llm-wiki - Persistent LLM-Maintained Markdown Wiki

Keyword: llm-wiki · obsidian wiki · research vault · knowledge base

Use this skill when the user wants knowledge to accumulate as a maintained markdown artifact, not be rediscovered from raw files on every query.

llm-wiki turns Andrej Karpathy's gist into an operational workflow. The core pattern is simple: keep raw sources immutable, let the LLM own the wiki layer, and encode the maintenance contract in AGENTS.md or CLAUDE.md.

When to use this skill

Bootstrap a new Obsidian or markdown vault for long-lived knowledge work
Convert raw articles, papers, transcripts, or notes into source summaries plus cross-linked wiki pages
Maintain index.md and log.md as navigational primitives before adding heavier search infrastructure
File good answers back into the wiki instead of losing them in chat history
Run periodic lint passes for broken links, orphan pages, stale claims, and missing synthesis pages
Add optional helpers such as Scrapling for URL ingestion, qmd for search, or Obsidian CLI for vault automation

Instructions

Step 1: Bootstrap the vault

Create the wiki skeleton first:

bash scripts/bootstrap-vault.sh /path/to/vault

The bootstrap creates:

raw/sources/ for immutable source markdown or copied files
raw/assets/ for downloaded images and attachments
wiki/sources/ for per-source summaries
wiki/entities/ and wiki/concepts/ for durable synthesis pages
wiki/queries/ and wiki/reports/ for filed answers and higher-value outputs
index.md, log.md, and AGENTS.md

Do not skip the schema file. The schema is what makes the agent act like a disciplined maintainer instead of a generic assistant. See references/architecture.md and references/schema-playbook.md.

Step 2: Treat `AGENTS.md` or `CLAUDE.md` as the operating contract

The schema should encode the rules that stay true across sessions:

raw/ is source of truth and must stay immutable
wiki/ plus index.md and log.md are LLM-owned working artifacts
Every ingest updates the source summary, relevant synthesis pages, index.md, and log.md
Every durable query answer gets filed back into wiki/queries/ or wiki/reports/
Lint passes must look for broken links, orphan pages, stale claims, contradictions, and missing page candidates

Keep the schema short and enforceable. A small contract that the agent actually follows is better than a giant policy file nobody will reread. The starter schema from bootstrap-vault.sh is intentionally minimal; refine it with references/schema-playbook.md.

Step 3: Ingest one source at a time until the workflow is stable

If the source is already local, place it in raw/sources/ and ask the agent to process it. If the source is a URL, use the Scrapling-powered helper:

bash scripts/ingest-url.sh /path/to/vault "https://example.com/article"
bash scripts/ingest-url.sh /path/to/vault "https://app.example.com/post" --mode fetch --wait-selector article
bash scripts/ingest-url.sh /path/to/vault "https://protected.example.com/post" --mode stealth --solve-cloudflare

Expected ingest touch points:

Save the immutable raw capture under raw/sources/
Create or refresh a source page under wiki/sources/
Update relevant entity or concept pages
Update index.md
Append a chronological entry to log.md

Prefer one-source-at-a-time ingest when starting. It forces the human to inspect what the wiki changed, surface missing conventions, and refine the schema. Once the workflow is reliable, batching is fine. Use references/ingest-playbook.md for the operating checklist.

Step 4: Query against the wiki, not the raw pile

When answering questions:

Read index.md first
Open the most relevant wiki pages
Drill into raw sources only when the wiki needs grounding or conflict resolution
Cite page paths and raw source paths explicitly
File durable outputs back into the vault

Create a reusable note stub for high-value answers:

bash scripts/new-query-note.sh /path/to/vault "How does Company A differ from Company B?" --question "How does Company A differ from Company B?"
bash scripts/new-query-note.sh /path/to/vault "Q2 product thesis" --section reports --citation "[[wiki/concepts/product-thesis]]"

Use wiki/queries/ for question-shaped outputs and wiki/reports/ for more durable synthesized artifacts such as memos, comparisons, or presentation backbones. More detail lives in references/query-and-filing.md.

Step 5: Lint and repair the wiki periodically

Run the local health check:

python3 scripts/lint-wiki.py /path/to/vault
python3 scripts/lint-wiki.py /path/to/vault --format json

This script focuses on structure, not truth. It checks required files and directories, broken wiki links, and orphan pages. Use the lint output as the starting point for a human-guided cleanup pass:

merge duplicate concepts
promote heavily referenced ideas into their own page
retire stale claims superseded by newer sources
add backlinks or summary pages where the graph is too sparse

See references/maintenance-and-scaling.md for the higher-level lint checklist.

Step 6: Add search and automation only when scale justifies it

index.md plus log.md is enough for small-to-medium vaults. Add heavier tools later:

Scrapling when URL ingestion becomes common
qmd when the index is no longer sufficient for wiki search
Obsidian CLI when you want shell-driven vault automation against a running desktop app
Dataview when you start relying on frontmatter-driven tables and lists
Git from day one for version history, branching, and reviewable diffs

Do not force embeddings, MCP, or browser automation on day one. The point of this workflow is that a simple markdown repo already compounds knowledge surprisingly well.

Examples

Example 1: Bootstrap a fresh vault

bash scripts/bootstrap-vault.sh ~/vaults/company-research

Example 2: Capture a static article into raw sources and create a source stub

bash scripts/ingest-url.sh ~/vaults/company-research "https://example.com/article"

Example 3: Capture a JavaScript-rendered page

bash scripts/ingest-url.sh ~/vaults/company-research "https://app.example.com/dashboard" --mode fetch --network-idle

Example 4: Create a durable query note

bash scripts/new-query-note.sh ~/vaults/company-research "Why this market is consolidating" \
  --question "Why is this market consolidating?" \
  --citation "[[wiki/concepts/market-structure]]"

Example 5: Create a report stub instead of a query note

bash scripts/new-query-note.sh ~/vaults/company-research "Q3 diligence memo" --section reports

Example 6: Run the structure lint

python3 scripts/lint-wiki.py ~/vaults/company-research

Best practices

Keep raw/ immutable. Corrections belong in wiki pages or follow-up source notes, not in rewritten raw captures.
Update index.md and log.md on every ingest, query filing, and lint pass. If these drift, the whole workflow gets harder to navigate.
Prefer a few strong entity and concept pages over hundreds of weak fragments. Compounding happens through synthesis, not page count.
Read index.md before doing expensive retrieval work. It is the simplest useful search system for moderate vault sizes.
Distinguish facts from synthesis. Source pages should be grounded; concept pages can be more interpretive.
File good answers back into the vault. If the answer mattered once, it will probably matter again.
Use git commits to separate ingest, query, and lint operations so the wiki stays auditable.
Keep the schema alive. Update AGENTS.md when you notice repeated drift or ambiguity.

llm-wiki