AI Stock Research Prompt
Ask an AI assistant to analyse a stock and it will happily invent the financials. This is a structured research prompt built to stop that — prefilled with your own position, signals and screener metrics, and copied to your clipboard from any ticker in st-ox.
The problem with asking an AI about a stock
Type "analyse NVDA for me" into a chatbot and you get something that reads like equity research: revenue figures, margins, per-share numbers, a valuation table, a confident verdict. The prose is excellent. The numbers are frequently wrong — not vaguely wrong, but specifically wrong, quoted to two decimal places against a quarter the model never actually looked up.
That failure is hard to catch precisely because it doesn't look like a failure. A hedge such as "approximately" or "assumed estimate" reads as diligence while doing the opposite, and once a made-up figure enters a comparison table it quietly feeds every ratio and score built on top of it.
The fix isn't a smarter model. It's a prompt that makes the model show its work, and a reader who knows which two or three things to check.
What st-ox puts in the prompt for you
The part a general chatbot can never reproduce is your own context. Every prompt st-ox generates arrives with the ticker's real, current data already filled in:
Live market data — price, 52-week range position, and whether the 200-day trend is rising or falling.
Your position — cost basis, return, trailing-stop level and whether it has been breached. Only if you hold it.
The entry-timing signal — the dip-buy verdict with the individual checklist rules that passed and failed, not just the colour.
Why it surfaced — if the ticker came from the screener, its composite score, factor sub-scores and the fundamentals behind them.
Candidate peers — other companies from the same screen, offered for the model to vet rather than asserted as comparable.
That last one is deliberate. Industry classifications are unreliable in both directions — a neobank can be filed under "Major Banks" beside JPMorgan, and a diagnostics company can be handed three biopharma peers — so the prompt names the classification it used and asks the model to reject peers that don't belong. In testing it does exactly that, with reasons.
Two depths, for two different questions
Quick take
One click, from any row
Three sections: forensic red flags, bull versus bear, and a short briefing. It answers the question a screener row actually poses — is this worth an hour of my time?
Available on every holding, watchlist and screener row.
Deep dive
The full report
All twelve sections, including moat, valuation against peers, DCF assumptions, catalyst calendar and management quality. A twenty-minute read.
Offered from a ticker's detail view, where you've already shown real interest.
What the report contains
The section list is fixed and mandatory — the model may not substitute an outline of its own. Structure is what makes an omission visible: you can see a section that came back thin, but you can't see a question that was never asked. Highlighted sections are the ones the quick take covers.
Company overview
Latest earnings call
Red flags
Competitive moat
Valuation vs peers
DCF assumptions
Catalyst calendar
Management quality
Bull vs bear
Plain-language checklist
The briefing
Composite score
Red flags runs to a fixed checklist rather than a free-form hunt: share count and what moved it, debt issued and repaid, buybacks and how they were funded, off-balance-sheet arrangements, dilution instruments, credit provisions against the growth of the book they provision, changes to non-GAAP definitions, and auditor or going-concern items. Every line must be answered — "nothing material" is required rather than a skipped line, because a gap you can see beats a gap you can't.
The rules that stop it making things up
These weren't written in advance. Each one exists because a real run broke it, and the prompt was tightened until it stopped.
Search first, then cite — with a source and a date.The report asks for latest-quarter figures, dated catalysts and insider ownership: things no model holds in memory. Without an explicit instruction it produces all of it anyway, fluently.
"Not verified" beats a labelled guess."Approximately", "roughly" and "assumed estimate" are named in the prompt and barred as substitutes. An unsourced figure may appear in a table if the cell says so — but it can never feed a ratio, a ranking or a score.
Compare like with like, and name the basis.Sell-side consensus is usually adjusted, not GAAP. Comparing a GAAP actual against an adjusted estimate turns an ordinary quarter into a dramatic "miss" — and revenue gross versus net of excise taxes does the same thing more quietly.
Quantify every beat and miss as a percentage.A revenue "miss" of $0.11M on $170.8M is 0.06% — rounding, not a finding. Anything under about 1% has to be declared noise rather than built into an argument.
