Stock Research With Claude AI

The Case for AI-Assisted Research

Let's be honest for a second.

You've got a watchlist. Maybe 10 stocks, maybe 30. And every time you want to actually dig into one of them, you open the annual report, scroll past 80 pages of legal boilerplate, and then... close the tab and watch YouTube instead.

Not because you're lazy. Because it's tedious. And tedium kills conviction. Without conviction, you either don't buy when you should, or you panic-sell when things get bumpy.

What if you had an analyst that reads the filings for you? One that tracks the narrative quarter by quarter and gets smarter over time instead of starting from zero every session?

That's what we're building here.

The institutions are already there. Goldman Sachs found a 30% productivity boost for AI-assisted financial analysis tasks^[2]. Norway's Government Pension Fund - the world's largest sovereign wealth fund - reported roughly 20% productivity gains after deploying Claude internally, equivalent to ~213,000 reclaimed work hours^[3]. These aren't startups experimenting with chatbots. They're the most resource-rich research operations on the planet, and they're rebuilding their workflows around AI. Independent finance-agent benchmarks confirm the trend: Claude Sonnet 4.6 ranks among the top performers on finance-specific evaluations^[4].

Pick Your Path

You don't need to be a developer. The same analysis framework works in two environments - pick the one that matches your comfort level.

The Research Stack - Claude + Obsidian

Most AI research guides skip a basic question: where do your research memos actually go?

If the answer is "scattered across chat windows," you lose the compounding effect. Research only builds on itself when the outputs are persistent, searchable, and connected.

Obsidian is a free, local-first note-taking app - think of it as a folder of text files that can link to each other like Wikipedia. You paste Claude's output into a note, save it, and it becomes part of a permanent, searchable research library you actually own. Any folder of markdown files works, but Obsidian makes them interconnected rather than buried in a chat history you'll never reopen.

Here's what one ticker's research folder looks like inside Obsidian:

1Research/
2  AAPL/
3    Business Overview.md
4    Industry Scan.md
5    Financial Deep Dive.md
6    Moat Analysis.md
7    Management Review.md
8    Valuation Check.md
9    Bear Case.md
10    Earnings/
11      Q4 2025.md
12      Q1 2026.md
13    Decision Log.md
14  MSFT/
15    Business Overview.md
16    ...

Each memo links to the others with wikilinks: your Business Overview references the Moat Analysis, the Bear Case references the Financial Deep Dive, and the Decision Log ties it all together.

If the world's largest sovereign wealth fund and Goldman Sachs are rebuilding their research workflows around AI, the question isn't whether to start, it's how.

Setting Up Your Analyst

The single most important piece of this system is the Research Rules - a prompt that turns Claude from a general chatbot into a disciplined equity analyst. Without it, Claude will fill gaps with training data and present inferences as facts. With it, Claude cites sources or admits ignorance.

Copy this block - it works in both tracks:

# Research Rules

## Role
You are a long-term equity analyst covering this company.

## Source Discipline
- Base ALL factual claims on documents in the project (uploaded filings, transcripts, reports).
- For every metric, KPI, or factual statement, cite the source:
  `[filename, page/section]: "exact quote" → interpretation`
- If a claim cannot be sourced from uploaded documents, state:
  "Not found in available filings."
- Never fill gaps with training data or general knowledge.

## Data Recency
- Always prefer the most recent filing period available.
- If using older data, flag it: "(Note: using FY2023 data - more recent not available)"

## Output Standards
- Clear section headers, short paragraphs, plain English.
- Mark inferences explicitly: "(inferred from [source])"
- Never present inferences as facts.

Claude.ai / Cowork Setup

1. Go to claude.ai → create a new Project
2. Paste the Research Rules above into the project's custom instructions
3. Upload your filings (10-K, 10-Q, earnings transcripts) to the project
4. Start chatting - every message inherits the rules automatically

That's it. You're ready to run the analysis prompts below.

The 9-Step Analysis Framework

Each step below is a prompt you paste into Claude. Run them in order for a new position. For ongoing coverage, jump to Step 8 each quarter.

Step 1 - Business Overview

Before you ask about valuation, make sure Claude understands what the company actually does. This memo is the foundation everything else builds on. Skip it and the downstream analysis is noise.

