Brooker Belcourt: Using Claude Code as a Financial Research Agent

Key Insights

• Extended compute time unlocks transformative research capabilities: While web-based AI interfaces are limited to 20-30 minutes of compute time, Claude Code can run for hours on your local machine, creating a massive capability gap that enables fundamentally different types of analysis (00:08:42).

• Natural language dashboards replace static PDF reports: Instead of spending 5 hours creating static Word or PowerPoint documents, financial analysts can now generate interactive, live dashboards with one command using Claude Code, MCPs, and Streamlit (00:07:23).

• Version-controlled prompts in GitHub are becoming valuable intellectual property: As prompts grow beyond the 8,000 character limit of web interfaces, storing them as Claude plugins in GitHub repos enables version control, sharing, and building of sophisticated research agents with custom investment philosophies (00:04:41).

• The software development layer has collapsed into the query layer: Building custom dashboards, accessing multiple data sources, and producing analysis now happen in a single natural language command rather than requiring separate development, data gathering, and reporting phases (00:08:05).

• LLMs need anti-consensus training: Large language models are naturally consensus-oriented, so financial analysts must explicitly program their research agents with contrarian investment principles like “trajectory matters” and “favor accelerating over decelerating businesses” (00:05:33).

Summary

Brooker Belcourt brings a decade of hedge fund experience and a decade in FinTech startups to demonstrate how Claude Code transforms financial research. Most recently running the finance vertical at Perplexity, he now consults with investment firms on AI adoption. Rather than focusing on Claude’s app-building capabilities, he showcases its power as a research assistant that can create sophisticated, interactive financial dashboards through natural language commands.

His demonstration centers on building an earnings preview for Meta using Claude Code’s extended compute time, MCP integrations (particularly Dilupa for institutional financial data), and Streamlit for interactive visualization. The workflow combines GitHub-stored prompts that encode investment philosophy, local file access to transcripts and notes, and API calls to financial data sources—all orchestrated through simple slash commands that generate custom dashboards in 20 minutes instead of the 5 hours traditional methods would require.

Main Topics

Background and Core Concept

Brooker positions himself as a non-engineer with deep financial expertise who has discovered a more powerful way to use Claude Code. His approach treats it as a research assistant rather than an app builder.

“I found it an incredibly powerful research assistant.” (00:01:00)

The key distinction is that Claude Code can access MCPs, all files on your computer, and run for extended periods—creating outputs that are interactive dashboards rather than static PDFs or emails.

Live Demo: Earnings Preview Dashboard

Brooker demonstrates creating an earnings preview for Meta with a single slash command. The dashboard includes:

• Beat/miss track record (12 quarters and 4 quarters)
• Revenue guidance analysis
• Guidance range versus actual results
• Interactive charts and tabs
• Mini Excel model with specific line items

“With one command, you can actually create this dashboard.” (00:01:46)

The dashboard is flexible enough to adapt to each company’s unique guidance metrics. Meta guides on revenue, expenses, CapEx, and tax rate, while other companies might only guide to revenue or EPS.

Timestamp for full workflow: 00:01:24 - 00:04:24

GitHub-Based Prompt Management

As prompts grow beyond 8,000 characters, Brooker stores them as Claude plugins in GitHub repositories with version control. The structure includes:

• Skills: Investment philosophy, data source priorities, preferred analysis methods
• Data source configuration: Exact directory paths for transcripts and notes
• Investment principles: “I like all the goblies. I like trajectory. I like accelerating businesses, not decelerating businesses.” (00:05:25)

This approach solves the prompt length limitation and creates reusable, shareable research frameworks.

“I’m turning Claude into this research agent that is running code to produce these custom dashboards.” (00:06:01)

Timestamp for GitHub walkthrough: 00:04:41 - 00:06:01

Natural Language Dashboard Specifications

Rather than specifying chart types, formatting, or layout details, Brooker writes in natural language what analysis he wants:

• “Beat miss track record”
• “12 quarters of beat rate, four quarters of beat rate”
• “Compare revenue guidance”

Claude Code determines whether to use bar charts, line charts, or other visualizations automatically.

