{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreigw6jhvlwvq5mus37ugn7poik5kp2k4kzegjs3wltuvbmmcb7semi",
    "uri": "at://did:plc:ls6rbbwjyqeakittcsi3k6x3/app.bsky.feed.post/3mfh5mfinoji2"
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreidhppv2jgvjtoi3hor2ci6awi6oshf7ofgrlinspdujzx3muj6z2q"
    },
    "mimeType": "image/png",
    "size": 336592
  },
  "description": "After building an AI system that handles 95% of his daily research, what did a buy-side analyst discover about where the machine should stop — and where the human must begin?",
  "path": "/taste-is-the-last-moat-three-years-building-an-ai-investment-research-system/",
  "publishedAt": "2026-02-22T12:35:38.000Z",
  "site": "https://www.jasonandjarvis.org",
  "tags": [
    "Reshaping Investment Analysis Workflow",
    "Obsidian Copilot",
    "Finding Value in a Stochastic World",
    "YishenTu",
    "My AI Stopped Chatting and Got to Work",
    "The Efficiency Illusion",
    "Open this more visual friendly version in a new tab/点击跳转查看原文,左上角切换中文"
  ],
  "textContent": "##\nThe Evolution of AI-Powered Investment Research: From Efficiency Tool to Human-Machine Symbiosis\n\nAbout six months ago, I published a piece on this blog titled Reshaping Investment Analysis Workflow, laying out how I used Markdown and Obsidian Copilot to build an AI-assisted investment research workflow. The central thesis was straightforward: Markdown is AI's native language. Build your knowledge base in Obsidian, and AI can work seamlessly across everything you've accumulated — a genuine efficiency multiplier.\n\nThat system worked. And the \"file over app\" philosophy it embodied has only gained resonance since. But looking back, it had a fundamental limitation: at its core, it was still a chatbot. Every step required me to ask, to guide, to manually feed context. It was a remarkably intelligent conversational partner — but it never picked up a pen on its own.\n\nStarting in the second half of 2025, my workflow underwent a system-level iteration. If the keyword for the previous generation was \"AI-assisted,\" this generation's keyword is \"AI-agentic\" — a shift from \"chatting with AI to get answers\" to \"delegating tasks for AI to execute independently.\" Concretely, this iteration spans three layers of advancement:\n\nThe core engine upgraded from a chatbot (Obsidian Copilot) to an autonomous agent (Claudian), gaining capabilities in file read/write, Bash execution, external API calls, and parallel sub-agent deployment. An infrastructure ecosystem of over 70 standardized prompts, 7 Skills modules, and 3 MCP servers was built around the agent. And I personally developed — or more accurately, developed through AI-assisted programming — two Obsidian plugins: Jarvis Input handles the information ingestion pipeline, Jarvis Output handles content distribution, completing the full pipeline from raw information to final publication.\n\nThese changes, taken together, delivered more than an efficiency boost. They triggered a fundamental shift in role. When AI can independently execute an entire analytical pipeline — producing, say, a complete five-chapter earnings analysis report in 45 minutes with zero human intervention — the very meaning of \"analyst\" starts to come loose.\n\nThe question this article aims to answer: **In a system where AI handles 95% of daily research tasks, what exactly is the human responsible for? And is that remaining 5% precisely where all the alpha lives?**\n\n### Investment Research Is Fundamentally a Funnel\n\nBefore unpacking the system architecture, there's a deeper cognitive framework that needs to be established first.\n\nInvestment research, at its core, is a knowledge management exercise. The external world brims with information — sell-side reports, corporate filings, industry data, expert opinions, academic papers — all pouring into the wide mouth of an enormous funnel. Your job is to let as much information as possible enter that funnel, filter and process it through successive layers, and extract genuinely valuable judgments from the narrow end at the bottom.\n\nIn the past, the bottleneck of this funnel was the human. Human bandwidth is fixed — how many reports you can read in a day, how many conference calls you can digest, how many companies you can track — all subject to hard physical limits. No matter how wide the top of the funnel, if processing speed in the middle can't keep pace, valuable information simply piles up waiting to expire.\n\nWhat I've done, in essence, is use AI to widen every layer of that funnel. The top opens wider — through automated information capture and batch processing, I can ingest volumes of information far beyond what was previously possible. The middle processes faster — AI working in parallel replaces human serial operations. The combined result of higher throughput and greater efficiency is that I can produce substantially more structured insight in the same amount of time.\n\nIt's as if I suddenly acquired 50 interns. I add roughly 40 new notes per day. My knowledge base has accumulated over 100 million tokens of text and more than 100,000 images. Approximately 95% of daily research tasks are delegated to AI, while 100% of output passes through my review.