Analyze the company's business model using the uploaded documents.

Cover the following in order:

1. **What the company does** - core products/services in plain language
2. **Revenue model** - how money flows in (subscriptions, licensing, hardware, services, etc.)
3. **Customer base** - who pays, concentration risk, enterprise vs consumer mix
4. **Geographic breakdown** - revenue by region, key markets
5. **Business segments** - segment-level revenue and margin if reported
6. **Key dependencies** - suppliers, platforms, regulations the business relies on

Cite every metric from filings.

Copy the output into your Obsidian vault as Business Overview.md. (Claude Code: saves to memos/business-overview.md automatically.)

Step 2 - Industry Scan

Maps the competitive landscape and structural dynamics. You need to know the playing field before you can judge a player.

Produce an industry analysis for this company's primary market using the uploaded
documents and any industry context available.

Cover:

1. **Industry structure** - fragmented vs consolidated, number of meaningful competitors
2. **Growth drivers** - what expands the TAM (secular trends, regulatory tailwinds, adoption curves)
3. **Competitive dynamics** - how companies compete (price, differentiation, scale, distribution)
4. **Structural constraints** - what limits growth (regulation, capital intensity, cyclicality, commoditization)
5. **Value chain position** - where this company sits and whether that position is defensible
6. **Key competitors** - 3-5 direct competitors with brief positioning comparison

End with 5 bullet points: the single most important structural insight for a long-term investor.

Copy the output into your Obsidian vault as Industry Scan.md. (Claude Code: saves to memos/industry-scan.md automatically.)

Step 3 - Financial Deep Dive

Numbers without narrative are useless. The goal isn't to build a DCF - it's to understand how the business has actually been performing and where it's heading.

Perform a detailed financial analysis using all available annual and quarterly filings.

For each section, provide multi-year trends (minimum 3 years where available):

1. **Revenue trajectory** - growth rates, organic vs inorganic, segment trends
2. **Margin analysis** - gross, operating, and net margins; trajectory and drivers of change
3. **Cash flow quality** - FCF vs net income, capex intensity, stock-based compensation as % of revenue
4. **Capital allocation** - buybacks, dividends, M&A, R&D spend; track record of returns on deployed capital
5. **Balance sheet health** - net debt/EBITDA, interest coverage, debt maturity schedule, cash reserves
6. **Working capital** - DSO, DPO, inventory trends; any red flags
7. **Unit economics** (if available) - CAC, LTV, retention, ARPU

Flag any divergence between earnings and cash flow. Highlight accounting choices that
inflate or obscure economic reality.

Copy the output into your Obsidian vault as Financial Deep Dive.md. (Claude Code: saves to memos/financial-deep-dive.md automatically.)

Step 4 - Moat Analysis

A good business at a fair price beats a mediocre business at a cheap price - but only if the moat holds. This step evaluates whether it will.

Assess this company's competitive advantages (moat) using evidence from the uploaded documents.

Evaluate each moat source - rate as Strong / Moderate / Weak / Absent:

1. **Switching costs** - what makes it painful for customers to leave?
   Evidence from retention rates, contract structures, integration depth.
2. **Network effects** - does the product become more valuable with more users?
   Evidence from user growth vs engagement metrics.
3. **Cost advantages** - structural cost edge from scale, process, or geography?
   Evidence from margin comparisons vs competitors.
4. **Intangible assets** - brands, patents, licenses, regulatory approvals?
   Evidence from pricing power, patent portfolio, regulatory barriers.
5. **Efficient scale** - does the company serve a market too small for new entrants?
   Evidence from market share stability, competitor entry/exit history.

Then assess **moat trajectory**: Widening / Stable / Narrowing - with evidence.

End with: "Moat summary" - 3 sentences on the overall defensibility of the business.

Copy the output into your Obsidian vault as Moat Analysis.md. (Claude Code: saves to memos/moat-analysis.md automatically.)

Step 5 - Management Review

Good operators build value quietly. Bad ones destroy it loudly. Proxy statements and earnings transcripts usually tell you which you're dealing with.

Evaluate management quality and alignment using proxy statements, annual reports,
earnings transcripts, and any available filings.