“I just I’m saying the stuff in all this natural language. And it’s doing all the work is to be like, well, should this be a bar chart? Should this be a line chart? How should I present this? I don’t have to talk all about that.” (00:06:23)

Timestamp: 00:06:09 - 00:06:50

The Compute Time Advantage

A critical chart shown during the presentation illustrates the gap between web-based AI interfaces (capped at 20-30 minutes) and Claude Code’s extended compute capabilities. This gap is what Brooker identifies as the reason for Claude Code’s surging attention.

“Ever since like the start of 2025, like we’ve expanded to offer LLMs like significantly more time to process and create answers. But the web apps have kind of stuck at 20 minutes. And so there’s this huge gap.” (00:08:59)

He emphasizes letting analysis run much longer on local machines rather than being constrained by web interface timeouts.

Timestamp: 00:08:42 - 00:09:22

Impact on Financial Analysis Workflow

When asked how long this would have taken as a hedge fund analyst earlier in his career, Brooker reveals the transformation:

“It’s crazy to imagine because this work for an earnings preview is like five hours of work compiling all these different sources together into one dashboard. And my dashboard would be like a Word document or a PowerPoint presentation. It wouldn’t be interactive. It wouldn’t be live.” (00:07:30)

The comparison highlights not just time savings but a fundamental shift in output quality.

“It’s like you’re building a Bloomberg or a dashboard system, like a Koi Fin for your entire process. And so now it’s the software development is combined with the query and the actual output, which is really cool.” (00:07:48)

Timestamp: 00:07:06 - 00:08:05

Actionable Details

Tools and Data Sources Mentioned

Institutional Data: - Dilupa MCP: Used throughout the demo for institutional-grade financial data, integrates directly with Claude Code

Retail/Accessible Data: - Perplexity Finance: Can download transcripts and all financials for free - Navigate to a company page - Go to “Earnings” tab for transcripts - Go to “Financials” tab to download all financial data - Also includes research reports

• Fiscal AI: API access for financial data, can be integrated via Claude Code skills

Development Stack: - Claude Code: Command-line interface with extended compute time - GitHub: For version-controlling prompts and Claude plugins - Streamlit: For rendering interactive dashboards on localhost - MCPs (Model Context Protocol): For connecting to data sources

Transparency/Publishing: - Autopilot: Platform where Brooker publishes all trades and ideas (same people who created the Pelosi tracker) - X (Twitter): Primary distribution for research and ideas

Workflow Steps

1. Create a Claude plugin in a GitHub repo
2. Define skills and investment philosophy in the plugin
3. Specify data source locations (e.g., directory paths for transcripts)
4. Use slash commands to invoke pre-built prompts
5. Claude Code accesses MCPs, local files, and APIs
6. Generates Streamlit dashboard running on localhost
7. Dashboard is interactive with tabs, charts, and live data

Example Companies/Use Cases

• Meta: Earnings preview demo (reports next week from recording date)
• ICE (Intercontinental Exchange): Mentioned as favorite company, “monopoly of exchanges”

Quotes Worth Saving

“I found it an incredibly powerful research assistant.” (00:01:00) — On Claude Code’s primary value for financial analysis

“It’s like you’re building a Bloomberg or a dashboard system, like a Koi Fin for your entire process. And so now it’s the software development is combined with the query and the actual output, which is really cool.” (00:07:48) — Describing how Claude Code collapses traditional workflow layers

“I find LLMs are just very consensus in the way they look at ideas. So I’m turning Claude into this research agent that is running code to produce these custom dashboards.” (00:05:40) — On the need to encode contrarian investment philosophy into AI research agents

“Ever since like the start of 2025, like we’ve expanded to offer LLMs like significantly more time to process and create answers. But the web apps have kind of stuck at 20 minutes. And so there’s this huge gap.” (00:08:59) — Explaining why Claude Code is gaining significant attention

“I think it’s really important to start building this IP of these prompts. I think they’re so valuable and GitHub is a great place to store that.” (00:12:12) — On treating version-controlled prompts as intellectual property