Building a Daily Automated Investment Research Workbench with the 15 Skills of WorkBuddy Xingyao Shuzhi

Abstract

Last time we talked about whether a single study can be trusted. This piece takes a different angle: put these 15 Skills together, and you can build a workbench that runs automatically every day—scanning announcements in the morning, scanning patterns during trading hours, and running backtests on weekends. At the end of the article, I'll say a word about how to get this batch of skills. 1. First, let's be clear: AI can't watch the market for you, but it can finish the work for you

Watching the market is something AI really can't do for you. It can't see the orders being placed and canceled in the order book, can't read the mood of the funds, and doesn't know which way the market will lean after a piece of news comes out. But beyond watching the market, there's still a mountain of work waiting: announcements for 300 watchlist stocks need to be gone through—who issued an earnings forecast yesterday, who got reduced, who has a lockup expiring; whether the stocks in hand have formed a pattern; whether last week's set of signals actually works. The common thread of this work is that it's large in volume, repetitive, and has fixed standards—exactly what AI is good at. So don't expect it to be a trader; letting it be the research assistant that never gets tired is more practical. 2. Four types of work, taken on separately by 15 skills

Break down daily investment research, and it's nothing more than four types of work. The first type is watching. Watchlist management stores the list locally, groups it, and lets you change and add at any time. It generates a display page with quotes and fundamentals on its own, so you don't have to manually refresh it every morning. The second type is reading. Three skills divide the work here. Announcement search filters announcements by category, downloads the original PDFs and converts them to Markdown, and supports stocks, ETFs, and convertible bonds. Smart news reading aggregates multi-source reports, deduplicates and merges them, categorizes by topic, and also produces objective statistics: which day had the densest coverage, the distribution of media sources, which topics are in the top ranks for heat, and how sentiment is moving period over period. Research report summaries pull out institutional ratings and target prices for statistics, and each entry can also be saved separately as Markdown for close reading. None of the three makes directional judgments; they only transcribe and count. The third type is finding. The pattern hunter has 34 built-in candlestick patterns, divided into four families—trend continuation, bottom reversal, top reversal, and consolidation breakout—and can run pattern radar scans on CSI 300/500/1000 constituent stocks and even all A-shares. It gives two layers of scoring: one for rule hits and one for shape similarity; all 13 candlestick periods are supported, and thresholds scale automatically by period. There's also an offbeat but useful feature: you hand-draw a shape, or screenshot a candlestick chart and throw it in, and it matches against the entire pool by contour point sequence. The fourth type is verifying. To verify an idea, first use 56 technical indicators or 90 fundamental indicators to calculate an indicator panel, then hand it to multi-indicator backtesting. Backtesting generates signals according to three rule types—position-based, crossover-based, and threshold-based; multiple signals can vote or be unioned; when matching trades, commissions, stamp duty, transfer fees, and slippage are all deducted; and finally you get two net value curves, one with costs and one without. The net value is passed to performance analysis, which produces five categories and more than twenty indicators, benchmarked against the CSI 300 by default, with net value curves, maximum drawdown, and monthly return heatmaps. If you think it's reliable, then throw it into paper trading for a while. All the data is supplied by the financial data skill, 13 major categories, with stocks traceable back to 2013. If you need to calculate factors, there's also factor analysis—the four major single-factor tests and multi-factor synthesis can all be done. Let me insert a word here about what WorkBuddy does in the middle. It doesn't let the model write formulas on the fly; instead, it calls these skills as fixed implementations: data retrieval goes through Galaxy's data sources, indicators go through the algorithms written in the skills, trade matching goes through the backtesting logic in the skills, and the model is only responsible for understanding what you want and organizing the results for you. This is also why the numbers it calculates are safe to use—the numbers given by general large models are pieced together. 3. Put together, it's three segments a day

