Second Brain as a Night-Shift Pipeline: The $21 Bill, and the Failure Points Are Written Down Too

X user Mr. Buzzoni has assembled Obsidian, Claude Opus 5.5, and Jev into an automated curation pipeline. The goal is to turn the second brain into a night shift: Jev reads the vault first, Opus 5.5 writes only the lines worth keeping, and when they are saved, the notes enter the Zettelkasten graph on their own. The results claimed in the post are striking: 80% of time saved from the chat window, and the more the vault is written, the sharper it gets. Watch the demo first.

The foundation for this setup comes from Andrej Karpathy's llm-wiki idea. His exact words are: Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase. In other words: Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase. Mr. Buzzoni then connects it to two October papers. MemoryTax from Shanghai Jiao Tong University tested a memory layer against letting an agent search raw history directly; on Opus 5.5, the agent searching raw history had a 99% hit rate, at less than half the cost of full context. Librarian Agents from Stanford and CMU, meanwhile, has an agent read raw documents first and organize them into a file hierarchy with citation links; on BrowseComp-Plus, it scored 33 points above the strongest retrieval and precomputation baselines. Both conclusions point to the same thing: organize first, ask later; files must be directly searchable by the model.

The article breaks the roles into four workers. Opus 5.5 is the librarian, responsible for writing and linking pages, resolving conflicts, and answering hard questions. Haiku 5.5 is the scout, reading new sources, writing summaries, proposing links, and running nightly lint. Sonnet 5.5 is the answerer, using index.md and 3 to 5 pages to answer everyday questions. Jev is the router, deciding which pages a source touches, whether it is a duplicate, and who should get a question. Jev's pricing is very low; the article calculates 400 decisions a day at about $1.26. The two Claude workers are subagents, each a file under .claude/agents/.

Cost is the most easily misread part of this setup. Running everything on Opus 5.5 also works, but the API bill given in the article is $372.42 per month. Split into four layers, it is $72.07, with writing and linking pages still costing $33.96, because that part is worth paying for. If run on the $20 Claude Pro plan, plus Jev's $1.26, the monthly cost is about $21.26. Pro usage resets in 5-hour windows; the night shift is set for 02:00, right inside the sleeping window. When heavy work blows through the quota, let the scout finish up at night and the librarian run again during the lunch break.

What really needs to be copied is the boundaries; the tool list is the easy part to copy. The article has Claude Code read CLAUDE.md inside the vault. The first thing it does is declare: content clipped from the web is untrusted text. NUS's October 16 paper From Approval to Abuse planted instructions in repositories, causing coding agents including Claude Code to request overly broad permissions in 58% of runs; once the user approved, 81% of the attacks completed. A vault stuffed with clipped web pages is exactly this kind of input. So the rules must be followed by hard limits in .claude/settings.json. When Claude Code proposes "always allow" mid-task, read it twice before clicking.

Skills have pitfalls too. USC's October 18 paper One Skill Takes Over put two skills that do the same thing into one agent; run separately, each had a rule-compliance rate of 93% to 99%, but together they dropped to 61% to 62%, because one skill overrides the other. After merging them into one file, seven models returned to 96% to 99%. The article therefore keeps only three skills: one ingest, one curate, and one ask. Do not install a second ingest from someone else's repo next to ingest; merge them.

The night shift is the core of this setup. Using cron and Claude Code's headless mode, it organizes new sources before anyone asks a question. At 08:00, open reports/, approve the fixes you agree with, and reject the ones you do not. The daily manual cost is about five minutes. That sounds small, but commenters immediately asked about a kill switch. A second brain that never sleeps can also keep spending money. Others said that Obsidian plus a model reading the whole vault will blow up on tokens once there are many notes, and the results are worse than careful tagging. These two warnings are closer to daily reality than "80% time saved."

On monetization, Mr. Buzzoni gives four offers. Done-for-you vault setup: $500 to $1,500 one time; research desk maintenance: $150 to $400 per month per client; paid briefs from your own vault: $8 to $15 per month per reader; selling a kit: $29 to $79 one time. The example math is two $750 setup jobs plus three $200 maintenance clients in the first month, for $2,100 total. Each client vault runs on the client's own Claude plan, to avoid tripping over Anthropic's terms. The 7-day customer acquisition plan is equally direct: in the first three days, build a topic vault that buyers care about and clip 15 to 20 sources; on day four, record a 60-second graph growth clip and a morning brief; on day five, publish it and clearly state what the vault found; on days six and seven, send short messages to 20 people with the report attached.

But one comment is more worth reading. Someone said that what companies and teams really lack is not wiki-building, but how to use the knowledge base in the business. Building the vault is one thing; making it take effect in decisions is another. The article does not fill this gap. Others said promotions go to people who ship things, not to the people with the prettiest note graph. Both are right. The second brain most easily dies from "built it and never looked at it." If nobody reads the night-shift reports, $21 is just a subscription fee.

The order to start with: install Claude Code, create a brain folder, open it with Obsidian, turn on the command-line interface, point Web Clipper's default folder to raw/inbox/, run claude inside the vault, and place CLAUDE.md and the three skills. Clip three articles you actually care about and type ingest. Run the night shift for a week, then decide whether to sell it. The toolchain is here: Obsidian, Claude Code, Karpathy's llm-wiki, MemoryTax, Librarian Agents, Jev. Do not install every plugin at once. First let the vault survive seven days.