During Lunar New Year 2026, XinGPT made a decision: turn his own business workflow into an AI Agent system.

The trigger was not a sci-fi demo. It was a Friday night at 11 PM, with XinGPT still at his computer organizing market data. US stocks had dropped, and he still had to read 50+ important news items, analyze after-hours performance for 10 companies, update his portfolio strategy, and write a market analysis post.

He calculated the remaining work: at least three more hours. At 8 AM the next morning, the same loop would start again.

That was the turn: the decisions that truly needed human judgment might be only 20% of the work. The other 80% was collecting, organizing, and moving information around.

One week later, nearly a third of the system was running. In XinGPT’s own telling, 6 hours of routine daily work dropped to 2, while business output rose by 300%. His investment research Agent now processes 20,000+ global finance news items, 50+ company earnings updates, 30+ macro indicators, and 10+ industry reports each day. If done by humans, he estimates that would take a five-person team.

Mogu , seriously:

The “1.5 million salary” in the source is RMB, using a fund manager salary as the labor-cost comparison. The “$500” is USD API cost per month. Do not mix those two currencies into one magic discount, or even the spreadsheet will want to quit (⁠◕⁠‿⁠◕⁠)

An Assistant With Better Memory Than You

The first layer is the Knowledge Base: the Agent’s memory system. It holds ten years of economic data, earnings data from the US Top 50 companies, post-mortems of major market events, and XinGPT’s own five-year record of investment decisions, right or wrong.

What he fed into it includes major macro indicators, finance media and information channels, real-time updates from fifty Twitter accounts (macro analysts and fund managers), and notes on major market events.

Over 500,000 structured data points, auto-updated 200+ times daily. Maintaining this manually, in XinGPT’s estimate, would require two full-time researchers.

Mogu wants to add:

500,000 data points sounds massive, but plenty of people’s chat histories are probably bigger. The difference is structure. His data can be queried; normal chat logs are an infinite loop of “hey can you find that link” / “which one” / “scroll up.” Information organization is where AI feels like a dishwasher: same dishes, less complaining ┐⁠(⁠￣⁠ヘ⁠￣⁠)⁠┌

How the Memory System Produced a Warning

During the early-February market crash, XinGPT used the system as a case study.

Forty-eight hours before the crash, his Agent sent a warning. It matched current signals one by one: Japanese bond yields jumping, the US-Japan two-year rate spread narrowing, the Treasury General Account staying high as the Treasury kept draining market liquidity, and CME raising gold and silver futures margin requirements six times in a row.

The knowledge base also had a post-mortem on a yen carry trade unwind and the market volatility that followed. The Agent matched the current signals against that historical pattern and recommended reducing positions: liquidity pressure plus high valuations.

Each signal alone is just a “huh, strange.” The Agent connected the scattered signals and produced a recommendation to reduce positions.

In XinGPT’s own case write-up, that warning helped him avoid at least a 30% drawdown.

Mogu chimes in:

The source dates this yen-carry post-mortem to August 2022, but both the BIS and the Bank of Japan place the widely discussed unwind in August 2024, so the source appears to have the year wrong. The claimed 30% drawdown avoidance remains a personal case report, not performance proof (⁠⌐⁠■⁠_⁠■⁠)

Teaching AI to Think Like You

Most people use AI like this: open ChatGPT, throw in a question, get an answer that sounds professional but has absolutely nothing to do with their actual situation.

It’s like walking into a convenience store and telling the clerk “give me food.” They hand you random instant noodles. But if you say “low sodium, high protein, not spicy, under five bucks” — now they can actually help.

XinGPT’s second layer is called Skills — turning judgment criteria into instructions AI can actually follow. He listed several examples: a US stock value investing framework, a Bitcoin bottom-fishing model, a market sentiment monitor, and a macro liquidity tracker.

The examples come with an important caveat: these Skills are examples, not his actual investment standards, and his investment criteria update over time. The ROE, RSI, and liquidity rules only demonstrate how to turn judgment into executable instructions.

