article|paper|podcast|video"
date_clipped: "YYYY-MM-DD"
ingested: false
wiki_pages_updated: []
---

---

Directory Map (updated 2026-07-28)

---

Synthesis Pages

LLM-maintained cross-domain pages. Update when domain content changes.

---

LLM Ownership Rules

LLM owns: Interview Prep/ (all of it — Content, Behavioral & Past Projects, Job Switch, Monthly Plan, Coach.md, Dashboard, Weekly/Daily Tracker retro data), Learning/ (all of it), index.md updates, log.md appends, per-folder index.md files.

Human owns: Private/Finance/, Projects/, Life/, raw/ (source selection), this CLAUDE.md schema, Archive/ (LLM may only add to it — auto-archive flagged stale content, never delete, never edit existing archived content).

LLM never modifies: Files in raw/ (except setting ingested: true), .obsidian/, .gitignore, anything already inside Archive/ (add-only).

On every write — MANDATORY, NO EXCEPTIONS:

1. Append to log.md: ## [YYYY-MM-DD] action | description — do this LAST, after all other writes
2. Check if index.md needs updating (new page → add entry + increment page_count)
3. Check if any synthesis pages need updating

Exemptions (no log required): writes to log.md itself, CLAUDE.md, raw/, Private/, Projects/, Life/.

Enforcement: .claude/hooks/log_enforcer.py blocks (exit code 2) any .md write outside the exempt paths that isn't followed by a log.md append — self-enforced by the LLM in response to the hook's reminder, not an automatic block on the write itself. Violation = incomplete task.