Boris Tane wrote a great piece about how he uses Claude Code. Research, Plan, Implement. Annotation cycles on a plan.md file. Tight, disciplined, effective.
But Boris is an engineering lead at Cloudflare. His world is code. His Claude Code is a pair programmer.
I'm not an engineer. I'm a builder who ships products, distills MBA courses, and runs a content operation — alone. For me, code is maybe 30% of the job. The rest is product judgment, content operations, knowledge management, and the unglamorous work of keeping a one-person company running without it collapsing under its own ambition.
So Claude Code is not my pair programmer. It's my back office. My general staff. The thing that makes a solo operator look like a small company — not by pretending, but by actually doing the work that a small company does.
Here's what that looks like in practice.
The Three AI Army
Most people use one AI. I use three. Each has a name, a role, and clear boundaries.
Max is a Zeabur-hosted bot running 24/7. The X grind — automated reply engagement across target accounts, three original posts per day on a 30-day content pipeline, daily performance reports. Max never sleeps, never gets tired, never forgets to post at 10pm Beijing time. He's the soldier.
Tony is a Discord bot. The work that needs a human trigger but not human execution — formatting and publishing X Threads, drafting X Articles, managing the Wave reply queue for high-value engagement opportunities. Tony is the workshop.
Claude Code is the local CLI. The brain. It reads my Obsidian knowledge base — 2,000+ notes, product philosophies, MBA course materials, personal frameworks — and turns raw knowledge into deployable ammunition. It extracts, translates, packages, and quality-checks. The armory.
And then there's me. I decide what to kill and what to ship.
This is not one AI doing everything. This is three AIs with non-overlapping responsibilities, each tuned to a different cadence. Max runs on a clock. Tony runs on a trigger. Claude Code runs on deep sessions. They never step on each other because they never share a lane.
Here's how the handoff works. Claude Code extracts a product philosophy note from my vault, translates it to English, compresses it into ten standalone quotes, and drops them into an ammo-staging folder. I review, approve seven, reject three. The approved quotes get injected into Max's 30-day content pipeline CSV — and for the next month, Max publishes them on schedule without me touching anything. The reply ammunition goes into Max's soul-snippets file, shaping how he engages under other people's posts. The Thread draft goes to Tony for formatting and publishing.
One deep session of Claude Code feeds the other two bots for weeks. That's the real leverage of this architecture — not any single bot, but the compounding effect when they're wired together.
30 Days, 3 Apps
In January 2026, I shipped three iOS apps in 30 days. One Page — journaling with invisible AI reflections. FileFlow — rule-based file organization that removes human decision from the loop. MARGIN — a schedule inbox that turns natural language into calendar events.
None of this was vibe coding. Every app passed my 15-rule product kill list before a single line of code existed. Most ideas die at rule one: "Can I replay a specific moment where I personally hit this problem?" If I can't see the cafe, the laptop, the half-finished coffee — I don't build it.
I write a complete product specification first. 3.6 pages of documentation per page of code. The spec is the product. Claude Code doesn't get "build me an app." It gets "read this spec, propose architecture, explain your choices, wait for approval."
Then we iterate on architecture for hours. Zero code. Pure decision-making. Data model. View hierarchy. State management. What the app explicitly does not do. Only after architectural clarity does implementation begin — and at that point, Claude Code handles roughly 80% of the grunt work. I handle the 20% that requires product judgment: "This interaction feels wrong." "This feature adds complexity we don't need." "Cut it."
MARGIN went from idea to App Store in four days. Day zero: spec. Day one: three hours of architecture debate with Claude Code, zero lines of code, complete structural clarity. Days two and three: implementation. Day four: edge cases, App Store screenshots, privacy policy, review notes. Zero third-party dependencies. 3,000 lines of Swift, backed by 11,000 words of documentation. Submitted at 11pm. Approved next morning.
One Page was faster — one day from concept to App Store submission. AI-generated reflections that feel like your own thoughts, slightly illuminated. If you didn't look under the hood, you'd never know AI was there. That's rule six: AI should disappear, not perform.
People ask how. The answer is not speed — it's a reusable iOS scaffold. Same architecture patterns, same zero-dependency philosophy, same documentation structure, deployed across all three apps. Claude Code doesn't reinvent the wheel each time. It builds from the same blueprint, and the blueprint is mine.
The scaffold is the compound interest of doing this more than once.
The Knowledge Distillery
This is the part no engineer talks about, because engineers rarely need it. Builders do.
I'm doing an MBA at Peking University's Guanghua School of Management. 53 courses. Hundreds of hours of lectures, transcripts, PDFs, discussion notes. The kind of material that sits in a Notion database and slowly dies.
I refused to let it die.
The setup: a 6-agent team running in parallel inside Claude Code. One data fetcher pulls raw materials from Notion via API. Three distillers work simultaneously, each processing different courses — extracting Concepts (reusable mental models), Skills (actionable frameworks), and Session Notes (distilled lecture summaries). One auditor checks source accuracy and flags fabrications. I coordinate.
The output: 173 Concepts, 52 Skills, 106 distilled session notes, 67 skeleton sessions, 44 Maps of Content. From 53 courses. Two weeks.
