Module 7: Some Claude Tips

MGMT 675: Generative AI for Finance

Kerry Back, Rice University

What We’ve Done

  • AI + code tool
    • financial analysis (mean-variance, dcf*)
    • deliverables (figures, excel, word, powerpoint)
    • small web apps (artifacts, pyodide)
  • Skills (reusable instructions)

Where We’re Going

M7: Some Tips (Today)

  • Final draft, not first draft
  • Enhance AI memory
  • Boris Cherny’s tips

M8: Web Scraping with AI

  • Scraping SEC EDGAR, news, transcripts
  • Static vs. dynamic pages
  • Magnolia example

M9: Build & Deploy Agents

  • The API: software connections to LLMs
  • Chatbots and web apps
  • Connecting tools
  • FastAPI and deployment

Where We’re Going (cont.)

M10: Corporate Deployment

  • Data security: on-prem or cloud
  • Guardrails and governance

M11: Retrieval Augmented Generation

  • How RAG works
  • Chunking, embedding, and vector search
  • Magnolia example*

M12: AI Sentiment Analysis

  • LLMs as sentiment classifiers
  • News and earnings call analysis
  • Trading signals from text

Final Draft, Not First Draft

Let AI Revise Its Own Work

Don’t review AI’s first draft. Instead:

  1. Ask AI to produce the work
  2. Ask it to critique itself and revise
  3. Repeat until no further revisions are needed
  4. Now review what it produced

You review a final draft instead of a first draft. This saves your time (though it burns tokens). The more consequential the work, the more revision cycles it deserves.

The /critique Skill

The /critique skill automates this. It spawns three parallel agents, each with a different lens, then synthesizes findings and applies revisions.

Correctness & Completeness

Are there factual errors, logical gaps, or missing information? Are claims supported and assumptions justified?

Clarity & Persuasiveness

Is the structure logical? Is anything confusing, redundant, or buried? Could the main message be stated more directly?

Devil’s Advocate

What are the strongest counterarguments? Where is the reasoning weakest? What would a skeptic challenge?

Download the critique skill and ask Claude to install it.

Enhance AI Memory

Saving Conversation Summaries

AI conversations have no memory of previous conversations — each new session starts from scratch. You can fix this by saving things to markdown (.md) files.

A summarize chat skill

  1. Download summarize-chat skill and ask Claude to install it.
  2. At the end of a chat, type /summarize-chat
  3. The summary is saved as a .md file in a chats/ folder. Each session, Claude reads the most recent summaries on startup.

Standard memory devices

  • CLAUDE.md — project-level instructions loaded every session
  • Skills (.md) — reusable task-specific instructions
  • Other files — you can always ask Claude to save any chat or part of a chat as a file

Markdown Quick Reference

Markdown is a plain-text format that uses simple symbols for formatting. AI reads and writes it natively.

# Heading
## Subheading

- Bullet point
- **Bold text**
- *Italic text*

1. Numbered list
2. Second item

> Blockquote

Markdown in AI Workflows

Where you’ll see it

  • CLAUDE.md — project instructions
  • SKILL.md — skill definitions
  • Session summaries — conversation logs
  • README.md — any GitHub repo

You can read any markdown file in Notepad, but you get a better presentation with a markdown renderer, for example, markdownlivepreview.com.

Claude Dispatch

Dispatch: Setup

Claude Dispatch lets you send a task from the Claude mobile app and have it run on your desktop — as either a Cowork or Code session, depending on the task.

Setup

  1. Update both apps first: Claude Desktop (Help → Check for Updates) and Claude mobile app (App Store / Google Play)
  2. Open Claude Desktop → Cowork tab
  3. Click Dispatch in the sidebar
  4. Scan the QR code from your phone
  5. Send a task — e.g., “fix the auth bug”

Dispatch: How It Works

Dispatch decides whether the task needs a Code session (dev work) or can be handled in Cowork (research, writing).

  • Your desktop must be awake and running Claude Desktop
  • Dev tasks → Code session spawned automatically
  • You get a push notification when done or when approval is needed
  • Full access to local files, MCP servers, and project config

Requires Pro or Max plan.

