Creator Intelligence — AI + Finance

The 2026 AI Investing Blueprint

Hook: If I were starting over in AI investing in 2026, here's the entire blueprint. No code, no Bloomberg terminal.

If I were starting over in AI investing in 2026, here's the entire blueprint. No code, no Bloomberg terminal. I spent years at Goldman and SoftBank deploying over a billion dollars, and most of the edge came from doing boring research faster than everyone else. Step one, stop predicting the market, start building a system. Every morning, paste an earnings call into the AI and ask one thing. What did management quietly stop talking about since last quarter? Step two, build a checklist the AI runs on every name. Revenue trend, margin direction, where the cash actually goes. Step three, the part nobody does. Verify. The AI gives you a confident answer, you go pull the filing and check it. That habit is the whole game. Comment PLAYBOOK and I'll send you the morning prompt. Educational only, not advice.

The Mistake Killing Your AI Investing Edge

Hook: The biggest mistake people make with AI and investing is going straight to what should I buy before building the muscle.

The biggest mistake people make with AI and investing is going straight to what should I buy before building the muscle. AI is not your stock picker. It's your junior analyst. You'd never let a brand new analyst trade your account on day one. You'd make them pull filings, summarize calls, build the tables, and check their work until you trusted it. That's the muscle you're skipping. I've led over a billion dollars in deals, and the people who win with AI aren't the ones with the cleverest prompt. They're the ones who ask narrow, verifiable questions and check the source every time. Start today. Have it summarize one earnings call, then read the transcript yourself and find what it missed. Save this and follow for AI investing systems that hold up. Educational only.

The Earnings Call Prompt That Reads Like a Hedge Fund

Hook: Copy this prompt and turn ChatGPT into the analyst that reads earnings calls like a hedge fund.

Copy this prompt and turn ChatGPT into the analyst that reads earnings calls like a hedge fund. Most people paste a transcript and ask, summarize this. That gets you a press release. Instead tell it, act as a skeptical buy-side analyst, compare this call to the last two, flag three things management emphasized more and three things they quietly stopped mentioning, then list every number they gave with no context. That last line is the whole game. Companies love dropping a big growth number with nothing to compare it to, and the AI catches it instantly. But here's the rule that keeps you safe. The AI points you to the question, you go to the filing for the answer. Never trade on the summary. Comment EARNINGS and I'll send the full version. Educational only.

AI Won't Pick Your Stocks. It Does This Instead.

Hook: AI won't pick your stocks. Anyone selling you that is selling you a slot machine. Here's what it actually does.

AI won't pick your stocks. Anyone selling you that is selling you a slot machine. Here's what it actually does. It does the two hours of homework before you form a view. Point it at a 10-K and it pulls the revenue mix, the margin trend, and every risk factor that changed from last year. Point it at two earnings calls and it tells you what management quietly stopped saying. Point it at your portfolio and it flags if you secretly own the same bet four times. None of that is a recommendation. It's the grunt work that earns the right to have an opinion. The AI builds the draft, you verify the numbers, you make the call. Comment RESEARCH and I'll send the starter prompts. Educational only, not advice.

6 Claude Finance Projects You Can Build Without Writing Code

Hook: You don't need to code. You just need Claude. Here are 6 finance projects you can build in an afternoon.

You don't need to code. You just need Claude. Here are 6 finance projects you can build in an afternoon. One, rate any earnings call and see how the tone shifted from last quarter. Two, summarize a 10-K and pull every risk factor that changed this year. Three, stress test a model by asking what has to be true for the bull case to break. Four, check if your portfolio is secretly the same bet four times. Five, flag where revenue and cash flow quietly stopped agreeing. Six, turn a wall of indicators into one plain sentence. One rule keeps you safe. Claude builds the draft, you open the filing and verify every number before you act. Comment PROJECTS and I'll send all six. Educational only, not advice.

5 Free Resources to Learn AI for Investing

Hook: These 5 free resources teach you AI for investing better than any two thousand dollar course.

These 5 free resources teach you AI for investing better than any two thousand dollar course, and I spent zero proving it. One, the filings themselves, every annual report is free. Two, the model makers' own prompting guides, they tell you how to get sourced answers instead of confident guesses. Three, public earnings call transcripts to practice asking what changed. Four, a free financial statements course, because you can't verify AI if you don't know what a clean balance sheet looks like. Five, your own error log, every time AI misses a number and you catch it. Notice the pattern. None of them pick stocks for you. They teach you to direct AI and check it. Comment FREE and I'll send the list. Educational only.

