LLM Wiki

← Docs hub

07 · Example workflows

Time
15 minutes total (read) / 1 hour each (do)
You'll need
A working wiki with at least 10 ingested sessions.
Result
Four concrete, end-to-end workflows you can adapt to your own use case.

Why this matters

The previous six tutorials taught you what the pieces are. This one shows what they look like stitched together. Pick the workflow that matches your day-job and copy it.

Workflow 1 — Daily developer loop

Who: solo developer using Claude Code for 80 % of their coding.

Cadence: one sync / day, one query / few days.

Morning:
  (start a coding session in Claude Code, normal work)

End of coding block:
  /wiki-sync

When you need to recall something:
  /wiki-query when did I last look at <topic>?

Weekly:
  /wiki-lint            (spot broken links, orphans)
  /wiki-candidates          (approve any candidates)

What you get: an always-current wiki that takes ~30 seconds of thought per day. The cost of entry is ~5 minutes; the payoff is six months of "oh yeah, I solved that in April" moments.

Workflow 2 — Team shared wiki

Who: a team where multiple people run Claude Code on the same repo.

Cadence: each person syncs locally, the wiki gets published to a shared site via GitHub Pages on every master push.

Each dev:
  clone the llm-wiki repo, run ./setup.sh
  /wiki-sync at the end of each work block
  git add wiki/ && git commit -S -m "wiki: <short summary>"
  git push

CI (on master push):
  python3 -m llmwiki build      (GitHub Pages workflow)
  → site/ published to <team>.github.io/llm-wiki

Everyone else:
  bookmark the published URL; search from Cmd+K

Non-destructive merging. Two devs sync to separate branches, each gets their own set of wiki/sources/<project>/<date>-<slug>.md files. Merging is conflict-free because every session has a unique timestamp.

Deploy guide: deploy/github-pages.md.

Workflow 3 — Bring-your-own Obsidian vault

Who: long-time Obsidian user who doesn't want to migrate 500 notes.

Cadence: daily sync writing into the existing vault; Obsidian stays the reading + editing surface.

Daily:
  llmwiki sync --vault "~/Documents/Obsidian Vault"

When a new entity lands:
  open Obsidian → navigate to Wiki/Entities/<name>.md → add prose

Next day's sync:
  sees the page, skips rewrite, appends new inbound links under
  ## Connections (idempotent)

What you get: auto-generated entity pages that become human-curated over time. Obsidian's graph view + backlinks + full-text search now cover your llmwiki output without any migration.

Deep dive: tutorials/06-bring-your-obsidian-vault.md.

Workflow 4 — Personal knowledge base with cost preview

Who: someone planning a large one-shot ingest (hundreds of past sessions) and worried about the API bill if they turn on Anthropic-backed synthesis later.

Cadence: ingest once (dummy backend), preview cost, then decide whether to enable Ollama or a cloud synthesis backend.

1. First sync uses the DummySynthesizer (local, free, canned text):
     llmwiki sync

2. Preview what an Anthropic-backed re-synthesis would cost:
     llmwiki synthesize --estimate

   Output (numbers vary):
     627 new sessions, prefix 3,944 tok
     Model: claude-sonnet-4-6 (first write)
       Prefix:    3,944 tok  $0.0148
       Fresh:     1,274 tok  $0.0038
       Output:    1,000 tok  $0.0150
       Total:                $0.0336
       + 626 subsequent sessions (cache hit):  $17.9484
     Batch total: $17.9820 (model claude-sonnet-4-6)

3. If the number is acceptable, configure a real backend:
     # For local / free: install Ollama + set synthesis.backend = "ollama"
     # For Anthropic: set synthesis.backend = "anthropic" (v1.2+)
     llmwiki synthesize

4. If the number is too high, stay on the DummySynthesizer. Your wiki
   still works end-to-end — pages are just seeded with canned prose
   until you're ready to pay.

Cache-control plumbing in llmwiki/cache.py ensures Anthropic prompt-caching reuses the stable prefix (CLAUDE.md schema + index + overview) across every session — that's where the 50-90 % savings come from.

References: reference/prompt-caching.md.

Pick one

Don't try to run all four. Start with Workflow 1 (solo / daily) for a week. Move to Workflow 2 (team) when a second person wants in. Move to Workflow 3 (vault overlay) when you stop touching Obsidian for three days in a row and feel the split. Stop at Workflow 4 if you're never going to synthesize against a paid model.

Verify your workflow is working

Three signals:

  1. You answered "wait, didn't I do this before?" with /wiki-query at least once this week.
  2. /wiki-lint has 0 errors every time you run it.
  3. Your session count on the home page is still growing after a month. (If it plateaus, you've either stopped using the agent or stopped syncing.)

Troubleshooting

"The wiki feels out of date" — when did you last /wiki-sync? If > 48 h, sync before investigating anything else.

"The answers are thin" — you're using the DummySynthesizer. Switch to Ollama (local, free) or configure an Anthropic backend in synthesis.backend. See reference/prompt-caching.md.

"Orphans keep growing" — you're creating entity pages faster than you're linking them. Spend 20 minutes with /wiki-update to link them into existing sessions.

That's the tutorial series

You're done. The complete set:

Next

→ Docs hub — reference, deployment, contributor guides, style guide, architecture.

Edit on GitHub ↗