Each Concept is a self-contained note in my Obsidian vault — linked to its source, cross-referenced against a deduplication registry, tagged for retrieval. Weber's Three Types of Authority. The Advocacy vs. Inquiry model. Gross margin horizons. These aren't class notes. They're reusable decision tools, queryable on demand.
This is Claude Code doing something that has nothing to do with writing code. Reading. Extracting. Structuring. Deduplicating. Linking. The kind of knowledge work that would take a research assistant months, compressed into days.
Was it clean? No. The Notion API threw 400 errors on unsupported block types. One course's data fetcher missed six out of seven subpages because it didn't recursively expand column lists. Session three of Management Economics was entirely fabricated — the model hallucinated detailed information economics notes from an empty Notion page. One agent's context window exploded on a 140,000-character transcript. Parallel agents created concept numbering collisions because they each read the registry's max value at the same instant.
Every one of these bugs was caught, documented, and fixed. The audit reports live in my vault. The lessons are baked into the pipeline's constitution — a governance document the agents must follow: "Distillers shall never fill gaps from general knowledge. If source data is insufficient, mark as skeleton. Fabrication is a fatal violation."
The pipeline is imperfect. But it exists. I can sit in a meeting, hear someone mention "performance-based legitimacy," and within seconds pull up a structured note that traces the concept from Max Weber through Chinese political theory to modern corporate governance — with citations pointing back to the original lecture transcript.
That's not a note-taking system. That's a second brain with better recall than the first one.
The Content Pipeline
Knowledge that stays in your vault is inventory. Knowledge that reaches people is leverage. The difference is a pipeline.
Every piece of content passes what I call the IP Consistency Guard — three questions. Does it teach, or does it share? (I share. I don't teach.) Does it sound like something I'd actually say? (Not generic, not corporate, not motivational poster.) Would I still want this online in six months?
Most content fails. That's by design. The guard is a kill mechanism, not a polish mechanism.
There's a hidden dimension here: the language barrier is the moat. My thinking happens in Chinese. My vault is in Chinese. The MBA materials, the product philosophies — all Chinese. But X is an English platform. The cross-language translation — not just words, but tone, cultural context, how a concept lands differently in Mandarin versus English — used to require a bilingual editor. Claude Code does it in the extraction step. Not perfectly, but well enough that my review pass is about tone adjustment, not translation correction.
The result: one person running what looks like a content operation. Daily posts, weekly threads, monthly articles, all in English, all sourced from a Chinese knowledge base. No ghostwriter. No content team. Just a vault, a CLI, and three bots.
What Boris Got Right, and What I Do Differently
Boris's Research-Plan-Implement cycle is excellent. His annotation pattern on plan.md is a genuinely clever workflow for engineering tasks. If you write code for a living, his approach is the one to study.
But building is not engineering.
Engineering is: given a well-defined problem, produce correct and maintainable code. Building is: decide what problem to solve, validate that the problem matters, define the boundaries, ship a product, extract knowledge from everything you learn, and turn that knowledge into content that attracts the next opportunity.
Boris uses Claude Code in one lane. I use it in four:
Product development — architecture, code generation, documentation, App Store prep. This is Boris's lane, and he's right about it.
Knowledge management — distilling 53 courses into reusable concepts, maintaining a 2,000-note vault, running multi-agent extraction pipelines. A lane engineers rarely need.
Content operations — extracting publishable material, translating across languages, formatting for platforms, quality-checking against brand voice. A lane that usually requires a team.
Strategic analysis — reading engagement data, generating hypotheses, proposing content strategy adjustments. A lane that usually requires a consultant.
Four lanes. One CLI. One human making the calls.
The Honest Part
Here's what doesn't work.
Agent context windows blow up on transcripts longer than 100,000 characters. You split sessions across multiple agents, and the coordination overhead eats the time you saved.
Bots drift. Max's reply selector occasionally fires on irrelevant tweets. Tony's Thread formatting needs manual cleanup more often than I'd like. The 30-day pipeline requires babysitting that I promised myself would be automated by now.
Multi-agent coordination is harder than it sounds. Six agents in parallel sounds impressive until two of them create concepts with the same ID number. Or one fabricates content because its source data was empty and it was too eager to produce output.
And I still review everything. Every article, every thread, every batch of reply ammunition. The AI produces drafts. I produce decisions. That ratio hasn't changed, and I don't think it should.
The Closing Argument
For engineers, Claude Code is a pair programmer. Boris captures that perfectly.
For builders, it's the back office of a one-person company. Research department, content team, knowledge manager, implementation engine — all from a terminal.
Three iOS apps in 30 days. 53 MBA courses distilled into 173 reusable decision tools. A content pipeline producing English articles from Chinese notes. Three AI bots with non-overlapping responsibilities. From the outside, it looks like a small company.
It's one person, one vault, and one CLI that turns thinking into systems.
The tool is the same. The scope of ambition is the variable.
I don't know if this scales. I don't know if it's replicable. I know it works — right now, at this stage, with this particular combination of an Obsidian vault, a CLI, three bots, and the stubborn conviction that one person with the right systems can do unreasonable things.
The apps are on the App Store. The concepts are in the vault. The content is publishing. And I haven't hired anyone.
I'm @UncleJAI. I ship iOS apps, build with AI agents, and document what I learn. If you're building alone, you might find something useful here.
