← Back to all playbooks

Brooker

BROOKER BELCOURT: Financial research with Claude
Start: 07:29:36
Sketch note for Brooker's session
TL;DR

Bypass web UI constraints by using Claude Code in the terminal to orchestrate GitHub-versioned prompts, local transcripts, and financial MCPs into interactive Streamlit dashboards.

Key Insights

GitHub-Versioned Prompting for Complex Workflows

07:34:18

Store long-form prompts and 'skills' in a GitHub repository as plugins to bypass the 8,000-character web UI limit and maintain version control over investment philosophies.

Local Compute vs. Web UI Caps

07:38:25

Claude Code in the terminal allows for significantly longer compute times (exceeding the ~20-minute web app cap), enabling the LLM to process deeper datasets and generate complex local applications.

Anti-Consensus Prompting Strategy

07:35:25

Inject specific, opinionated investment principles (e.g., 'prefer accelerating trajectories over decelerating ones') into the system prompt to prevent the LLM from defaulting to generic, consensus-driven financial summaries.

Interactive Streamlit Outputs

07:32:41

Instead of static PDFs, instruct Claude to generate local Streamlit apps that allow for interactive data manipulation, tabbed guidance analysis, and on-the-fly Excel mini-models.

Systems

Automated Earnings Preview Dashboard

  1. Connect Claude Code to a financial data MCP (e.g., DeLupa).
  2. Define a local directory path for company transcripts and financial files.
  3. Create a slash command in Claude Code that calls a versioned prompt from GitHub.
  4. Instruct Claude to generate a Streamlit application script based on the data.
  5. Run the Streamlit app on localhost to visualize beat/miss track records and guidance analysis.
Tools: Claude Code, Streamlit, GitHub, DeLupa MCP

Tools

Claude CodeStreamlitGitHubDeLupa MCPPerplexity FinanceFiscal AI