One Of These Ten Prompts Needs No Market Data. It Is The One About Your Money.
No card this time, no graphic. A one-line post — “Turn Claude into Your Stock Market Analyst… 10 powerful prompts” — and then eleven comments from the author containing the whole thing.
Two notes before anything else. First, we read nine of the ten prompts plus the general instructions; prompt number 1 would not load for us without a Facebook login, so it is missing from everything below and we have not guessed at it. Second, and more usefully: this is the most carefully guarded prompt set this page has published, and the reason is the eleventh comment, which almost nobody will copy.
The source post — “Turn Claude into Your Stock Market Analyst”, BANI Academy. Ten prompts and a block of general instructions, all posted in full in the author’s own comments.
The best thing here is the comment nobody will copy
It is headed “General Instructions (Paste before any prompt)”, and it reads:
“Use plain English and explain financial terms. Confirm the company, ticker, exchange, and currency. Cite sources beside factual claims, with reporting periods and publication dates. Prefer company filings and investor relations materials. Separate verified facts, calculations, and assumptions. State the timestamp of market prices. If you cannot access current data, ask me to upload it. Never invent missing figures.”
Those last three sentences are the whole discipline of using a language model on anything with a price attached, and they are sitting in a comment under a post, in a place designed for people to scroll past.
Here is why they matter, stated plainly. A model’s training stops at a date, and that date is behind today. Without a live tool — a search, a connector, a file you paste — it does not know this morning’s price, last night’s filing, or yesterday’s guidance revision. And it will not feel uncertain about that. Asked for a P/E ratio it does not have, the structurally natural thing for a text model to produce is a plausible P/E ratio.
“Ask me to upload it” gives it somewhere else to go. That is the same trick the author used on a research-prompt deck we looked at recently, where every prompt carried a legal way to say I don’t know. Here it is one paste that covers all ten.
Diagram: where each prompt’s facts come from
Sort the nine prompts we could read by the question that actually determines whether the answer is any good: who supplies the numbers?
| Where the facts come from | Which prompts | What can go wrong | How you check it |
|---|---|---|---|
| You must paste or upload it | #8 technical setup (“using uploaded daily price and volume data”) | Your file is wrong, stale, or not split-adjusted | The prompt itself says to validate the dataset and check for stock splits |
| The model must fetch or you must supply it | #2 financial health, #3 earnings, #4 comparison, #5 valuation, #7 catalysts | Confident figures with no source behind them | Cite-every-figure, timestamps, and the “never invent” rule from the general instructions |
| You supply the argument, it supplies the attack | #6 thesis stress test, #10 monitoring framework | It argues well against a thesis built on figures you never verified | Verify the thesis first; the critique inherits the inputs |
| No outside data at all | #9 position sizer — one prompt of the nine | Nothing external. Only your own inputs and the arithmetic | A calculator. See below |
The only prompt that needs nothing is the one about your money
Prompt 9, the Risk-First Position Sizer, takes: account size, maximum account risk as a percentage, entry price, stop-loss price, target price, and a maximum allocation percentage. Every one of those is a number you type. There is no ticker in it. No price feed, no filing, no news.
So it is the only prompt in the set that cannot be wrong because of stale or invented data. It can only be wrong because of your own inputs, which you can read back. That is a real and underrated property, and it lands on the prompt that decides how much money is at stake.
The pattern generalises past investing: in this deck, the prompts that need the least from the world are the most reliable, and the glamorous ones are the ones whose answer quality was set entirely by what you fed in. A valuation engine sounds like analysis. It is arithmetic performed on assumptions you supplied, compared against a price the model may not have.