Which AI to Use for Which Research Job

Hook: ChatGPT, Claude, Perplexity, Gemini. Running every market question through one of them is quietly poisoning your research.

ChatGPT, Claude, Perplexity, Gemini. Running every market question through one of them is quietly poisoning your research. After leading over a billion dollars in deals, here's how I actually split it. A long filing where I need exact citations with page numbers? The model that reads huge documents and stays anchored to the source. Live news I can click and verify? The one built for search, not the one guessing from memory. A messy valuation or a second order chain? The slow reasoning model that shows its steps. A fast rewrite of my notes? The cheapest one. Match the tool to the job, then verify against the primary source every time. The edge was never the model. It's knowing which one to point at which question. Follow for the workflows I actually run. Educational only.

A Quant Concept Most Investors Get Backwards

Hook: Most investors think volatility means risk. That one backwards idea is costing them.

Most investors think volatility means risk. That one backwards idea is costing them. Volatility is the price of the option, not the danger itself. Here's the 30 second version. When fear spikes, the cost to hedge spikes with it, and that premium is often richer than the actual risk being priced in. The pros don't try to predict the move. They sell the overpriced fear and manage the exposure around it. You don't need the math today. You need the lens. AI can build the model in minutes. You keep the judgment. Follow Wall Street Prompt for the concept of the week. Educational only, not investment advice.

This Week in AI x Markets, in 60 Seconds

Hook: Three AI headlines that actually move markets this week, in 60 seconds.

Three AI headlines that actually move markets this week, in 60 seconds. First, the big AI capex number. Remember it's a cost for the buyer and revenue for the supplier, so always read both sides of the trade. Second, the model launch everyone's posting about. Ask which company's margin it actually changes, not how impressive the demo is. Third, the quiet adoption story. The kind that shows up in a 10-K next quarter, not just in your feed. Here's the filter that runs all three. If a headline can't change a number in someone's filing, it's entertainment, not research. Follow for next week's three. Educational only.

Comment ASSISTANT and AI Builds Your Research Desk

Hook: Comment ASSISTANT and I'll send you the setup that turns ChatGPT into a 24/7 equity research assistant.

Comment ASSISTANT and I'll send you the setup that turns ChatGPT into a 24/7 equity research assistant. While you sleep, it reads a filing, pulls the revenue mix, flags what changed from last quarter, and drafts the exact questions you should ask on the call. It is not picking stocks. It is doing the two hours of grunt work before you even sit down. You wake up, you verify against the source, you decide. That split is the whole edge. Comment ASSISTANT, follow, and I'll DM you the setup. Educational only, not advice.

I Built an AI Earnings Analyst in One Afternoon

Hook: I built an AI earnings analyst in one afternoon, and it caught something the headlines missed.

I built an AI earnings analyst in one afternoon, and it caught something the headlines missed. I fed it two earnings calls, this quarter and last, and asked one question. What did management quietly stop talking about? It flagged a margin comment that simply vanished between the two calls. That's the tell. I did not trade on it. I went to the actual filing and checked it myself. That's the move. The AI finds the thread, you pull it. That's how you use this without getting fooled by a confident summary. Comment ANALYST and I'll send you the prompt. Educational only, not advice.

The 5 AI Tools I Open Every Morning Before I Look at a Stock

Hook: These are the 5 AI tools I open every morning before I look at a single stock.

These are the 5 AI tools I open every morning before I look at a single stock. One, a long-context model for reading filings, it stays anchored to the document instead of guessing. Two, a search tool for live news I can actually click and verify. Three, a transcript reader to compare this quarter's call to last quarter's. Four, a spreadsheet assistant that builds ratios I can audit, not a black box. Five, a notes tool that stores every sourced finding so the work compounds. Notice what's missing. Nothing here tells me what to buy. They clear the busywork so I can spend my attention on judgment. That's the stack. Comment MORNING and I'll send it. Educational only, not advice.

I Gave the Same 10-K to Four AIs. Only One Didn't Make Up a Number.

Hook: I gave the same 10-K to four different AIs. Only one didn't make up a number.