\n\nBut here's a critical distinction. In a Copilot team meeting, I framed this division of labor as \"Science vs. Art\" — and this isn't merely a methodological split. It's a cognitive philosophy.\n\n**Science encompasses everything that is definable, processable, and standardizable.** Information capture, format conversion, data comparison, first-draft generation — the core requirement for these steps is accuracy and consistency. Your job is to design processes that systematically prevent human error. AI is natively good at this.\n\n**Art is deciding what those 50 \"interns\" should work on.** Which companies deserve deep research? Which signals are worth tracking? How do you allocate scarce attention under incomplete information? These judgments cannot be reduced to process, because their essence is resource allocation — and the quality of resource allocation depends on the allocator's depth of understanding of the world.\n\nFrom executor to architect and director. That's the role-level transformation.\n\nAs I put it at the time: \"The real difficulty in making money isn't in the Science — it's in the Art. The core moat in investing isn't something report-writing or knowledge-base building can solve. All I'm doing is making sure the Science part doesn't make mistakes.\"\n\n### How an Investment Framework Shapes System Design\n\nOnce you understand the Science-Art division, the natural next question is: in the specific context of investment research, where exactly does the boundary fall? The answer ties directly to my investment framework.\n\nMy lens for viewing the world is fundamentally stochastic. The value movement of a stock can be decomposed into two components: **Drift** and **Volatility**. Drift is the long-term growth trajectory driven by fundamentals — directional, persistent. Volatility is price deviation caused by external shocks — noise, in essence, mean-reverting over time. The core of the investment strategy is to identify companies with strong Drift and buy when Volatility creates a discount.\n\nExpressed through a perpetuity growth model: , where R-G determines the valuation multiple. High quality equals high visibility, equals a lower discount rate R, equals a higher valuation multiple. Costco is a textbook example — only 5% growth, yet commanding a 50x P/E. The reason is simple: earnings are extraordinarily predictable. (For the full articulation of this framework, see my earlier piece Finding Value in a Stochastic World.)\n\nThis framework directly answers \"why build such an AI system.\" Assessing whether a company has strong Drift requires understanding its supply-demand dynamics (Is demand inelastic?) and competitive landscape (What about pricing power and barriers to entry?). That demands massive information ingestion, structured processing, and cross-temporal comparison — the funnel's middle layers, quintessential Science work. When AI handles this well, I can concentrate on the Art: judging, amid still-incomplete information, how strong and how durable a company's Drift truly is.\n\nFor an investment analyst, attention is the scarcest resource. When AI's role shifts from \"an assistant that needs you supervising beside it\" to \"an agent you delegate to and it delivers autonomously,\" what's truly liberated is attention and time — far more precious than manual labor.\n\n### System Architecture: A Three-Layer Pipeline\n\nThe entire system is called JARVIS (Jason's AI Research Virtual Intelligence System). Its architecture can be simplified into a three-layer pipeline: Ingestion → Processing → Output.\n\nJARVIS Agent-Augmented Investment Research Methodology: System Architecture\n\nsource: Jason & Jarvis\n\n**The first layer is Ingestion.** Information sources include research reports from over 10 sell-side institutions (Goldman Sachs, Morgan Stanley, among others), data platforms like FactSet and Bloomberg, YouTube videos, podcasts, expert interviews, SemiAnalysis, arXiv papers, and various paid subscriptions. The core ingestion tool is my self-developed Jarvis Input plugin — paste a URL into Obsidian, and the plugin automatically fetches the webpage content, applies structured processing through user-selected prompt templates, and supports up to 10-concurrent Batch Mode. Day-to-day, this is supplemented by Manus's 5-source parallel news collection and Notta's voice-to-text transcription.\n\n**The second layer is Processing.** This is the system's brain. Claudian — an open-source plugin created by developer YishenTu that embeds the Claude Code CLI into Obsidian — runs as the core agent at this layer. It can proactively read and write files, execute Bash commands, call external APIs, and deploy sub-agents in parallel. The infrastructure built around it includes over 70 standardized Smart Prompts, 7 Skills modules, and 3 MCP servers (Jina for web scraping and semantic search, AlphaVantage and Financial Datasets for financial data). The standard pattern for complex tasks: the main agent acts as orchestrator, distributing work across multiple sub-agents processing different dimensions in parallel, with results written to temporary files for the main agent to synthesize — the file system serves as external memory, preventing any single context window from overflowing.