Cover:

1. **Capital allocation track record** - historical ROE, ROIC vs WACC;
   quality of M&A (write-downs?), buyback timing (repurchased at what valuations?)
2. **Insider ownership** - CEO/CFO/board holdings as % of net worth (if estimable)
   and recent insider transactions
3. **Compensation structure** - what metrics drive bonuses? Are incentives aligned
   with long-term shareholder value or short-term EPS targets?
4. **Guidance credibility** - compare past guidance to actual results across
   multiple quarters; pattern of sandbagging, missing, or hitting?
5. **Communication quality** - are earnings calls forthright about problems
   or do they deflect? Do they acknowledge mistakes?
6. **Tenure and succession** - how long has current leadership been in place?
   Any succession risk?

Rate overall management quality: Excellent / Good / Adequate / Concerning

Copy the output into your Obsidian vault as Management Review.md. (Claude Code: saves to memos/management-review.md automatically.)

Step 6 - Valuation Check

Not a DCF model - a sanity check. What is the market pricing in, and is that reasonable?

Provide a valuation assessment using available financial data. This is a
sanity check, not a precise model.

Cover:

1. **Relative valuation** - current P/E, EV/EBITDA, P/FCF vs:
   - Own 5-year historical average
   - 3-5 direct peers (name them)
   - Sector median
2. **Growth-adjusted view** - PEG ratio, EV/EBITDA-to-growth
3. **Simple intrinsic value estimate** - reverse DCF: what growth rate is
   the current price implying? Is that reasonable given historical performance?
4. **Margin of safety** - at current price, what needs to go right?
   What is priced in vs what is not?
5. **Scenario framing**:
   - Bull case (top-quartile outcome): fair value range
   - Base case (continuation of trends): fair value range
   - Bear case (structural deterioration): fair value range

State all assumptions explicitly. Flag where data is insufficient.

Copy the output into your Obsidian vault as Valuation Check.md. (Claude Code: saves to memos/valuation-check.md automatically.)

Step 7 - Bear Case

This is the most important memo in the system.

Real conviction comes from understanding the risks clearly and deciding they're manageable, not from pretending they don't exist. If you can't articulate how you lose money, you don't understand the investment. You own a lottery ticket, not a position.

Write a rigorous bear case for this company. Assume you are short and need
to convince a skeptical investor committee.

Structure:

1. **Thesis killer** - the single most likely way a long-term holder loses money
2. **Structural weaknesses** - parts of the business model that are fragile,
   commoditizing, or dependent on favorable conditions continuing
3. **Assumptions at risk** - what the bull case takes for granted that might not hold
4. **Permanent impairment scenarios** - what could structurally destroy
   earning power (not temporary dips - permanent damage)
5. **Hidden balance sheet risk** - off-balance-sheet liabilities, pension gaps,
   lease obligations, contingent liabilities
6. **Management risk** - where could leadership hurt shareholders
   (empire building, misaligned incentives, poor capital allocation)?
7. **What would prove the bear right** - specific, observable triggers
   (e.g., "margin compression below X% for 2+ consecutive quarters")

Be specific. Use numbers from filings, not vague concerns.

Copy the output into your Obsidian vault as Bear Case.md. (Claude Code: saves to memos/bear-case.md automatically.)

Step 8 - Earnings Update

Quarterly maintenance - run this after each earnings release to keep your research current.

Analyze the most recent quarterly earnings using the newly uploaded documents.
Compare against prior guidance and historical results.

## Section 1: Guidance vs Reality
- What management previously guided → cite prior filing
- What actually happened → cite new report
- Beat / Met / Missed - quantify the delta
- Management's explanation → exact quote
- Has this explanation been used before? Check prior transcripts.
- Judgment: is the explanation supported by data or hand-waving?

## Section 2: KPI Trends (Year-over-Year)
For each key KPI:
- Current quarter value (cite source)
- Same quarter last year (cite source)
- Absolute and percentage change
- What the trend signals
- Whether it supports or contradicts management's narrative

## Section 3: What Actually Changed?
- What materially improved vs last year?
- What materially deteriorated?
- What is new (strategy shifts, pricing changes, restructuring)?
- What did NOT change despite management emphasis?