I gave the same 10-K to four different AIs and asked for the revenue mix. Only one didn't make up a number. The fast chat models gave me clean, confident answers, and two of them invented a figure that wasn't in the filing. The one that got it right was the model built to stay anchored to the document, and even then I checked every line against the source. Here's the lesson. The smoothest answer is the most dangerous one, because it reads finished so you stop checking. Match the tool to the job, force it to cite the exact line, and verify before anything touches a memo. The edge is the workflow, not the model. Follow for the workflows I actually run. Educational only.

Copie Este Prompt e Tenha um Analista de Equity Research 24/7

Hook: Copie este prompt e transforme o ChatGPT no seu analista de equity research particular, 24 horas por dia.

Copie este prompt e transforme o ChatGPT no seu analista de equity research particular, 24 horas por dia. A maioria pergunta, a TSLA é uma boa compra, e recebe uma resposta vaga e sem fundamento. Prompt ruim, tese ruim. Um analista de verdade não responde sim ou não. Ele monta um caso estruturado, tese otimista, tese pessimista, caso base. Então dê esse papel exato pra IA. O prompt, atue como um analista buy-side, para esta empresa me dê o caso otimista, o pessimista e três coisas que mudariam sua visão, e cite a fonte de cada afirmação pra eu poder verificar. Salve isso pra parar de receber resposta morna. Mas lembre, a IA monta o rascunho, você verifica o número na fonte e decide. Comente ANALISTA que eu te mando o prompt completo. Apenas educacional, não é recomendação.

3 Questions Claude Answers About a Stock Better Than a Junior Analyst

Hook: Here are 3 questions Claude answers about a stock better than most junior analysts.

Here are 3 questions Claude answers about a stock better than most junior analysts, and you don't need to code. One, what changed in this 10-K versus last year? It pulls every added, removed, and reworded risk factor in seconds, something that takes a human an afternoon. Two, what did management stop talking about? Feed it two earnings calls and it surfaces the topic that quietly disappeared. Three, where do the numbers stop agreeing? It flags where revenue and cash flow diverge, the classic quality tell. None of these ask should I buy. They do the reading so you can do the thinking. And you still open the filing to confirm every answer. Comment QUESTIONS and I'll send the prompts. Educational only, not advice.

Comment SETUP: The Exact AI Research Setup I Run Every Morning

Hook: Comment SETUP and I'll send the exact AI research setup I run before the market opens.

Comment SETUP and I'll send the exact AI research setup I run before the market opens. It's not one magic prompt. It's three tools wired together. First, the AI reads overnight filings and news and tags anything that touches a name I follow. Second, it drafts a one-line summary of what changed and why it might matter. Third, it lines up the questions I should answer before I act. By the time I sit down, the busywork is done and I'm spending my time on the decision, not the search. The AI surfaces, you verify, you decide. That's the whole setup. Comment SETUP, follow, and I'll DM it. Educational only.

If I Had to Relearn Investment Research From Scratch in 2026

Hook: If I had to relearn investment research from scratch in 2026, this is the exact AI roadmap I'd follow.

If I had to relearn investment research from scratch in 2026, this is the exact AI roadmap I'd follow. Week one, learn to read a filing with AI. Not summarize it, interrogate it. What changed, what's missing, what needs verifying. Week two, learn the comparison move. Feed it two earnings calls and find what management stopped saying. Week three, build one repeatable workflow end to end. Pull the numbers, check them against the source, output a one-page memo. Week four, learn to break your own thesis. Make the AI argue the bear case in the same session. That's it. No coding, no expensive terminal. Just reps until it's muscle memory. Comment ROADMAP and I'll send the full plan. Educational only, not advice.

The Craziest AI Finance Research of 2025, No Equations

Hook: The craziest AI finance research of 2025, explained without a single equation.

The craziest AI finance research of 2025, explained without a single equation. Researchers showed that a model reading thousands of earnings calls could flag the language that tends to show up before guidance gets cut, just from how management hedges. No ticker, no prediction. Just a pattern in how people talk when they're nervous. Here's why it matters for you. The edge was never a secret number. It's noticing the shift before the crowd does, and AI is very good at catching shifts in language at scale. But it spots the pattern, you confirm it in the filing. The machine reads, you judge. Follow Wall Street Prompt for the research that actually matters. Educational only.

I Can't Believe I Just Built This: An AI That Reads Every 10-K on My Watchlist

Hook: I can't believe I just built this. An AI that reads every 10-K on my watchlist the day it drops.

