Short answer: I manage my stock portfolio with artificial intelligence (AI) by writing my buy, sell and cash rules into one file, keeping every open decision in a second file, and asking Claude or ChatGPT to run those rules every Monday on live screener data. The AI does the checking and the remembering. I make every decision and place every order myself. Below are the exact steps, the mistakes, and 9 prompts you can copy.
This article describes how I run my own portfolio. It is general information, not personalised investment advice. The rule settings below are mine; set your own.
Why I did this: I get lazy, and my biases creep in
I've invested my own money since 1987. I know the rules that work for me. And I still break them.
When I'm busy, I skip a review. When I'm nervous about the market, I drift to cash. When I like a company, I find a reason to keep it after my rule says sell.
Here's a real example. A few weeks ago I was holding a lot of cash. My reasons: momentum "was not very strong" and I was worried about an AI bubble. That felt like caution. It was a mood.
So I asked the AI to help me turn that mood into a rule. We wrote a cash rule based on three stock market indices and their moving averages (I explain it below). On the review date all three indices were above their averages. My own rule said: be fully invested. My gut said: wait.
That's the whole point of this setup. A written structure decides, not my mood on a Monday morning. The AI is the part that never gets tired of checking it.
The setup: 3 files, 2 idea stores and 1 connector
You need five things. None of them is complicated.
- A rules file. A plain text (markdown) file with every buy, sell and cash rule, with the numbers spelled out. You can dictate it in Claude or ChatGPT and it will write the file for you.
- An actions log. A second text file with every open decision, a date for each one, and every exception you have made.
- Your portfolio. A watchlist in the screener website is enough to start. I also keep a spreadsheet with my buy prices and all my sold trades going back years.
- Your idea stores. The company notes you write in the screener, and, if you use it, Evernote.
- The connector. A Model Context Protocol (MCP) connector lets Claude or ChatGPT read the Quant Investing screener directly: 22,000+ companies, over 110 ratios and indicators, your watchlists and your notes.
I run this in Claude. I've also tested the connector in ChatGPT (I back-tested a strategy there) and it worked just as well. Keep the two files where your AI can read them: in a Claude Project, a ChatGPT Project, or a folder Claude can open. If your AI can't save files, ask it to give you the full updated file and replace the old one yourself.
Step 1: Clean your data first
Before I wrote a single rule, I asked the AI to check my portfolio spreadsheet. I thought it was fine. It wasn't.
The AI tested every row against simple checks: number of shares times price equals the position value, the local price converted to euro correctly, the holding period matched the buy and sell dates, every trade had a strategy label. Then it ranked what it found by how much each error changed my results.
It found real mistakes. On Company H, a current holding, the euro purchase price had been typed in as the local price, so the gain was overstated. On Company D, the purchase value counted only half the shares sold. One error was big enough to turn a sold trade from a gain into a loss.
The AI also got things wrong. It flagged missing dividends on funds that never paid a dividend while I held them. I told it, and it stopped. That's normal: the AI doesn't know your portfolio until you tell it.
Prompt 1: check my portfolio data
I have attached my portfolio file. [Or: read my watchlist "My portfolio" on the Quant Investing screener.] Check every row for errors and missing values that would make my returns or performance wrong: - number of shares x price = position value - local price converted to my home currency correctly - buy and sell dates match the holding period - every holding and sold trade has a strategy label Rank what you find by its effect on my results in my home currency, biggest first. Do not fix anything yet. Ask me short multiple-choice questions about anything only I can know.
Rules that run on bad data give bad answers. Ten minutes here saves you from selling on a wrong number.
Step 2: Write your rules with the AI
I already had rules. Most of them lived in my head, and a few in spreadsheet columns. I asked the AI to read my files, tell me what it understood, and point out anything vague. Then I dictated each rule and the AI wrote it down.
The useful part was the pushback. Here's how four of my rules changed.
Trailing stop with a 2-week grace period
First idea: sell when a stock falls 20% from its highest price. But I like to watch a stock for a few days after it hits the stop.
