The secret for ordinary people to build their own AI investment research workflow: Make good use of the three tools—Harness, Skill, and GitHub.
```
Large models have long since evolved from novelty to tools that more and more investors, researchers, and knowledge workers use every day.
But while everyone uses Doubao, GPT, Claude, some people have already enabled AI to deeply participate in their research, analysis, and decision-making processes; others remain at “help me summarize this article.”
Does the gap come from the model itself, or from how it’s used?
On June 2, Thea Xue, author of the WeChat blog "Thea's AI Diary" and former Microsoft search algorithm engineer, was a guest in the VIP Lounge to share her thoughts on AI workflows. Replay → How else can AI investment research be played?
In her view, large models are just the starting point. What really determines the difference in efficiency is not the model itself, but whether the user can harness the model (Harness), distill their workflow (Skill), and use the open-source ecosystem (Github) to continually expand their boundaries of capability.
She summarizes this as AI empowerment’s "three axes": Harness, Skill, and Github.
The following is a summary of the livestream content.
1. Harness: Many don’t use AI well—not because the model isn’t strong enough, but because they don’t know how to harness it
Let’s first talk about Harness.
Harness originally refers to the reins used when riding a horse. Many people feel large models are very powerful—they can do almost anything—but they’re like a wild horse without reins: strong, but not necessarily going where you want. To make large models work for us, the first step is to put the reins on.
A while ago, Liang Wenfeng said something that went viral:
Model + Harness = Agent
In other words, to use a large model well, essentially you just need to solve two issues: Model and Harness.
Model is relatively simple. For most general tasks, just pick the most capable model. But there’s one special situation—when models are connected to different data sources.
For example, I once had several models search for recent AI sharing events in Beijing. Yuanbao mostly found small events suitable for individual developers and entrepreneurs, while Claude and GPT found mainly large industry conferences.
The reason is simple. Yuanbao can access content in WeChat public accounts, and their anti-crawling mechanisms are strict, so Claude and GPT have trouble directly accessing that data. So if you want to search for information in the public account ecosystem, Yuanbao is usually better.
Different models have their own specialties.
Yuanbao is best for WeChat ecosystem content; Gemini excels on YouTube, Google Scholar, etc.; Doubao can access more Douyin-related material; GPT covers Twitter, Reddit, and overseas communities well; Claude is especially good at writing code and long-form text.
When writing articles, I tend to prefer Claude. GPT is also strong, but currently is somewhat verbose; for structured text output, I personally prefer Claude.
Having talked about choosing a model, let’s discuss Harness.
Harness covers many topics. Today, I’ll mainly share three parts:
First, how to provide information to AI;
Second, how to control AI’s direction;
Third, after getting results, how to give AI feedback to make those results more reliable.
First, providing AI information.
Everyone knows large models can hallucinate. When it comes to numbers, references, or professional content, they may make up data or sources. Also, their training data has time boundaries; for latest info, web search is often needed.
Another issue: large models tend to give “average answers” by default.
To fix these, we must actively provide information to AI.
There are mainly two kinds of information here.
The first is real-world information.
Such as newest financial reports, announcement PDFs, full research reports, etc.
If it’s professional analysis, I strongly suggest you feed these objective materials to the model, instead of asking naked questions.
Often, you feel AI’s analysis is logical because the info is public and it can find answers in training material. For less public data, it often can only guess.
So, when handling professional subjects, provide objective info to the model first.
The second is your own information.
This includes your competence circle, risk preference, investment boundaries, expression habits, etc.
For example, you ask:
Should I buy CATL?
The model will probably give you a standard “optimistic for the long term, cautious in the short term” answer.
But if you let it know beforehand:
I already have 40% of my portfolio in new energy.
The advice will be totally different.
Because now it considers your investment boundary, not just an “average investor.”
Of course, not all tasks need personal info.
For things like reading financial reports or breaking down tables, personal profile isn’t crucial. But if you want AI to write research reports in your style, or give advice matching your habits, then language style, analytical frameworks, and risk preferences are very important.
Many people say:
“I’m not sure what my investment style is.”
No worries.
You can have the AI interview you.
Let it ask questions to learn your investment habits, risk preferences, and decision logic, then help organize an investor profile. Save it for repeated use.
A tip here:
When describing yourself, focus on facts, not adjectives.
Instead of saying:
I’m good at energy research.
Tell it:
Which companies you’ve researched over the past years, what judgments you’ve made.
Let AI analyze your competence circle from facts—it’s often more accurate.
Next, controlling direction.
If you don’t provide direction, AI will still give an average answer.
For example:
Help me analyze CATL.
Compared to a prompt with specified scope, structure, and output requirements, the result will be very different.
So when using AI, you must learn to add constraints.
I usually divide constraints into four types:
First: boundaries.
For example, only discuss battery business, only analyze 2026, ignore overseas business, etc.
The clearer the boundaries, the less AI will stray.
Second: structure.
Tell it which framework you want replies in.
Otherwise, it will just say whatever comes to mind.
Third: elements.
For example, must cite page numbers, must provide tables, must include risk warnings, etc.
Fourth: length.
Many overlook this, but length can directly impact quality.
With the same question, a 300-character and a 3000-character answer may have very different depth.
Constraints can be used in more advanced ways.
For example, role setting.
I once wrote an article making AI play roles like Goldman Sachs researcher or Renaissance quantitative analyst.
It can't know genuine internal workflows of those firms, but it can invoke relevant styles and expressions from its training data, making outputs more like those institutions’ frameworks.
Another way is intentionally broadening boundaries.
I previously wrote about Claude’s “God mode”—basically, explicitly allowing it to break conventions and think from less mainstream perspectives, to help generate more brainstorming-type inspiration.
So constraints aren’t always about narrowing borders—they can also mean deliberately opening them.
After lots of practice, you’ll see constraint methods go far beyond boundaries, structure, elements, and length—the forms are very flexible.
After setting constraints, if you want further control, you need to manage AI’s attention.
My habit is: one task, one window.
That is, discuss each issue in its own window. For instance, while doing company analysis, don’t ask about dinner in the same conversation. In the short term, it seems fine, but over time the context may mix up. During company decisions, it might wrongly think you care a lot about food & beverages, affecting answers.
So best practice: separate windows for different tasks.
Conversely, for big, long-term projects, don’t split too much. Often, maintaining one window is better than eight. A long conversation keeps the overall memory, letting AI continually understand your prior discussions.
Note large models have context length limits. If you talk too much, it may forget earlier material. My method: regularly write a handover document.
Have AI summarize the current project's core, conclusions, and next steps. Later, let it reread the doc and resume. This helps preserve your chat progress.
Also, re-emphasize rules during conversations.
If you want responses controlled to a certain length, remind it every time; if conclusions need page numbers, repeatedly emphasize. As chats drift off-topic, restate questions to guide back.
You can even tell AI at the start:
If you see I’m drifting, pull me back.
This keeps its attention on the current mission.
When AI replies aren’t satisfactory, use adversarial follow-up questions to push for deeper answers.
First follow-up: ask "why."
For example:
Is this conclusion based on evidence? Where’s it from? What’s the weakest link in this chain?
This lets AI attack its own response, recheck if it went deep enough. If not, it often improves itself.
Second: add hypothetical conditions.
For example:
What if the opposite situation happens? If this variable changes drastically, does the conclusion still hold?
The more extreme the hypothetical, the more AI is forced to reconsider and provide deeper answers.
Third: demand verifiable content.
Only when it gives specific info, data, or sources can we judge if it’s superficial or not.
But note: don’t make follow-ups too emotional.
If you simply ask:
Are you sure?
It may just say “yes.”
Better to give conditions, examples, and constraints—so it has the opportunity to overturn previous conclusions and give more authentic results.
Regarding conversations with large models, my third point is feedback.
AI’s answers naturally have two biases: hallucination and flattery.
Hallucination can be reduced by providing information; flattery needs feedback and cross-verification to correct it.
For example, put one model’s conclusion into another model, let it check if it holds up. For complex tasks, I often open three large models at once, and only act when their views are mostly consistent.
It’s similar for writing code. In Claude you can invoke Codex; sometimes I have Claude and Codex discuss together—when their results agree, then implement. This makes code more robust.
For decision questions, I recommend cross-verifying with models of different “bloodlines”; for code, focus on using Claude and Codex.
Another feedback method: send real world results back to AI.
AI’s knowledge comes from the web or historical texts—it doesn’t truly experience your reality. When it makes a judgment or suggestion, it’s best to feed your actual results back to it.
It will gradually learn your true situation and tailor its advice more to you.
So, before each conversation, do four self-checks:
First, did I pick the right model?
Second, did I provide enough information?
Third, am I consistently directing AI’s attention?
Fourth, after getting results, did I send them back for feedback?
If you do this every time, it can feel tedious. Is there a simpler way? That brings us to the second axe: make a Skill to handle repetitive work.
2. Skill: Turn repetitive tasks into personal capability assets
Skill is simple—it’s a reusable prompt pack.
That is, you package what you say to AI and the processes you want it to run as a fixed thing. In future, just call the Skill’s name, and it will automatically execute your flow and produce results.
Without Skill you have to restate structure and instructions every time; with Skill, you directly enter analysis per a fixed process.
Why are finance professionals especially suited to using Skill?
Because investment research processes are highly standardized, ideal for packaging. Financial analysis data sources and output formats are relatively fixed and repeated frequently, so compounding effects are obvious.
For example, reading a company’s financial report. If you redo the workflow each time, it’s exhausting. Better to make a Skill for reading financials.
Many people feel making Skills is troublesome, but actually, making a reading financial report Skill from zero only requires a few rounds of conversation with AI.
Step one: discuss workflow with AI.
You can directly ask:
I want to make a Skill for reading company financials. Can you help me design the workflow by asking questions?
It’ll ask about content to be read, PDF source, key metrics, output format, etc.
Step two: inject your personal style.