I can't believe I just built this. An AI that reads every 10-K on my watchlist the day it drops. The second a filing hits, it pulls the revenue mix, flags every risk factor that changed from last year, and drops a one-line summary of what actually moved. It took me an afternoon and zero code. Before, that was a junior analyst's whole week. Now here's the part I won't skip. It does not decide anything. It hands me a sourced draft, and I open the filing and verify every number before it touches a memo. The build is the easy part now. The judgment is still the job. Comment BUILD and I'll send you the setup. Educational only, not advice.

This One Move Turns Claude Into a Markets Research Analyst

Hook: This one move turns Claude from a chatbot into a markets research analyst.

This one move turns Claude from a chatbot into a markets research analyst. Stop giving it open questions like is this stock good. Give it a role and a checklist. Tell it, you are a skeptical buy-side analyst, here is the filing, for every claim cite the exact line, separate what management said from what the numbers show, and flag anything you cannot verify. The difference is night and day. A vague prompt gives you a confident guess. A role plus a checklist gives you sourced research you can actually audit. It still doesn't make the call. You do, after you check the source. Comment DESK and I'll send the full analyst prompt. Educational only.

5 ChatGPT Shortcuts I Use Daily to Research a Stock Faster

Hook: 5 ChatGPT shortcuts I use every day to research a stock in half the time.

5 ChatGPT shortcuts I use every day to research a stock in half the time. One, paste the filing first, then ask, so it answers from the document and not from memory. Two, force the citation. Tell it every number needs the exact line it came from or leave it out. Three, ask for the diff, not the summary. What changed versus last quarter. Four, make it tag each line as fact, inference, or opinion so you know what to trust. Five, end every session with, now argue the bear case. These aren't tricks. They're the difference between a confident guess and research you can defend. Save this and follow for the AI workflows I actually use. Educational only.

You Won't Be Replaced by AI. You'll Be Replaced by an Analyst Who Uses It.

Hook: You won't be replaced by AI. You'll be replaced by an analyst who uses it. Here are the 3 habits that separate them.

You won't be replaced by AI. You'll be replaced by an analyst who uses it. Here are the 3 habits that separate them. One, they don't ask AI for answers. They ask it to do the reading, then they form the view. Two, they verify everything. Every number the model gives gets checked against the filing before it counts. Three, they automate the boring 80 percent so they can spend their judgment on the 20 percent that actually moves a decision. The analyst who fears the tool gets slower. The one who directs it gets leverage. The model builds, you judge. Comment HABITS and I'll send the breakdown. Educational only, not advice.

The End-to-End AI Investment Research Workflow (One Filing to a Desk-Ready Memo)

Hook: Most analysts treat AI like a vending machine. They put in a vague question, they take out a confident paragraph, and they ship it.

