How We Use AI at Work

PyRVA Meeting — September 2025

Presenter: Chris Way

Topic: Claude, specifically Claude Code and how it integrates with Python and software projects.

---

Getting Started with Claude Code

• Install Claude Code:

  npm install -g @anthropic-ai/claude-code

• On first run, Claude will:

  • Ask for permissions (whether you trust the code you’ve written).
  • Run as a text-based command line interface.

• Setup steps:

  • You select a directory where Claude operates.
  • The first step is creating a CLAUDE.md file with instructions for Claude.
  • Running /init generates an initial CLAUDE.md based on the project.

---

Working with CLAUDE.md

• Claude writes its findings in Markdown so you can review them.

• Chris emphasized:

  • Keep CLAUDE.md small → large files degrade quality.
  • Borrowed ideas from Kent Beck’s experiments with AI, including using a very long system prompt.
  • Differences between CLAUDE.md and plan.md can help separate context from planning information.

---

Context Window

• Value of Claude often depends on the size of the context window.

• Command /context shows context usage.

• Problems:

  • When context is compacted, answers get worse.
  • Summarization reduces quality.

• Recommendation: keep the context window as small as possible.

---

Modes and Commands

• Modes:

  • Shift+Tab opens plan mode and other modes.
  • Acceptance modes: auto vs. one at a time.

• Useful commands:

  • /context → show context window usage.
  • /cost → shows cost estimates. (Claude Pro has extra costs.)
  • Restarting: just type go and Claude reloads from CLAUDE.md.

• Permissions: stored in .claude/ so they persist between sessions.

---

Workflow and Refactoring

• Chris used Claude to perform a massive refactor:

  • Claude continuously ran tests.
  • It successfully merged a refactor with pending code changes.

• Output variability:

  • Test names and plans can be surprising.
  • Quality varies significantly depending on the input.

• Predictability:

  • Run times are inconsistent — Chris sometimes lets Claude run unattended.

---

Git & Parallel Work

• Chris uses git worktree to manage multiple Claude sessions:

  mkdir ../movie_alt
  git worktree add ../lost_password -b feature/lost-password

• This enables parallel work on multiple branches.

• Ultimately, Chris must handle the merge himself.

• Architecture tip:

  • Chris structures code in a “vertical slice” style.
  • Each slice contains changes in isolation, making it easier for Claude to manage.

---

Commands, Agents, and Extensions

• Claude Commands:

  • Stored in .claude/commands/.
  • Example: prime.md for reusable macros.
  • Helpful when repeating the same instructions.
  • Growing library of community commands (e.g., from the humanlayer repo).

• Agents & Subagents:

  • Stored locally in .claude/ or system-wide.
  • Can call each other and even span multiple checkouts.
  • Useful for extending Claude’s capabilities.

• MCP Servers:

  • Chris hasn’t used them yet.
  • Designed to extend LLMs in specific, structured ways.

---

Observations & Takeaways

• Claude serializes an understanding of the codebase into CLAUDE.md, enabling a stronger starting point for future changes.
• