Use Cases

Understand a new codebase

You join a team. The repo has 200 files across 15 modules. Instead of reading code file by file:

cd the-repo && llm-wiki .
llm-wiki query gods              # what are the core abstractions?
llm-wiki query community 0       # what's in the biggest cluster?
llm-wiki query path Auth Database # how does auth reach the DB?

Open graph.html to visually explore — click nodes, filter by community, trace paths.

Build a research corpus

Drop papers, notes, tweets, blog posts into a folder. Build once, query across all of them:

llm-wiki add https://arxiv.org/abs/2401.12345 --author "Smith et al."
llm-wiki add https://karpathy.ai/blog/llm-wiki
llm-wiki .
llm-wiki query search "attention mechanism"
llm-wiki query path "transformer" "diffusion"

The graph shows connections between papers you didn’t know were related.

Personal /raw folder (Karpathy workflow)

The original concept: one folder where you drop everything.

~/raw/
  papers/attention-is-all-you-need.pdf
  notes/meeting-2026-04.md
  screenshots/architecture-whiteboard.png
  code/prototype/main.py

cd ~/raw && llm-wiki .

Code gets AST extraction. Docs get structural parsing. Use /wiki . in Claude Code for deep extraction of PDFs and images.

Vietnamese historical documents

Tested with Thượng Chi Văn Tập (Phạm Quỳnh) and Nam Phong Tạp Chí:

DOCX (43K chars Vietnamese text): 3 hub nodes → 103 entities after agent extraction
Scanned PDF (1007 pages): 0 text via pypdf → 86 entities after agent vision
HEIC scans (13 images): 0 text → 29 entities with character names, places, themes

Agent mode preserves Vietnamese labels and extracts culturally specific entities.

Documentation audit

Run on your docs folder to find:

God nodes — most referenced concepts (are they well-documented?)
Orphan nodes — concepts with no connections (missing cross-references?)
Surprising connections — unexpected links between docs
Community structure — do your docs cluster the way you expect?

llm-wiki query gods    # what's most referenced?
llm-wiki query stats   # how much is INFERRED vs EXTRACTED?

Living wiki for a team

The graph compounds over time. Each developer session adds knowledge:

# Day 1: initial build
llm-wiki .

# Day 5: new files added, rebuild (cache skips unchanged)
llm-wiki .

# Day 10: file an insight from debugging session
mkdir -p wiki-out/ingested
echo "# SharedConnection: GraphStore and MemoryStore share SQLite connection pool" \
  > wiki-out/ingested/insight_shared_connection.md
llm-wiki .   # insight now in graph

# Weekly: health check
llm-wiki lint

Claude Code integration

After building, Claude answers questions without re-reading source files:

# Install skill
mkdir -p ~/.claude/skills/my-llm-wiki
cp "$(python -c 'import my_llm_wiki; print(my_llm_wiki.__path__[0])')/SKILL.md" ~/.claude/skills/my-llm-wiki/

# /wiki .     → structural + agent extraction (DOCX, PDF, images)
# /wiki lint  → graph health check
# "what connects GraphStore to Settings?" → Claude uses llm-wiki query

The skill implements the full living wiki cycle — monitor, rebuild, lint, write-back.