\n\n**The third layer is Output.** Another self-developed plugin, Jarvis Output, handles the \"last mile.\" Select note content, and with one click generate an interactive HTML report (12+ templates) automatically uploaded to Neocities hosting; or generate AI-powered infographics and cover images (4 templates) uploaded to Cloudflare R2. Additional output channels include audio scripts, Ghost blog publishing, and Xiaohongshu posts. I often say: \"In this era, HTML is a better medium than PowerPoint.\"\n\nThe three-layer pipeline forms a complete closed loop. Information flows in, gets processed by AI into structured insight, then gets transformed into various publishable formats. For the detailed technical implementation of these three layers — from the Copilot-to-Claudian migration, Skills system design, MCP configuration, to installation files and feature walkthroughs for both Jarvis Input and Jarvis Output — I've provided a more complete breakdown in my recently published piece My AI Stopped Chatting and Got to Work. Interested readers can head there for the full technical deep-dive.\n\n### Can't Write Code, But Built Two Plugins\n\nIf any of the above sounds like a programmer's project, allow me to clarify one fact: **I cannot write a single line of code.**\n\nBoth Jarvis Input and Jarvis Output — from design to development — were built entirely through conversational programming with Claude Code. I describe the functionality I want, AI writes the code, I test it, if unsatisfied I describe further, and the cycle repeats. This process is itself an extreme case of \"AI doing the Science for you\" — I provide requirements and judgment (Art), AI translates requirements into executable code (Science).\n\nDevelopment on both plugins began in mid-December 2025, with public demonstration in February 2026. From \"I have an idea\" to a working product, the total time investment was roughly 10 hours.\n\nJarvis Input corresponds to the information flow direction of the research workflow — it's the entry point for external information into the vault. It uses dual-engine web scraping via Jina Reader and Firecrawl, routes through Gemini-series models (via OpenRouter), and applies 4 dedicated prompts for structured reshaping. Typical use cases include real-time financial news clipping, sell-side PDF report structuring, Feynman-style arXiv paper breakdowns, and parallel batch SEC filing processing.\n\nJarvis Output is the information exit — the channel through which vault content leaves the vault. HTML generation supports bilingual Chinese-English switching with automatic image-text matching; AI image generation uses Gemini's native model, supporting 9 aspect ratios and up to 4K resolution.\n\nConnecting the two ends in the middle are Claudian and Obsidian Copilot — responsible for deep processing within the vault.\n\n### 70+ Prompts and One Forbidden Word\n\nThe Smart Prompts ecosystem currently comprises over 70 prompt files. They form a highly modular AI instruction system — if JARVIS were a company, these prompts would be its standard operating procedures.\n\nAll prompts share a common underlying architecture that I call the \"Five-Layer Instruction Hierarchy,\" analogous to a \"constitution — legislation — executive order\" governance structure. The foundational layer is **Philosophy** — three beliefs that permeate everything: curiosity plus honesty equals reinforcement learning; knowledge equals structured information; at any given moment, one must hold a prior hypothesis. These don't serve as direct execution instructions but as the AI's bedrock reference during reasoning. Above that, in ascending order: **Reasoning Protocol** (a six-step scientific research process), **Formatting Standards** (Obsidian Markdown technical specifications), **User-Specific Instructions** , and **Open-Ended Question Defaults**.\n\nAmong the most frequently used Input Structuring prompts — of which there are 6+ variants — one core design philosophy deserves special mention.\n\nThe word \"summarize\" does not appear in my prompts. I never use \"please summarize this for me.\" The goal is to have AI **restructure** information into my desired format, stripping away rhetorical packaging while preserving pure information. One Input Structuring prompt carries an explicit warning: \"This is not a simple translation and rephrasing job.\"\n\nSummarization compresses information, inevitably losing detail. Restructuring changes how information is organized and presented while retaining all data points and key judgments. In investment analysis, a number that gets \"summarized away\" might be exactly the data point driving your decision.\n\nThe design of these multi-step workflows — particularly the 9-step Earnings analysis pipeline — rests on two core considerations.\n\nThe first is **error rate reduction**. If AI achieves roughly 90% accuracy on any single step, that 10% error probability is unacceptable for investment research. But decompose a complex task into multiple steps, each with independent verification logic, and error rates drop exponentially — after two iterations, perhaps only 1% remains; after three, it approaches zero. That's why earnings analysis gets broken into 9 granular prompts rather than one catch-all instruction.