Most analysts treat AI like a vending machine. They put in a vague question, they take out a confident paragraph, and they ship it. That is not a research process. That is a guess wearing a suit. The people building serious AI research systems are doing something completely different. They are not asking one big question. They are building a pipeline. A research operating system, where each stage hands clean work to the next. Today I want to take that idea and lay it on a research desk. One filing in. A desk-ready memo out. Five stages. I am not going to tell you what to conclude. I am going to give you the assembly line, and the discipline that keeps every stage honest. Here is the frame, and it is the thing most people miss. A research workflow is not a prompt. It is a sequence. The reason your AI output feels generic is that you are collapsing five different jobs into one request. You ask it to read, interpret, structure, draft, and check, all in one breath, and it does all five badly. Separate them. Each stage has one job, one output, and one handoff to the next. When you build it that way, the quality compounds instead of blurring. Stage one. Intake. Before the model reasons about anything, you hand it the source. The actual ten-K. The actual transcript. The actual eight-K. Not a question about the company. The document itself, loaded up front. Because the moment you ask first and make it go look, it starts filling gaps with whatever sounds right. So the first stage produces nothing but a clean inventory. What did I just give it. What period does it cover. What sections are in it. You are orienting the system before it interprets a single number. Stage two. Extraction. Now you pull the figures, and you pull them with a leash on. The instruction is not summarize this. The instruction is extract only the figures stated in this filing, and tag each one with the exact section it came from, because I need an auditable source. That phrasing matters. You are telling the model the why, not just the what. Revenue. Margins. Segment mix. Guidance language. Risk factor changes. Every number lands in a table with a source tag next to it. Anything the model cannot tie to a section, it flags as not found. Not invented. Flagged. Stage three. Interpretation. This is where the deltas live. You take this period against the prior period and you ask one disciplined question per line. What changed, and is that change positive, negative, or unclear. Not what does this mean for the stock. What changed in the document. Revenue driver shifted here. Margin compressed there. Management softened the language on this segment. Each delta gets a label, and the unclear ones stay unclear. You are not resolving ambiguity at this stage. You are surfacing it. Stage four. Drafting. Now, and only now, the model writes. It takes the extracted figures and the labeled deltas and turns them into a structured note in your house style. Feed it your past memos so it matches your voice, not a robot's. But here is the rule that does not move. Every line in that draft traces back to a source tag from stage two. If a sentence has no source behind it, it does not belong in the memo. The draft is a reorganization of verified inputs. It is not a new set of claims. Stage five. The audit. This is the stage almost nobody runs, and it is the one that protects you. You take the draft you just produced and you hand it back, with the original filing, to a fresh pass. One question. Does this memo make any claim the filing does not support. Point to it. That catches the number that drifted a decimal. The guidance you remembered slightly wrong. The inference that quietly hardened into a fact between stage three and stage four. The system checks its own work against the primary source before you ever read it. Now let me make the whole thing concrete. You are handed a quarterly filing the night before you cover the print. Stage one, you load the ten-K and the transcript and the system tells you what it is holding. Stage two, it extracts revenue, margins, segment data, and guidance, each tagged to a section. Stage three, it lines this quarter against last and labels every change positive, negative, or unclear. Stage four, it drafts the note in your format, every line sourced. Stage five, it audits the draft against the filing and flags two claims that overreach. You open it in the morning, you read five minutes of structured work instead of fifty minutes of raw filing, and you spend your real time on the two flagged claims. That is the difference between faster reading and faster being wrong. Be honest about where this breaks, because it does break. The model can be confidently wrong at any stage. It can attribute a real number to the wrong section. It can sound certain about something that is not in the document at all. That is exactly why the pipeline is built the way it is. The source tags are not bureaucracy. They are how you catch the hallucination before it reaches the memo. And the work lives as plain files. Tables, notes, drafts. That means your process is portable. You are not locked into one tool. If a price moves or a feature disappears, you carry the whole workflow somewhere else and keep going. So here is where I land. The edge was never asking AI a smarter question. The edge is the assembly line you wrap around it. Intake, extraction, interpretation, drafting, audit. Five stages, each with one job, each handing clean work to the next. And through all five, the line holds. The system builds the memo and it builds it fast. It does not make the call. You own the assumptions, you own the classification on every unclear line, and you verify every number against the primary filing before it leaves your desk. Human judges. AI builds. The pipeline is the speed. Your judgment is still the standard.

Derived shorts: The Vending Machine Mistake: Why Your AI Research Reads Like Everyone Else's · Five Stages, Not One Prompt: The Research Pipeline Analysts Skip · The Last Stage Nobody Runs: Make The AI Audit Your Own Memo

Build an AI Research Agent for Finance That Is Not a Stock Picker

Hook: The most dangerous AI agent in finance is the one that works perfectly and tells you exactly what to buy.