Restate everything without the one-time item.A report can flag a large one-off settlement in one paragraph and lead with a "43.7% beat" in the next, while stripping the settlement turns that beat into a miss. Both facts were present; nothing joined them.
Bridge the guidance forward.Subtract what's already delivered and say whether the implied remainder accelerates or decelerates. When a stock falls on a quarter that beat, this is almost always the reason.
Reconcile the report against itself.A growth rate in the summary that contradicts the earnings table two sections earlier; a share count that can't be squared with the buybacks reported beside it. The arithmetic is checkable without leaving the page — so the prompt requires it.
An unverified line scores no severity.Severity describes a finding, and "I did not look" is not one. Scoring it manufactures a moderate concern out of an absence, so those lines are excluded from the total and counted separately instead.
How to use it well
Three habits catch almost everything the rules can't:
Use an assistant that can search the web. This is the single biggest factor in output quality, ahead of which model you pick. A model with no web access cannot honour a sourcing instruction no matter how firmly it's worded — it will cite a filing it never opened.
Spot-check two or three of the links. A citation is not a verification. In testing, one otherwise-careful report sourced a share price to an article about an entirely unrelated company — and a link that looks authoritative is harder to catch than a missing one.
On anything you'll act on, run it twice and compare. Two runs that disagree about a filed fact — a share count, whether a buyback happened — tell you immediately which figures need checking at the source. This is the most reliable control there is, and it costs one extra paste.
Why st-ox doesn't call the AI for you
The prompt is copied to your clipboard and nothing else happens. There's no server-side call to any AI provider, no API key to supply, and no data leaves st-ox for this feature.
That's a deliberate design choice rather than a missing feature. It keeps the app free of per-user AI costs, it lets you use whichever assistant you already trust and pay for, and — most importantly — it keeps the resulting report your research. st-ox supplies structure and your own data; it doesn't issue an opinion about a security.
Frequently asked questions
Can ChatGPT analyse a stock reliably?
Only if it can search the web, and only if the prompt forces it to cite what it finds. From memory, a language model produces fluent, specific, wrong financials that look exactly like the real ones. The fix isn't a better model — it's requiring a source and a date for every claim, and "not verified" where none exists.
What is an AI stock research prompt?
A structured instruction set that turns a general assistant into an analyst working to a fixed format — overview, earnings call, red flags, moat, valuation, DCF assumptions, catalysts, management, bull vs bear, checklist, briefing and composite score. The fixed structure is what makes an omission visible.
Does st-ox send my portfolio to an AI company?
No. The prompt is copied to your clipboard and nothing more — no server-side AI call, no API key, no data leaving st-ox. You choose where to paste it.
What's the difference between the quick take and the deep dive?
The quick take is three sections and answers the triage question a screener row poses. The deep dive is the full twelve-section report — a twenty-minute read — so it's offered from a ticker's detail view rather than from every row.
How do I know the AI hasn't made the numbers up?
Use a search-capable assistant, spot-check two or three of the links, and on anything you'll act on, run it twice and compare. Two runs that disagree about a filed fact tell you at once which figures to verify at the source.
Which stocks can I run this on?
Any holding, watchlist entry or screener row across the US and European markets st-ox covers. The one deliberate exception is the intraday Day Trade scan — a fundamental research report contradicts the premise of a momentum scan, so no button appears there.
Is this investment advice?
No. The prompt forbids a buy or sell recommendation, and the output is research for you to weigh. st-ox is not a licensed adviser, and the composite score is a synthesis of subjective judgements that's only comparable against companies scored in the same conversation.
Try it on a stock you're actually looking at
Open any ticker in st-ox and copy a research prompt prefilled with its live data, your position and the signals behind it. Free, no ads, no API key.
Open st-ox free →
For information and educational purposes only. Not investment advice, and not a recommendation to buy or sell any security. Reports generated by third-party AI assistants may contain errors — verify any figure you intend to rely on against the company's own filings.