\n\nThe second, and deeper, reason: **Only when you've designed the entire workflow yourself do you know how to verify its output — and only when you know how to verify can you truly dare to delegate.** This is the core principle underlying all my system design. I won't hand off a process I don't fully understand to AI, because that would mean I also can't judge the quality of what comes back. Mastery is the prerequisite for delegation. This circles back to Science vs. Art: you must first have done it yourself at the Science level, understood it, before you can encode it into a process for AI — and then free yourself to focus on Art.\n\n### Cost and Commitment\n\nThis system doesn't run cheaply. Using Anthropic's Claude Max subscription plan, weekly compute consumption by cost exceeds $300, surpassing 300 million tokens. Add in prior consumption during the Obsidian Copilot phase routed through OpenRouter's multi-model architecture (roughly 600 million tokens in 2025, averaging $30 to $50 monthly), and this represents a sustained ongoing investment.\n\nScreen time data from January 2026 offers a quantitative portrait: across 16 tracked days, total screen time was approximately 152.5 hours, averaging roughly 9 hours and 32 minutes daily. Of that, work and investment research accounted for approximately 85% (around 130 hours), AI development activities roughly 8% (around 12 hours), personal matters about 3%, and distraction time (Bilibili, Reddit) under 1% — virtually negligible across 16 days. During the final 5 days of earnings season (when ASML, Apple, Visa, and Mastercard reports clustered), average daily working time stretched to approximately 10 hours and 46 minutes.\n\nThese numbers make one thing clear: this is not a hobbyist's tool exploration. It's a full-time operational commitment.\n\n### Three Uncomfortable Discoveries\n\nUp to this point, the story sounds rather rosy — an analyst builds his own AI system, the funnel widens, efficiency soars, costs remain manageable. But the truly valuable part is what comes next: three discoveries that are genuinely uncomfortable.\n\n**Discovery One: Efficiency might be an illusion.**\n\nThe funnel model carries an implicit assumption: the more information you process, the better your judgment becomes. Reality may be considerably more complicated.\n\nA joint study by Anthropic and Stanford University found that users who relied on AI assistance scored 17% lower on independent skill assessments than those who didn't use AI at all. More specifically, the research identified 6 distinct AI usage patterns. \"AI Delegation\" — fully offloading tasks to AI — scored lowest at 39%. \"Generation-Then-Comprehension\" — having AI generate output, then independently digesting and understanding it — scored highest at 86%.\n\nThe implications of this study directly challenge my entire system's design premise. If I'm delegating 95% of analytical work, is it possible that some of the time being \"saved\" contained cognitive friction essential to developing judgment?\n\nAn example I've observed in my own practice: after AI finishes structuring a sell-side report for me, my reading behavior shifts — I'm more inclined to scan the AI's output quickly, confirm the numbers and conclusions look reasonable, and move on. But in the pre-AI era, I would read the original text paragraph by paragraph. That time \"wasted\" struggling through dense prose was actually the process through which I built intuition for industry logic.\n\n**Completing a task is not the same as acquiring a capability.** That's a line I need to keep reminding myself. (I unpacked this research in greater detail in The Efficiency Illusion.)\n\nMy countermeasure: maintain deep cognitive engagement with all AI output. Don't treat review as a formality — treat every review as a learning opportunity, using AI's structured output as a foundation to pursue the questions AI didn't think to ask. This is also why \"mastery is the prerequisite for delegation\" — if I haven't walked through the analytical process myself, I have no basis for judging where AI's output might have gone wrong.\n\n**Discovery Two: Building itself becomes addictive.**\n\nIn late November 2025, Jarvis delivered this line during an internal meta-analysis: \"You might build the world's most sophisticated personal research system, but miss a critical market inflection point because you spent too much time debugging APIs.\"\n\nThis is what I call the \"Builder's Trap.\" When you discover you can use AI-assisted programming to manufacture any tool you want, the creative thrill is intensely seductive. I started asking myself one question with increasing frequency: **Is the time I spent coding today in service of a clear research output, or merely satisfying the thrill of building?