The most dangerous AI agent in finance is the one that works perfectly and tells you exactly what to buy. Because the day it is wrong, and it will be wrong, you will have already stopped checking it. There are tutorials all over the internet right now walking hundreds of thousands of people through building their first automation agent. Most of them are good builds. But the agent most finance people will copy points at the wrong target. They aim it at a decision. Buy this. Sell that. Today I want to show you how to build the agent an analyst should actually build first. A research agent. One that reads everything you read, faster, and never once tells you what to do. I will break down the three parts of any agent, then point all three at a research job instead of a trading job. Start with the architecture, because every agent is the same three things. A brain. Tools. And the rules that govern both. Get those three right and the agent is an assistant. Get them loose and the agent is a liability with a confident voice. Part one. The brain. The brain is a model plus memory. The model is the reasoning. The memory is what it holds across time. In most demos the agent is told something, and one step later it has forgotten it completely. The fix is memory. For an analyst, memory is not a nice-to-have. It is the entire point. Your coverage list is the memory. The names you track, the thesis you wrote last quarter, the conclusions you already reached. You write that down once, so when you open the agent on a Sunday it already knows what you told it on Monday. A brain with no memory is a chatbot. A brain with memory is a junior analyst who remembers the file. Part two. The tools. This is where finance people get burned, so listen closely. The tools are what the agent can actually reach. In a typical build, you wire it into a sheet and it can write rows. For a research agent, you flip the default. Every tool is read-only first. It can read a new ten-K. It can read an eight-K. It can read an earnings transcript. It can read your tracker. What it cannot do, on day one, is write a buy or a sell anywhere, send anything, or touch execution. And here is the rule underneath it. Assume that if the agent can do something, eventually it will, whether you told it to or not. So you do not write an instruction that says never trade. Instructions are not capabilities. You remove the key from the key ring. Market data and filings, read-only. Execution, not connected at all. You control the permission layer. The agent controls nothing you did not physically hand it. Part three. The rules. The system prompt. This is the part people skip, and it is the part that matters most. For a research agent, the rules say three things and they are not optional. One. Cite the source, the section, and the date on every single line. Two. Classify every change as positive, negative, or unclear. No vague summaries. Three. Confirm before you record anything, and never act on a guess. Those three rules are not features you add for polish. They are the standard that keeps the agent from quietly becoming a stock picker. Because notice what they force. An agent that has to cite a source and confirm before it writes cannot hand you a naked recommendation. It can only hand you sourced changes and open questions. Now here is the most useful thing about building one of these. The first real run should fail, and it should fail in a way that teaches you. In the tutorials this idea comes from, the first run writes its output, but it writes it without confirming, and it stamps the wrong date. The model was not broken. The instructions were loose. The fix was two lines. Always confirm, even when the data looks complete. And pull the date dynamically from the actual document, instead of letting the model guess. Map that onto research. An agent that drafts a filing summary without confirming, with the wrong reporting period stamped on it, is not a faster analyst. It is a faster way to put a wrong number in a memo. Same fix. Tighter rules. Cited sources. A confirmation step before anything is recorded. So let me give you the whole build, end to end. A morning research agent for your coverage list. The brain holds the list and your prior conclusions. The tools read filings and transcripts, read-only, nothing else connected. The rules force citation, classification, and confirmation. Every morning it produces one brief. For each name, it gives you what was filed, the exact source and section, the date, what changed, the change labeled positive, negative, or unclear, and one line on what to verify next. It does not say buy. It does not say sell. It does not name a price target. It hands you a desk-ready brief and a list of open questions, and you read it in five minutes instead of fifty. Now be honest about the limits, because the agent has real ones. It can misread a filing. It can attribute a number to the wrong section. It can sound completely confident and be flat wrong. A black-box agent can hallucinate, and a badly scoped one can reach data you never meant to expose. So the read-only default is not timidity. It is the design. And the cite-the-source rule is not red tape. It is how you catch the hallucination before it costs you anything. You supervise this agent the way you would teach someone to ride a bike. You watch it run. You correct it. Only once it is reliable do you let it run on its own. A skill you have not corrected is a skill you cannot trust. Here is the line that does not move. The agent owns construction and speed. You own the assumptions and the decision. Every claim it makes traces to a primary source you can open yourself, on EDGAR, in the transcript, in the filing. Every brief ends with what to verify, not what to do. The machine prepares you. It does not decide for you. So the takeaway is simple. The edge was never the agent that picks the stock. That agent is a black box you should not trust, and the day it breaks you will not even notice. The edge is the agent that does the reading and shows its work, the one you can supervise line by line. A stock-picker agent is a bet you cannot audit. A research agent is a junior analyst you can. Build the second one first. Human judges. AI builds. This is educational only. Not investment advice. The agent structures the research. You make the call.

Derived shorts: The Most Dangerous Agent In Finance Is The One That Works · Three Parts Of A Research Agent: Brain, Tools, Rules · Why The First Run Should Fail On Purpose

The AI Productivity Stack That Turns One Memo Into a Full Research Workflow

Hook: You are paying for four AI tools and running everything through one of them. That is how a research desk quietly produces confident, unchecked work.