If you know your usual workflow, tell it; if not, have it keep asking questions via multiple choices or tables to help you sort it out.
Step three: add preferences.
If you want it to use a particular teacher’s financial analysis method, or your company’s template, tell it.
I made a basic Skill for reading financial reports. It produces an overall analysis, company business profile, business breakdown, financial analysis, rating, etc. It can present in your desired format. If your firm has a unique style, give it the template to mimic.
Skill methods aren’t complicated—and once you start, you’ll realize “everything can be a Skill.”
Reading financials, research reports, industry comparisons, quick reading of announcements—any repetitive and fixed-flow task can be packaged as a Skill.
Don’t aim for perfection at the start.
Skills don’t need to be perfect initially. Make one, use it once, find problems, then have AI optimize further. Iterate until it’s good enough, then keep it.
The point of Skills is to distill repetitive work, so every future task can save lots of time.
3. Github: Integrate global capability into your workflow
Finally, let’s talk about Github open-source projects.
Github can be seen as the world’s biggest “tool marketplace.” Many needs you can imagine have likely already been addressed, and often well. Many talented people release their tools as open source, so you can pick what you want from this marketplace.
Finance pros can use Github projects for many things: reading financials, reading market data, building agents, etc.
After opening Github, you can search for relevant projects. When selecting, I usually look at a few metrics.
First is Star count.
This shows how many people bookmarked the project. More stars means more popularity.
Second is Fork count.
Fork means how many people downloaded the project—shows real-world use.
Third is update time.
If the project is being updated, it’s maintained. We try to pick these. Otherwise, if your system updates but the project isn’t maintained, it may stop working.
Some projects have discussion zones. The more discussions, the more active it is—these are good to prioritize.
What if you can’t code at all?
No problem. After finding a project, you can tell your programming tool:
I need to use this project—please run it for me.
Today’s natural language programming tools can read code and help you set up and run it.
Of course, Github has countless projects—not all are usable out of the box. Broadly, some are fully usable, some offer conceptual inspiration, some need deep customization.
So, when you see a new project, have a large model analyze it—let the model help judge its suitability for your workflow.
If it fits, use it. If not, learn from its ideas—see if it can inspire your workflow.
In summary, today you can do three things:
First, write your investment research background documentation;
Second, try making your own Skill;
Third, find an open-source project on Github, and let a large model judge if it can integrate with your workflow.
4. About AI investment research: Thea answered some of the most frequently asked questions
Q: If you can't code, how do you start building your own AI investment assistant?
Thea: My suggestion is not to try making an Agent right away, but to start with a Skill.
A Skill is essentially a reusable set of prompts and workflow. You can directly tell AI:
I want to make a Skill to read financials—help me design the flow.
It’ll lead you step by step—how to read PDFs, extract data, analyze content, and output results.
For those unable to code, this is a good starting point. Once you distill a fixed workflow and can repeatedly invoke it, you already have a simple version of an Agent.
Q: How to use statements from tech leaders like Jensen Huang, Elon Musk to aid investment research?
Thea: I usually break the process into four steps.
First is data collection: company websites, personal social media accounts, Reddit discussions, etc.
Second is extracting core content—key new terms, repeatedly mentioned keywords, and content already getting actual feedback.
Third, build logical chains—what industries and technologies do these keywords link to, what future application scenarios may arise.
Last, map to specific companies—find which firms may benefit from these changes.
This works for Jensen Huang, Elon Musk, and any consistently outspoken people—listed company executives, OpenAI management, even industry experts can be tracked similarly.
Q: Will AI replace researchers? Which abilities are harder to replace?
Thea: I think the easiest to replace are repetitive tasks and work without personality.
If everyone uses the same model, prompts, and workflows, their results will increasingly converge.
What's hard to replace are those with personal traits.
With the rise of AI, I intentionally keep some abilities.
One is reading long texts. AI can summarize anything now, but if you only read summaries, you’ll lose the ability to process complex material over time.
Another is decision-making. Many things I let AI analyze, but the person taking responsibility is still me. So I intentionally keep my independent judgment and decision-power.
Q: When facing AI advice, do you trust it or yourself more?
Thea: I reference a lot of AI suggestions, but won’t rely completely on AI to decide.
Sometimes I record AI’s verdicts and my own choices, then review later to see if my thinking was at fault or AI had biases.
I use Skills designed for brainstorming and decision support. They don’t give direct answers, but help clarify what I really want through continuous questioning.
I always feel AI is best as a decision-supporter, not a decision-maker.
Instead of asking:
What should I do?
It’s better to guide it so it grows into a decision partner—understanding and approximating how you think.
Risk Reminder: The Master Class platform selects third-party compliant agency experts to teach investment research theory courses, the content does not constitute any specific product trading or investment advice. Opinions presented in platform courses are only for learning and reference, do not represent Wallstreetcn’s viewpoints, nor address users’ specific investment goals, financial status or needs. Markets are volatile and uncertain, the platform does not accept liability for losses incurred due to reliance on course views or information. Investment involves risks, please make decisions cautiously. ```