All nine, with the guard each one carries
| # | Prompt | What it needs | Its own guard clause |
|---|---|---|---|
| 2 | Financial Health Audit — three fiscal years, a table of revenue through to diluted shares | Filings for the company and its competitors | “Cite the source for every reported figure” |
| 3 | Earnings Intelligence Report — release, statements, call transcript | A specific quarter’s documents | “Separate GAAP results from adjusted figures”; analyst expectations only “when independently verified” |
| 4 | Competitive Edge Analyzer — rank several tickers | Consistent reporting periods across companies | “Identify metrics that could create misleading comparisons” |
| 5 | Stock Valuation Engine — bear, base, bull with formulas shown | Assumptions from you, a verified current price | “Compare with the latest verified market price and its timestamp”; “do not present estimates as guaranteed outcomes” |
| 6 | Investment Thesis Stress Test — five strongest counterarguments | Your thesis, plus filings and industry data | Rank objections by “evidence quality”; name the single biggest blind spot |
| 7 | Market Catalyst Radar — a dated timeline of material news | News and filings in a date range | “Distinguish confirmed facts from speculation”; no event “guarantees a particular stock-price movement” |
| 8 | Technical Setup Analyzer — moving averages, support, resistance | A price and volume file you upload | Validate the dataset; check for missing observations and stock splits; distinguish history from hypothetical setups |
| 9 | Risk-First Position Sizer — shares, capital, planned loss, reward-to-risk | Nothing but your six numbers | “Hypothetical paper trade”; “reject inconsistent inputs”; explain “why a stop-loss cannot guarantee the exact exit price” |
| 10 | Long-Term Investment Monitor — five indicators with thresholds | Verified baselines and reporting frequencies | “Clearly labelling any analyst-defined thresholds as assumptions”; next earnings date “only if verified” |
Ten sourcing guards, zero arithmetic guards
Read that right-hand column again. Every guard in the deck is about where a fact came from: cite it, verify it, timestamp it, label it an assumption, do not invent it. It is a sourcing discipline, and it is a good one.
Not one clause asks anybody to check a calculation.
Which matters most precisely where the sourcing guards do not apply. Prompt 9 has no sources to check — it is pure arithmetic from your six inputs — so its failure mode is not a fabricated fact. It is a quiet slip: a share count rounded the wrong way, a risk percentage applied to the wrong base, a reward-to-risk ratio computed from the target instead of the entry. Prose arithmetic is where language models are least dependable, and none of the deck’s guards is pointed at it.
Prompt 5 has the same exposure in a longer form. “Show the formulas, calculations, and implied per-share values” is excellent instruction design — it makes the working visible. Visible is not the same as checked.
The checks to add, by what can go wrong
| If the prompt’s risk is… | The deck already has | What to add |
|---|---|---|
| A fact that is not there (#2, #3, #4, #7, #10) | Cite every figure, prefer filings, timestamps, never invent | Open one cited source and check one number. If the citation is right, the rest usually is. If it is wrong, stop |
| A file that is wrong (#8) | Validate the dataset, check splits and gaps | Confirm the row count and the last date before reading a single indicator |
| An argument built on unchecked inputs (#6, #10) | Rank by evidence quality, label assumptions | Verify the thesis’s two load-bearing numbers before asking for the critique |
| A calculation (#5, #9) | “Show the formulas” — and nothing else | Recompute it. Put the six inputs in a spreadsheet, or ask a second conversation to redo it from the inputs alone and compare |
Four honest caveats
- This article is about prompt design, not investing. Nothing here is investment advice, we are not licensed to give any, and a blog about growing a YouTube channel is not where it would come from if we were. What we can usefully say is where a prompt’s facts come from and what happens when they are missing.
- We read nine of the ten prompts. Number 1 would not load without a Facebook login. It may be excellent; it is simply absent from the tables and from the count of guards.
- We did not run any of them. No ticker was analysed. This is a reading of the prompts against what a model can and cannot know, not a test of output quality.
- Training cutoffs are the mechanism, not a metaphor. A model is trained up to a date and then stops. That is a real property, it is behind today by months, and it is the reason “state the timestamp of market prices” is the sharpest instruction in the whole deck.
The one-line version
Ten prompts, and the most trustworthy one is the one that asks the world for nothing. Prompt 9 takes six numbers you typed and returns arithmetic — no ticker, no feed, no filing — which makes it the only one immune to stale and invented data, and also the one standing closest to your money.
Everything else is bounded by what you fed in. So the deck’s real instruction is not any of the ten: it is the eleventh comment, pasted first, every time. If you cannot access current data, ask me to upload it. Never invent missing figures. Then add the line it is missing — recompute anything that came out as a number — and you have a set of prompts that fails loudly instead of quietly.
If you want the general form of this problem, we wrote it up as prompts that ask for facts the model does not have; and for the opposite case — what a well-guarded deck looks like — the research prompts that give the model somewhere to put “I don’t know”.
The thirty-second test. Run prompt 9 twice with the same six inputs, in two separate conversations, and compare the share count and the reward-to-risk ratio. If they match, check one of them against a calculator anyway. That tells you more about whether this deck is usable than reading all ten.