AI pushback: "A few days" is easy to stretch again and again. How many Mondays, exactly?
Final rule: I check the 20% trailing stop (dividends included) every Monday. At the first breach I write the date in the actions log and wait at most 2 more Mondays. If the stock is still 20% or more below its high on the second Monday, I sell. No exception can keep a stock past that. Before a breach counts, I check that the fall isn't simply a large dividend that hasn't been recorded yet.
Example: Company A was 22.1% below its high. The dividend check showed no dividend waiting, so the breach counted, with a sell date 2 Mondays later.
The cash rule: Green, Amber, Red
First idea: "Stay out while momentum isn't clear, and be careful about an AI bubble."
AI pushback: how much cash you hold matters more than any stock-picking rule. Do you want a fixed market test instead of a judgement call?
Final rule: I watch three indices: MSCI World, the S&P 500 and the STOXX 600 (I live in Europe and spend in euro). I compare each one's last close with its 100-day and 200-day simple moving average (SMA), the average close of the last 100 or 200 trading days.
| Regime | Condition | What I do |
| Green | All three above the 100-day and 200-day SMA | Invest up to 100% |
| Amber | One or more below the 100-day SMA, all above the 200-day SMA | Cash may rise to 20%; new buys only with cash above 20% |
| Red | One or more below the 200-day SMA | No new buys; cash grows only from sells |
One line in my rules file matters most: a personal view, for example an AI-bubble worry, does not override the regime. If I want to override it, I have to write down why.
F-Score review, only when the data is complete
First idea: if a company's Piotroski F-Score (a 0 to 9 score of financial strength) falls below 5, review it.
AI pushback: a low score can come from missing data, not a weaker business. The screener has a field that counts the missing values.
Final rule: review only when the F-Score is below 5 and no values are missing. Otherwise the company goes on a data-gap list. On my first run, Company D (score 4, nothing missing) went to review, while three others with scores of 3 to 4 had 2 missing values each and did not.
Scale up winners (a rule I am still testing)
First idea: if a stock gains 20% and has strong momentum, double the position, up to 5% of the portfolio, with a tighter 10% trailing stop.
AI pushback: it led with the evidence against: the rule isn't back-tested, a normal dip can hit a 10% stop, and doubling every candidate would cost far more cash than my Amber regime allowed. It also listed its own design choices for me to confirm instead of quietly deciding them.
What I changed: the AI first ranked candidates by valuation. That's not what I wanted. This rule rides momentum, so I removed the valuation test. And because the market was Amber, I added half positions, not full ones.
Prompt 2: interview me and write my rules file
Interview me about how I manage my portfolio, one topic at a time: what I buy, how big each position is, when I sell (stop loss, losing positions, valuation, weaker fundamentals) and how much cash I hold and when. Lead with the evidence against any rule that is vague, or that I could bend when I'm busy or nervous. Then write my rules in a markdown file called portfolio_rules.md: - every number spelled out - the sell tests in a fixed order - which rule can never be overridden - a short glossary Show me the draft in the chat before you save it.
The AI's most useful job here is asking "exactly how many?" until a habit becomes a rule you can't wriggle out of.
Run your own Monday review from inside Claude or ChatGPT
The AI Screener connects Claude or ChatGPT to the Quant Investing screener: 22,000+ companies, over 110 ratios and indicators, your watchlists and your company notes. Every prompt in this article works with it, and you keep every decision.
100% money-back guarantee for the first 30 days.
Step 3: Keep your ideas where the AI can read them
Good ideas get lost. I read a fund manager's letter, find a strong argument for a company, and three months later I can't remember where I saw it.
Now I paste that passage into the company's note on the screener. For example: "Fund Manager X, first-quarter letter: why they bought Company M." The connector can read every note I've written, with its date, for any company. When I review my portfolio or my research list, the AI reads the notes too.