You are paying for four AI tools and running everything through one of them. That is how a research desk quietly produces confident, unchecked work. The failure mode is silent. You hand one general tool a job it was never built for, it gives you a clean-looking answer, and you never trace it back. Today I want to fix that. I am going to show you how to build an AI productivity stack for finance, and then I am going to show you how that stack turns a single memo into a full, repeatable research workflow. The frame here comes from a Jeff Su walkthrough of his own stack, and it maps onto a research desk almost perfectly. The core idea is this. Your desk is not one job. It is five. One job, one kind of tool, then you verify. Stop asking which AI is best. Start asking, best for which job. Job one. Research and synthesis. This goes to a strong general-purpose model, and the thing you are buying is instruction-following. The best one takes a long checklist and does not silently drop steps. Why does that matter for you. Your research prompt is a checklist. Compare this quarter to last. Pull the revenue drivers. Pull margins. Pull guidance. Pull the risk factor changes. Flag any shift in management language. A model that quietly skips three of those is worse than useless, because it looks finished and it is not. So your first synthesis pass goes to the model that actually follows orders. Job two. Reading mixed material. This goes to a multimodal model, one that takes audio, video, slides, and text together. Picture your real inputs the morning after an earnings event. The call recording. The slide deck. A photo of an exhibit you snapped off the printout. Most tools read only the text. One model takes all of it at once and tells you what was discussed and what changed. For you that is the difference between transcribing for an hour and reading a synthesis in a few minutes. So ingesting calls, decks, and exhibits goes to the multimodal tool. Job three. The last mile. The draft you actually send, and the small data pull you keep punting on. This goes to a draft-quality model, the one where the first attempt is closest to done. Two things matter here. First, voice. Feed it your past memos and it writes in your house style, so the draft reads like your desk and not like a machine. Second, working code. The kind of quick script you always hand to someone else. Cleaning a CSV. Reshaping a comp table. Charting a series. The draft-quality model gets you a first pass on both. So polishing the memo and writing the quick pull goes here. Job four. The fast fact. This goes to a search specialist, and there is a clean line you should hold. Reasoning models are built to think. Search tools are built to fetch. So you do not ask a search tool to reason. You ask it for one specific, current fact, right now. Was guidance reiterated or cut. What is the current share count. When is the next print. Think of it as the scalpel that checks the thing the reasoning model just told you. The chatbot drafts the thesis. The search tool verifies the facts inside it. Job five. This is the one I care about most. Checking your work against your own sources. This goes to a source-grounded tool, one that answers only from the documents you upload and has no outside knowledge to invent from. Here is the move. Before you publish, you upload your draft memo plus the actual filing, and you ask one question. Does this memo make any claim the filing does not support. That catches the quiet errors. The number that drifted a decimal. The guidance you remembered slightly wrong. The inference that hardened into a fact. So checking your draft against your primary sources goes to the source-grounded tool. Now here is the part that makes the stack worth building. Once you have these five jobs mapped, a single memo stops being a one-off and becomes a repeatable workflow. Watch how one filing flows through the whole stack. You drop the call recording and the deck into the multimodal tool and get a synthesis of what changed. You hand that, with the filing, to the instruction-following model and run your full research checklist against it. You take the structured output to the draft-quality model and it writes the memo in your voice and pulls the comp data with a quick script. You take every number it produced to the search tool and verify the live facts, the share count, the guidance, the next date. And finally you hand the finished draft and the filing to the source-grounded tool and ask it to flag any unsupported claim. One memo. Five stations. A workflow you run the same way on every name you cover, so the next memo is faster than this one. Now the heads-up most tool videos skip. A source-grounded tool is only as good as the sources you feed it. Bad source in, confident wrong answer out. In finance that is the whole game. And notice what none of these five tools did anywhere in that workflow. None of them decided anything. They drafted, they read, they fetched, they checked. They did not pick a name. They did not set a price target. You own the assumptions. You own the source list. You own the call. So the rule does not move. The tool builds. You verify against the primary filing. No source, no claim. The tool names are workflow examples, not endorsements. AI structures the research process. It does not pick the stock, and it does not replace the filing. Treat every output as a draft until you have traced the number back to the primary source yourself. The goal is to be better prepared, not more confidently wrong. Map your five jobs to the right tools, run one memo through all five, and you have a research workflow you can repeat every week. Human judges. AI builds. This is educational only. Not financial, legal, tax, or investment advice.

Derived shorts: One Memo, Five Jobs: The Stack Finance Pros Run Wrong · Stop Asking Which AI Is Best, Ask Best For Which Job · The Source-Grounded Check That Catches Your Quiet Errors