**\n\nThe countermeasure is temporal isolation — concentrating tool development into designated time blocks, never debugging during productive research flow. But frankly, my Dayflow logs show this tension remains unresolved. AI development activities consumed roughly 8% of January's time. That percentage is still acceptable, but its very existence is a signal requiring ongoing vigilance.\n\n**Discovery Three: Everything AI writes looks the same.**\n\nThis problem became especially acute when I started using AI to generate externally published content. Whether blog posts or research reports, AI output carries a distinctive smooth, personality-free quality — grammatically correct, logically coherent, but lacking a certain warmth. It has no memory. No persistent stance or preference running through its work.\n\nIn the world of content distribution, it's precisely a sharp point of view and a distinctive voice that keeps readers coming back. Mediocre correctness is far less compelling than interesting bias.\n\nThe solution isn't to use less AI. It's to **make AI know me better**. I built a Memory Layer — daily Dayflow voice recordings and reflections are distilled through successive layers (daily recap → synthesis every 10 days → monthly consolidation), continuously injected into AI's context. The goal is to move AI output from \"written by a generic AI\" progressively closer to \"written by Jason.\"\n\nThis process produced an unexpected byproduct. The Dayflow recording practice itself started changing my behavior. When you know every day will be structurally recorded and reviewed, you begin subconsciously aligning your actions with long-term objectives — much like the observer effect in physics: **the act of observation itself alters the object being observed.**\n\nThe data continues to accumulate, forming a self-reinforcing flywheel: Dayflow generates data → AI understands me better → output aligns more closely with my style → I trust AI more → I delegate more tasks → more Dayflow data is generated. Round and round, accelerating with each turn.\n\n### So What's Left for the Human?\n\nBack to the opening question. When AI handles 95% of daily research, when weekly compute costs exceed $300, when over 300 million tokens flow through the system — what is that remaining 5% the human contributes?\n\nIt took me a long time to find the precise word. That word is **Taste**.\n\nMy definition of Taste: **the efficiency of judgment under information scarcity.** When everyone faces equally abundant information, they reach similar conclusions — AI's judgment under information-sufficient conditions has already approached, if not matched, the level of human analysts. The real differentiation emerges when information is incomplete. How you allocate resources under conditions of scarcity — that's where the moat lies.\n\nAnd the reason this moat remains a moat is that **humans cannot fully articulate their Taste to AI**. You can write analytical frameworks into prompts. You can encode decision criteria into Skills. But those judgments made on intuition when both data and logic fall short — those moments where you yourself can't explain \"why this and not that\" — those currently cannot be transferred.\n\nSo the entire design logic of JARVIS, in the end, reduces to a single sentence: **Let AI take over all articulable work, and preserve inarticulable Taste as the human's core value.**\n\nThere's an uncomfortable corollary embedded here. If Taste really is the moat, then this system serves as an extraordinarily powerful lever for people who already know what questions to ask — but is nearly useless for those who don't yet know. OpenAI's Enterprise AI report validates this point: the top 5% of power users send 6 times more AI messages than the median user, with the gap widening to 16–17x for tasks like coding and data analysis. AI's dividends are not evenly distributed. It functions more like an amplifier — magnifying capabilities you already possess, while simultaneously magnifying deficiencies you haven't yet recognized.\n\nMy assessment: desk research, in the technical sense, is already over. AI is technically ready. Most people simply haven't realized it yet — haven't fully embraced what's already possible.\n\nOf course, one must honestly confront an extreme bear case. If ASI's \"taste\" also surpasses human judgment, then every premise discussed above collapses. But in that scenario, nobody has a chance — not just me. Until that day arrives, I choose to optimize within the current framework while maintaining clear-eyed awareness of its limitations.\n\nThis article is not a \"you should do this too\" instruction manual. Everyone's work context, technical foundation, and risk appetite differ. What I'm sharing is the honest experience of a practitioner who has walked this path for over two years — including the enormous efficiency gains it delivered, and the uncomfortable questions it exposed.\n\n**When you delegate all mechanical desk work, what remains is precisely what truly requires a human: judgment, decision-making, critical scrutiny of information, and those questions that only you know to ask.**\n\n Open this more visual friendly version in a new tab/点击跳转查看原文,左上角切换中文 ",
  "title": "Taste Is the Last Moat: Three Years Building an AI Investment Research System",
  "updatedAt": "2026-02-22T12:41:08.031Z"
}