A 131-Line Prompt Gets Opus 5.5 to Work Node by Node—Don’t Thank Him Yet

Mr. Buzzoni posted on X: a Karpathy-style prompt turns Opus 5.5 into a graph that advances node by node through a job. 131 lines, 14 tagged sections. His exact words were “send it to Opus 5.5 and thank me later.” The tweet includes a video where you can see the shape of the workflow.

Let’s be clear: the public tweet only gives this much information; the original prompt is not expanded in this material. What actually has details is a long-form article he cited, with a very direct title, How to Turn Claude Opus 5.5 + Obsidian Into $2,100 a Month With Karpathy's Second Brain (Full Guide). The article is about another thing: using Claude Opus 5.5 as a librarian to maintain an Obsidian folder as a searchable, linkable wiki, then using night-shift tasks to organize it automatically. Only when you look at the two together does the logic become complete: the node-based workflow breaks down tasks, and the second brain remembers what happened between tasks.

The comments get to engineering questions earlier than the main text. Chatis asks how 14 tagged sections handle loops—can a node call itself, or is it strictly one-pass? Raul says splitting a prompt by node can save context, but latency blows up with deep recursion, and asks whether subgraphs were tested. Teddet cares about the skeptic’s kill rate: after 30 days, is it 70% or 90%? None of this is flattery; these are questions you should ask before using it. Another reply is more cutting: “karpathy-inspired” is doing more work than all 131 lines combined. Sarcasm aside, it hits on a fact: labels cannot replace implementation.

The second brain in the long article, however, has a concrete implementation. Karpathy’s llm-wiki idea is just one sentence: Obsidian is the IDE, the LLM is the programmer, and the wiki is the codebase. Mr. Buzzoni breaks it into a brain folder: raw/inbox holds clippings, raw/assets holds images, wiki holds generated pages, CLAUDE.md writes boundaries and rules, .claude/agents holds scout and librarian, and scripts/night.sh runs the night shift. Since 1.12.7, Obsidian has a CLI that can check for orphans, unresolved, and deadends; these mechanical checks don’t spend model tokens.

The architecture is four workers revolving around Opus 5.5. scout uses Haiku 5.5 to read new sources, write summaries, and propose candidate pages; router uses Jev to make yes/no judgments, such as whether a source is duplicate and whether a question should go to Sonnet or Opus; answerer uses Sonnet 5.5 to answer everyday questions; librarian uses Opus 5.5 to write pages, connect links, and handle conflicts. The article cites two October papers to support this division of labor. MemoryTax comes from Shanghai Jiao Tong University, with 419 rounds of dialogue and 152 questions; one conclusion is that memory quality is limited by the writer—a strong reader cannot rescue what a weak writer missed. Librarian Agents comes from Stanford and CMU: first let the agent organize raw documents into a linked file hierarchy, then answer questions, scoring 33 points above the strongest retrieval baseline on BrowseComp-Plus. Both papers point to the same practice: put expensive organizing before asking questions, and leave files where the model can search them directly.

The cost accounting is also detailed. The $20 Claude Pro plan already includes Claude Code, Obsidian is free, and Jev making 400 decisions a day costs about $1.26 a month, for a total cost of $21.26. The split-API plan, estimated at 5 new sources, 20 questions, and a 500-page vault per day, costs about $72.07 a month; using Opus 5.5 for everything costs $372.42. What’s expensive is writing and linking; search and classification can be pushed down to cheaper models. The article deliberately keeps the librarian line unchanged, and the point is clear: writing quality is where you should spend money.

The security section is not decoration. The article cites NUS’s From Approval to Abuse: instructions buried in a repository can make coding agents, including Claude Code, request overly broad permissions in 58% of runs, and after the user approves, the attack completion rate is 81%. So raw/ must be read-only, and settings.json should deny Edit(raw/), Write(raw/), Bash(rm:), Bash(curl:), and WebFetch. Another paper, USC’s One Skill Takes Over, says that when two skills do the same thing, compliance is 93–99% individually, but drops to 61–62% together; only merging them into one file brings it back to 96–99%. This explains why the article insists on one job per skill, and three skills are enough.

The monetization section is the easiest place to get carried away. The article gives four options: one-time vault setup $500–1500, research-desk maintenance $150–400 a month, paid briefing $8–15 a month, toolkit $29–79. The sample math is 2 setups at $750 plus 3 maintenance clients at $200 in the first month, totaling $2100. This number is an example, not a floor. A commenter, kuch., puts it closer to reality: “second brain only pays when u actually open it, most ppl build the vault and never look at it again.” Building the vault is only the first step; spending five minutes a day reading reports and approving fixes is where the system starts to work.

If you just want to try the smallest closed loop, the article’s starting point is: install Claude Code, create the brain folder, paste CLAUDE.md, clip three articles you care about, and type ingest. The night-shift script uses cron to run mechanical checks at 02:00, has scout lint at 02:05, has librarian curate at 02:20, and generates the morning report at 02:50. The next day, open reports/, approve what you agree with, and reject what you don’t. This process isn’t mysterious, and it isn’t cheap enough to be zero, but it turns organizing from an occasional impulse into a fixed procedure.

Back to that tweet. The 131-line prompt and 14 tagged sections are an entry point, not the destination. The real value is in the constraints between nodes: who reads, who writes, who approves, and where it stops after failure. The graph growing in the video looks impressive, but the loops, latency, and kill rate asked about in the comments are what determine whether it can be used long term. Don’t rush to thank him yet; run it for a week and see whether you open your vault every day, or whether it becomes an expensive Obsidian folder three days later.