I've also kept articles, web clippings, company news and market views in Evernote for years. Evernote now has its own AI connector, so the AI can search that archive as well. If you are an AI Screener or a Best Deal subscriber, the Automate Your Investment Life ebook that comes with your subscription shows how Evernote fits into an investing workflow.
One practical point: on Claude's free plan you can add only one custom connector. To use the screener and Evernote together in Claude, you need a paid plan. ChatGPT needs a paid plan for the screener connector.
Prompt 3: what do my notes say?
Read my company notes on the Quant Investing screener for every company in my watchlists "My portfolio" and "Research". If Evernote is connected, also search it for each company name over the last 12 months. For each company, give me 2 lines: 1. what my notes say (quote the note and its date) 2. whether anything in them strengthens or weakens the case under my rules in portfolio_rules.md Use only my notes and the screener data. Do not add outside opinions.
Your research stops sitting in a drawer and becomes part of every review.
Step 4: An actions log that sorts every holding into 3 boxes
This is the file that changed the most for me. Every Monday, each holding lands in one of three boxes:
- Sell now, the rule fired. For example, a stock past its trailing-stop grace period, or an exchange-traded fund (ETF) in a momentum strategy that closed below its moving average.
- Review now. For example, a stock held 6 months or more that is still at a loss and behind the index.
- Already looked at, next look on a set date. An exception I've written down, with a reason, an end date and a condition that re-opens the case early.
The third box is the one that saves time. Company C was 20.6% behind the index after 2 years. My rules said review it. But it was still very cheap (Qi Value 1) and financially strong (F-Score 8). So I kept it, wrote the exception, and set it to re-open in April 2027, or earlier if the trailing stop reaches 20%, the Qi Value goes above 10, or the F-Score falls below 5.
Before this file, the same companies came up every single week. Now a decided case stays decided until its date.
Prompt 4: set up my actions log
Create a markdown file called portfolio_actions_log.md with these sections: 1. Market regime: date, index levels, regime 2. Open actions: date, company, rule, next step 3. Exceptions: company, rule, reason, end date, condition that re-opens the case early 4. Watch list 5. Data gaps 6. History: closed cases with the date and the result Rules for the file: - every exception must have an end date - a case with an active exception does not come up again before its end date, unless its re-open condition happens Read portfolio_rules.md and fill in sections 2 and 4 for my current portfolio.
Nothing comes up every week, and nothing gets forgotten.
Step 5: Tag every trade with its strategy
Not every idea I buy comes from the same place. Some are newsletter picks, some are coat-tail ideas (a stock bought because an investor I respect holds it), some are income stocks.
So every holding and every sold trade in my spreadsheet carries a strategy label. That lets me ask one question I couldn't answer before: how has each strategy done over the last 3 years, counting what I still hold and everything I've sold?
That answer tells me which idea sources earn their place, and which ones I keep out of habit.
Prompt 5: how has each of my strategies done?
Every holding and sold trade in my portfolio file has a strategy label in the column "Strategy". For each strategy over the last 3 years, show: - number of trades - % of trades with a gain - average return per trade - the biggest loss - the result against [my benchmark index] over the same holding periods Show current holdings and sold trades together, then separately. Tell me which strategies have too few trades to judge.
You judge your idea sources on your own record, not on how they make you feel.
Step 6: The Monday review
This is the routine that runs every week. The AI does the checking. I do the deciding.
| Step | What the AI does | What I do |
| 1. Open the actions log | Lists every action due today or earlier | Read it |
| 2. Market regime | Looks up the index closes and moving averages on the web, sets Green, Amber or Red | Check the source and date |
| 3. Screener data | Pulls F-Score, missing values, Qi Value, momentum and price history for every holding in one step | Nothing |
| 4. Sell tests | Runs them in the order in my rules file, skips cases with an active exception | Decide each case |
| 5. Buy list | Lists candidates if the regime allows new buys | Choose, size and place the orders at my broker |
| 6. Update the files | Writes decisions, exceptions and dates into the actions log | Check the changes |
The screener connector has no index data, so the AI gets the index levels from the web. Always check the source and the date it gives you.