Using each skill on its own is only a productivity boost; stringing them together is where it gets interesting. Half an hour before the open. Open the watchlist, have it pull quotes and fundamentals for the whole list, then scan announcements, news, and research reports one by one. Focus on whether there were earnings forecasts, reductions, lockup expirations, or rating changes from last night to this morning. This step used to require opening three websites and flipping through them for half an hour; now it's a one-sentence job. After the close. Run the pattern radar to see whether anything in the pool you follow has just broken out of a box or formed a head-and-shoulders bottom. After that, run a single-stock pattern diagnosis on your holdings to confirm nothing has broken down. On the weekend. This is the most valuable segment. Write the ideas you've accumulated over the week into signals, run backtests, look at the dual net value curves, look at Sharpe, look at maximum drawdown, and then decide whether to put it into paper trading. After running these three segments, the time you spend shifts from flipping through materials to reading conclusions. And every run uses the same standards, so when you look back three months later, the data is comparable and won't fail to line up because you changed formulas at the time. 4. What the running output looks like

The output isn't a pretty speech; it's something you can use next. Indicators and factors are calculated as structured data—date, code, category, and the value of each indicator are all there, and can go directly into your spreadsheet or feed the next step. Pattern scanning, performance analysis, backtesting, paper trading, and asset allocation produce HTML reports, with net value curves, drawdown intervals, monthly heatmaps, and trade records all inside; they can be archived directly or sent to others. For announcements, news, and research reports, the original files land locally and are converted to Markdown before being handed to the model to read; both the reading process and the original text are preserved. This point is quite key. It gives you raw materials and semi-finished products, not conclusions already made for you. 5. A few pitfalls that are easy to step into—let's get them out front

Adjusted prices. Before calculating technical indicators, you must always use forward-adjusted prices; otherwise, once the ex-rights date passes, MACD and moving averages will all be broken. This step is built into the technical analysis skill—don't forget it when building your own factors. Look-ahead bias. Q1 reports are only disclosed at the end of April; if you use them to calculate valuations on March 31, that backtest is fake. The fundamental analysis skill uses forward filling based on the actual announcement date, and the options strategy only uses information from T-1 and earlier when selecting contracts—both are for this reason. Total share units. In AmazingData, total shares are recorded in units of ten thousand shares; to calculate market value, you need to multiply by ten thousand. With this one order of magnitude off, the price-to-sales ratio can go wrong in baffling ways. Don't install all the skills at once. With all 15 installed, skill matching in WorkBuddy conversations tends to get messy. Install whichever you need, and add more once things run smoothly. 6. Boundaries—still worth a word

These 15 skills provide data, calculations, and validation; they do not provide judgments, nor do they touch real money. Indicator-type skills only output values. Pattern Hunter will say that this stock has formed a head-and-shoulders bottom with a similarity of 0.82; it will not say you should buy. Strategy backtesting, options strategies, asset allocation, and paper trading will produce net values, signals, and weights, but all of it stops at the simulation layer; paper trading does not connect to live accounts and does not generate real orders. The entire set of capabilities does not provide automated trading, does not provide live programmatic trading, and does not output directly executable buy or sell instructions. In other words, the workbench is built, but the one sitting in it making decisions is still you. And that's right—judgment is, after all, the most valuable part of this business. 7. If you want to build your own setup, the skills are free

Normally, to use this batch of skills, you first have to open an account with a brokerage, meet asset and risk assessment requirements, and then separately apply for data permissions—the whole process takes several days. The 2026 Xingyao Shuzhi AI Simulation Competition starts on November 2, and registration is already open. During the competition, these 15 skills can be used in WorkBuddy without thresholds; just install the skills and configure your account and password, and you can run them. Real data, real tools, simulated trading—no real money involved, ranked by simulated returns during the competition period. To get the free skills and a competition account, go to the Galaxy Securities Xingyao Shuzhi official account; the registration entry and account activation instructions are both there. You can treat it as a public live-fire exercise in AI investment research, or as a reason to force yourself to build the workbench. There are only four days left before the competition starts—enough time to set up your environment and run two stocks through the full workflow. Risk disclosure: The tools described in this article only provide tool capabilities such as financial data queries, indicator calculations, strategy generation, backtest analysis, performance analysis, and paper trading. They do not provide automated trading, do not provide live programmatic trading, do not output buy or sell instructions, do not predict securities price movements, and do not constitute any investment advice or return commitment. The competition rules, schedule, and awards are subject to the information published on the Galaxy Securities Xingyao Shuzhi official account. Markets carry risk; invest with caution.