Keeping It Alive: Patterns That Stick

Why Most Second Brains Die

The personal knowledge management space has a well-documented graveyard problem: people build elaborate systems, use them intensively for a month, and then quietly abandon them. The systems do not fail because they are bad. They fail because they create more friction than they eliminate. Ambitious tagging hierarchies, daily review rituals, elaborate linking conventions — within a few months the system itself becomes the work.

The LLM Wiki pattern is specifically designed to avoid this. But it still has failure modes worth knowing about:

• Abandoning CLAUDE.md updates. If you stop updating CLAUDE.md when you change a convention, Claude defaults to generic behavior in new sessions. The operating manual must stay current or the system drifts toward a generic chatbot.
• Over-engineering the schema. It is tempting to create elaborate category hierarchies in index.md or complex frontmatter templates. Start simple. Add structure only when the absence of it actually causes a problem you encounter.
• Dumping without ingesting. Dropping files in raw/ is only half the step — they need to be ingested. If raw/ accumulates files without the Dream Sequence running, the wiki stops growing. Make sure the schedule is actually active.
• Using it like a search engine. The second brain is not for finding something specific you know you saved. It is for synthesizing things you only half-remember or have not explicitly connected. If you are using it to locate a specific article, that is a symptom of treating it like a notes app.

The Minimal Viable Workflow

The simplest workflow that keeps the system alive and improving:

1. Feed it casually. When you read something interesting, clip it or paste the link and say "add this." Do not think about which category it belongs to — Claude figures that out during ingest.
2. Ask questions, not searches. When you want to remember something or think through an idea, ask: "what do I know about X?" or "what are the tradeoffs between A and B?" The answer synthesizes everything you have ingested, not just keywords.
3. Save good answers. When Claude gives you a synthesis you will want again, say "save that." It becomes a permanent page.
4. Let the Dream Sequence run. Do not worry about contradictions, orphan pages, or stale claims. The weekly schedule handles all of that. Your job is not maintenance — it is curation.
5. Capture sessions. At the end of any substantive Claude conversation, say "save this session." The things you think through in conversation are often the most valuable inputs to the knowledge base.

Five actions. Three of them are things you already do naturally (read things, ask questions, think out loud). The system handles the rest.

How the Brain Compounds Over Time

The compounding effect is not obvious in the first week. In week one, you have a folder structure and a few wiki pages. At one month, you have a coherent index and pages that are starting to reference each other. At three months, you have a knowledge base with real density — pages that have been updated multiple times as new sources refined the picture. At six months, the value is qualitatively different.

Karpathy describes his own experience: one research wiki grew to approximately 100 articles and 400,000 words of synthesized content without him writing any of it directly. The compounding is quiet until suddenly it is not: you ask a question and get an answer that synthesizes six months of reading in a single paragraph, with citations to pages you barely remember ingesting.

This is the target state. Getting there requires only consistency, not effort. Feed it what you read. Ask it questions. Save the good answers. The rest is automatic.

Your Job vs. the LLM's Job

The clearest summary of the division of labor:

• Your job: Source curation (deciding what is worth adding), direction (deciding what questions are worth asking), quality judgment (deciding when an answer is good enough to save).
• The LLM's job: Everything else — summarizing, cross-referencing, filing, bookkeeping, updating, synthesizing, maintaining consistency, resolving contradictions.

The moment you start doing the LLM's job — manually editing wiki pages, manually updating cross-references, manually maintaining processed.md — you have introduced friction. Not because manual editing is wrong, but because it accumulates. Systems fail when maintenance becomes your job. Keep your job to judgment calls; let the LLM handle the bookkeeping.

The Five Commands You Actually Need

The entire interface to your self-maintaining knowledge base:

• add this — paste a URL or text and trigger ingest
• what do I know about ___ — query the brain with a topic or question
• save that — file a good query answer back as a permanent wiki page
• save this session — capture the current conversation's key takeaways
• run dream sequence — manually trigger the health check (or let the schedule handle it)

Five commands. That is the entire interface to a knowledge base that maintains itself, compounds over time, and gets smarter every time you use it.