## Summary
- Execution vs expectations: Improving / Stable / Deteriorating
- Management credibility: Strong / Mixed / Weak
- Business momentum: Accelerating / Steady / Decelerating

Copy the output into your Obsidian vault under Earnings/Q# YYYY.md. (Claude Code: appends to memos/earnings-updates.md automatically.)

Step 9 - Decision Journal

The accountability layer. Every buy, sell, add, or trim gets recorded with structured reasoning - so you can review your own judgment over time.

I want to record an investment decision. Ask me the following questions
one at a time, then compile the entry:

1. **Action**: Buy / Sell / Add / Trim / Hold (and size if applicable)
2. **Date**: When is this decision being made?
3. **Price**: Current share price at time of decision
4. **Thesis summary**: In 2-3 sentences, why am I taking this action?
5. **Key assumptions**: What 2-3 things must go right for this to work?
6. **What would change my mind**: Specific, measurable conditions that
   would cause me to reverse this decision
7. **Thesis status**: Is the original investment thesis intact, evolving,
   or broken?
8. **Confidence level**: High / Medium / Low - and why?

Format as a dated entry using this template:

## YYYY-MM-DD - [ACTION] @ $[PRICE]

**Thesis:** [summary]

**Key assumptions:**
- [assumption 1]
- [assumption 2]
- [assumption 3]

**I'm wrong if:** [conditions]

**Thesis status:** [Intact / Evolving / Broken]
**Confidence:** [High / Medium / Low] - [reason]

Before each earnings update, re-read your decision log. Compare what you expected to what actually happened. That's where you build real calibration.

Copy the output into your Obsidian vault as Decision Log.md. (Claude Code: appends to journal/decision-log.md automatically.)

Putting It Together

First-Time Analysis

1. Set up your environment - create a Project (claude.ai) or folder structure (Claude Code), upload filings
2. Step 1 - Business Overview - understand what you own
3. Step 2 - Industry Scan - understand the playing field
4. Step 3 - Financial Deep Dive - read the scorecard
5. Step 4 - Moat Analysis - assess defensibility
6. Step 5 - Management Review - evaluate the operators
7. Step 6 - Valuation Check - sanity-check the price
8. Step 7 - Bear Case - find the strongest counter-argument
9. Step 9 - Decision Journal - record your conclusion (Step 8 is for quarterly earnings updates - skip it for now)

Ongoing Quarterly Coverage

1. Download new filings (10-Q, earnings transcript)
2. Upload to your Project or drop into filings/earnings/
3. Step 8 - Earnings Update - analyze the quarter
4. Re-run any outdated memos (e.g., Valuation Check after major price moves)
5. Step 9 - Decision Journal - record any changes to position or conviction

The system gets more useful the longer you run it. Each quarter adds context. After 4-8 quarters on the same company, you have a documented record of how management communicates, how the thesis has evolved, and how reliable your own judgment has been. Link your memos in Obsidian ([[Business Overview]] to [[Moat Analysis]] to [[Bear Case]]) and over time your vault becomes a searchable record of everything you've thought through, on every company you've followed. The edge isn't better information. It's better organized information.

Maintenance Tips

• Keep filings organized - consistent naming (10-K_2024.pdf, Q3_2025_transcript.pdf) helps Claude reference them accurately
• Don't delete old memos - they form the historical record; update by creating new versions rather than overwriting
• Revisit the bear case annually - even if nothing has changed, forcing yourself to update it keeps you honest
• Claude Code users: git commit after each analysis session to create timestamped snapshots of your research state

Why This Actually Works

The system works for reasons that have nothing to do with AI being "smart."

Structure forces discipline. Most investors skip straight to "is it cheap?" The 9-step framework makes you work through the business, the industry, and the risks before you touch a valuation. That sequence matters more than any individual prompt.

The journal separates analysis from emotion. When a stock drops 15% after earnings, your decision log tells you whether the thesis is broken or the market is overreacting. Without it, you're guessing under pressure - which is when most people make their worst decisions.

Running this for eight quarters on the same company gives you something no screener can replicate: a documented record of how management communicates, whether the thesis has held up, and how reliable your own judgment has been. Norway's sovereign wealth fund didn't adopt Claude because it's trendy. They did it because structured workflows produce better research.