Prompt 6: my Monday review
It's my Monday portfolio review. Read portfolio_rules.md and portfolio_actions_log.md first, and do every action dated today or earlier. 1. Look up the latest close of [MSCI World, S&P 500, STOXX 600] and their 100-day and 200-day simple moving averages on the web. Tell me the source and date. Set the market regime by my cash rule. 2. From the Quant Investing screener, pull my watchlist "My portfolio" with: price, Piotroski F-Score, F-Score missing values, Qi Value, Price Index 3 months and Adjusted Slope 90 days. Use the price history to find the highest close since each buy date for my trailing stop. 3. Run my sell tests in the order in my rules file. Skip any company with an active exception. 4. Give me the action list in the chat, grouped as: Sell now (rule fired), Review now, Next review date, No action. For each line, give the rule and the numbers behind it. Do not decide for me. I make every decision.
When I want a new rule, I test it on the whole portfolio within minutes of writing it.
Prompt 7: test a new rule before I adopt it
I want to add a new rule: [describe the rule in your own words]. Before you write it: 1. Give me the evidence against it. 2. List every design choice you would have to make, with your suggestion, for me to confirm. 3. Run it on my current portfolio and show which companies pass or fail, and why. Do not add it to portfolio_rules.md until I say so.
The review takes the checking off your plate, so you only spend time on the decisions.
Step 7: "Update the reference files": the habit that makes it work
This is the simplest step and the most useful one.
Every time I change a rule, and every time I finish a review and make decisions, I end the chat with one sentence: "Update the reference files." The AI writes the new decisions, exceptions, dates and rule changes into the two files.
So the next time I open a new chat, in Claude or in ChatGPT, it reads the files and starts exactly where I stopped. It knows what needs a review, what I've already looked at, and what waits until next quarter. I never re-explain my portfolio, and nothing gets lost between chats.
Prompt 8: update my reference files
Update the reference files. Add today's decisions, exceptions (with end dates and re-open conditions) and any rule changes to portfolio_rules.md and portfolio_actions_log.md. Move closed cases to History with the date and the result. Show me every change, line by line, before you save.
The AI forgets everything when a chat ends. Your files don't.
Step 8: Make the AI argue against you
An AI tends to agree with you. So I ask it not to.
When I wanted to keep a losing stock for another quarter, I asked for the case against keeping it first. The answer was blunt: "Two overrides on one stock usually means the case is being argued, not tested." I still decide, but now I decide with the other side in front of me.
The second question I use often: "Are you 100% sure?" When I asked it after a cost comparison, the AI listed what was exact, what was estimated and what was missing. Two of its answers changed.
Prompt 9: the case against
My [rule] says sell [Company A], but I want to keep it. Give me the case against keeping it first, then the pros and cons. If I still keep it, write the exception with a reason, an end date and a condition that re-opens the case. Follow-up for any analysis: Are you 100% sure? List what is exact, what is estimated and what is missing.
You get a second opinion that isn't in love with your stocks.
What went wrong, and how I fixed it
This didn't work perfectly on the first try. These were the main problems:
- The AI called my deliberate choices errors. It said comparing all my holdings with one European index was wrong. I explained why I do it (I live in Europe and spend in euro). It wrote the reason into my rules, so it never raised it again.
- A process step broke my spreadsheet. My rules file said "copy the row" when I sell a stock. That broke the spreadsheet's checks, because the columns were in a different order. The AI traced every cell that pointed at a holding and rewrote the step as an 11-step checklist. All 450 checks passed after the next sale.
- The AI used a bad source. For one comparison it took figures from a review website, which were wrong. When I gave it the official page, the answer changed.
- The AI added things I didn't want. It moved some values to a new place in my spreadsheet. I said no, and it put them back exactly as they were.
Check the AI's work the first few times. Each correction goes into the files, so the same mistake doesn't come back.