Boris Cherny’s Tips

Boris Cherny is the creator and Head of Claude Code at Anthropic. In a series of X posts over several weeks, he offered tips on how he and his group use Claude Code.

The tips were compiled by a fan and made available online here — dozens of tips organized into 10 parts covering workflows, memory, automation, and more.

A few highlights:

  • Run multiple tasks simultaneously instead of waiting for one to finish
  • Set up automatic quality checks (e.g., verify data after every change)
  • Write detailed CLAUDE.md files — better instructions mean less repetition

His tips can be installed as a Claude Code skill — paste the URL in Claude Code and ask it to create a skill from the content.

Claude Plugins

What Are Plugins?

A plugin bundles multiple Claude Code components into a single installable package:

Component What it does
Skills Slash commands & auto-triggered prompts
Agents Custom sub-agent definitions
Hooks Event handlers (e.g., lint on file write)
MCP servers External tool integrations
LSP servers Language intelligence (type checking, etc.)
Output styles Custom formatting for agent output

Installing and Sharing Plugins

Installing

  • /plugin install from a marketplace or git repo
  • Commands are namespaced: /my-plugin:hello
  • Versioned via plugin.json manifest

Why plugins?

  • Skills are great for personal, single-project use
  • Plugins are designed for team distribution and cross-project reuse
  • Available from the official marketplace, team marketplaces, or any git repo
  • Browse community plugins: anthropics/claude-plugins-official

The SaaSpocalypse

In early 2026, AI agent and plugin announcements triggered a historic selloff in enterprise software stocks — roughly ~$2 trillion in SaaS market cap evaporated between January and March. Wall Street dubbed it the SaaSpocalypse.

AI is attacking the interface and workflow layer, not the data layer:

Replacing the UI

Many SaaS products are forms, dashboards, and screens sitting on top of databases. AI agents let users skip navigating 20 screens and just state what they want in natural language.

Automating Workflows

Task tracking, status updates, if-then automations, data transfer between systems — exactly what agents do natively.

Collapsing Per-Seat Pricing

If 10 AI agents do the work of 100 human reps, you need 10 Salesforce seats, not 100. The revenue model collapses even if the product is still used.

What Survives?

AI isn’t replacing the database — it’s replacing the screens people use to interact with it, and the humans who sit behind those screens.

Defensible positions

  • Systems of record in regulated industries (finance, healthcare, law)
  • Platforms with accumulated proprietary data and feedback loops
  • Integration hubs managing connections across many systems

The market damage (as of March 2026)

  • IGV (software ETF) down ~25% YTD
  • Forward P/E compressed from ~34x to ~22x
  • Software traded at a discount to the S&P 500 on a forward P/E basis — a first
  • Salesforce -31%, Adobe -32%, Atlassian -35%

Sources: TechCrunch, CNBC, Bloomberg, SaaStr; data as of mid-March 2026.

Version Control with Git

Why Git Matters for AI Workflows

Git is like Track Changes for an entire folder — it saves every version so you can always go back.

Why it matters with Claude

  1. Safety net — it is much safer to let Claude edit your files when you can easily roll back to any previous version
  2. Claude does the work — Claude can commit changes, review past commits, and recover old versions for you — no need to learn git commands yourself

Getting started

Ask Claude to:

  1. Set up Git — ask Claude; it will install and configure it if needed
  2. Initialize a folder as a git repo
  3. Commit changes whenever you reach a good stopping point

That’s it. Claude handles the commands. You just say “commit this” or “show me what changed”.

Back Up to GitHub

Git keeps snapshots locally. For an external backup, push to GitHub (free account).

Setup (one time)

  1. Create a free account at github.com
  2. Ask Claude to install the GitHub CLI (gh)
  3. Run gh auth login — it opens your browser to sign in (no tokens or passwords to manage)
  4. Ask Claude to create a GitHub repo and push your folder to it

Ongoing

  • After committing locally, say “push to GitHub”
  • Your work is now backed up in the cloud
  • You can also browse history, compare versions, and restore files from the GitHub website