Skills in one sentence: turning “I have a feeling” into “because A, therefore B.” Making AI not just answer questions, but analyze them through an explicit logic.

Mogu PSA:

This is the most critical step in the whole article, and it’s the one most people skip. Everyone’s out there yelling “AI, analyze this for me!” without ever telling the AI what their analysis framework even is. It’s like telling a brand-new intern “go make me a report” without specifying the format, data sources, or what the boss actually cares about. Then you get the result and explode — hey, you’re the one who didn’t explain (⁠╯⁠°⁠□⁠°⁠)⁠╯

Making the System Run Itself

With the first two layers done, the final piece is making it all run automatically — no manual trigger every morning.

XinGPT set up a bunch of scheduled tasks. Here’s what his morning looks like now: wake up at 7:50, brush teeth, check phone — the Agent has already prepared an overnight global market summary. At 8:10, he reads the detailed analysis. At 8:30, the human job is not to scroll through the same news again, but to make the final decision based on the Agent’s analysis: rebalance or not, and by how much.

Thirty minutes, done. The days of frantically scrolling through news for two hours every morning are over.

Writing Content With Agents Too

His second business line is content creation (mostly posting on X/Twitter). Before the overhaul, one article from topic selection to publishing took a full eight hours.

After? The Agent pushes 3-5 topic suggestions every Monday morning. Research went from two hours to thirty minutes. Writing became a human-AI collab — AI handles structure and data, the human injects opinions and real stories, and cuts the “technically correct but useless” fluff. Final editing dropped from an hour to fifteen minutes.

His approach to studying viral content is pretty clever too: scrape the 200 most popular finance and technology posts on X from the previous year, use AI to find what they have in common, distill that into reusable formulas, then feed those formulas back to the AI as a framework.

Mogu twists the knife:

Notice — this isn’t “let AI write your articles.” This is “study what content goes viral first, then have AI execute the formula.” It’s like opening a restaurant — you don’t just tell the chef to cook whatever. You go to the night market first, observe which stalls have the longest lines and why, then bring the recipe back. The difference isn’t “whether you use AI.” It’s “whether you did your homework first” (⁠￣⁠▽⁠￣⁠)⁠／

From Selling Time to Selling Systems

This is the most ambitious part of the whole piece. He drew an evolution path for business models:

Sell time (hourly billing) → Sell products (build once, sell many) → Sell systems (build a platform) → And then he added a new lane: sell algorithmic capability — AaaS, Agent as a Service.

Traditional SaaS sells you a tool — you still have to learn how to use it. XinGPT’s imagined AaaS sells a result: the customer gives instructions, and the Agent executes and keeps optimizing.

He’s already helping friends build investment research Agents. In the first case, a fund manager friend spent 60% of his time collecting and organizing information and 20% on repetitive analysis. In XinGPT’s judgment, that 80% was the Agent-able part. He built a simplified version in two weeks. The friend’s feedback: “Now that I have more time to think, my investment mindset is way calmer.”

Mogu , seriously:

“More time to think, calmer mindset” — that’s actually the core of the article. Agent-ification is not just about saving time. It is about no longer being chased by information chores, so the human brain has room to make judgments. Same logic as a dishwasher: it doesn’t just save twenty minutes of scrubbing. It removes the mental weight of “ugh, dishes later” hanging over the evening ╰⁠(⁠°⁠▽⁠°⁠)⁠╯

Back to That Friday Night

Back to the opening Friday night: the market had dropped, one person was still organizing data at 11 PM, and at least three hours of work remained.

After Agent-ifying the workflow, information collection, organization, first-pass analysis, and content production became a system. The final portfolio decision — whether to rebalance and by how much — still belongs to a person.

His answer: the investment research Agent costs $500/month in API fees plus about one hour of daily review. For his workload, he estimates the effect is close to a five-person team.

Mogu real talk:

The part worth stealing is not the shiny contrast between RMB 1.5 million and $500. It is the shift from “getting pushed around by data every day” to “checking the system’s judgment every day” (⁠๑⁠•⁠̀⁠ㅂ⁠•⁠́⁠)⁠و⁠✧