What the connector adds, and its limits
What I couldn't easily do before:
- Pull the F-Score, the F-Score missing values, the Qi Value and the momentum fields for every holding in one step, and apply my rules to each one.
- Read every company note I've written, by company and date.
- Run a brand-new rule on my whole portfolio within minutes, and run it again after every change.
What it does not do:
- It has no index data and no ex-dividend dates. The AI looks those up on the web or you type them in.
- It has no analyst forecasts. The screener is built on reported company figures.
- It doesn't trade. I size and place every order myself at my broker.
- It doesn't decide. Every sell, hold and buy in this article was my decision.
Two more uses worth trying: ask the AI to compare your broker's dealing costs per market (currency conversion is often the biggest cost, not the commission), and to list possible tax ideas for the year. Check any tax idea with your own tax adviser before you act on it.
The connector brings the data, your files bring the rules but the judgement of course stays with you.
How to start in 10 minutes
- Connect the screener to Claude or ChatGPT. The setup guide How to connect Claude to Quant Investing takes about 15 minutes.
- Put your current holdings in a watchlist on the screener.
- Run Prompt 2 to write your rules file. Don't aim for perfect. Write what you do today.
- Run Prompt 4 to create your actions log.
- Next Monday, run Prompt 6, then Prompt 8.
If you want a smaller first step, start with my weekly portfolio review with Claude, a single Monday prompt that shows your biggest movers and why. For more ready-made phrases, see the Quant Investing MCP commands library.
Remember: The first version of your rules will be rough. Running them every Monday is what sharpens them.
Put your own rules to work this Monday
With the AI Screener, Claude or ChatGPT reads your watchlists, your company notes and the screener data, runs your rules and gives you the action list. You make the decisions. The 9 prompts above are ready to copy.
100% money-back guarantee for the first 30 days.
Frequently asked questions
Can AI manage my portfolio for me?
Not in the way I use it. The AI checks data against rules I wrote and reminds me what is due. I make every decision and place every order myself.
Does this work in ChatGPT as well as Claude?
Yes. I run it in Claude, and I tested the Quant Investing connector in ChatGPT with the same results. ChatGPT needs a paid plan to use the connector.
Do I need a spreadsheet?
No. A watchlist on the screener is enough to start. A spreadsheet helps once you want to track sold trades and strategy results over several years.
What data can the AI get from the screener?
Over 110 ratios and indicators for 22,000+ companies, daily price history, your watchlists, your alerts and your company notes. It has no index data, ex-dividend dates or analyst forecasts.
Should I copy your rule settings?
No. The 20% trailing stop, the 2-week grace period and the cash rule are my settings for my portfolio. Use the prompts to write rules that fit how you invest.
Glossary
- Artificial intelligence (AI): here, a chat assistant such as Claude or ChatGPT.
- Model Context Protocol (MCP) connector: a link that lets the AI read data from another service, here the Quant Investing screener or Evernote.
- Markdown file: a plain text file with simple formatting that any AI can read.
- Trailing stop: the fall from the highest price since you bought, dividends included.
- Simple moving average (SMA): the average closing price of the last N trading days.
- Piotroski F-Score: a 0 to 9 score of a company's financial strength. Higher is better.
- F-Score missing values: how many of the 9 F-Score tests couldn't be calculated. 0 means the score is complete.
- Qi Value: the Quant Investing value rank. Lower is cheaper.
- Price Index 3 months: today's price divided by the price 3 months ago. Below 1.0 means the price fell.
- Adjusted Slope 90 days: the strength and steadiness of a price trend over 90 days. Higher means a stronger, steadier uptrend.
- Exchange-traded fund (ETF): a fund that trades on a stock exchange like a share.
- Exception (override): a written decision not to follow a rule for one company, with a reason and an end date.
This article is general information about how I run my own portfolio, not personalised investment advice. Share prices go up and down, you can lose money, and past results don't guarantee future results. Your investment decisions are yours. Read the full Quant Investing disclaimer.
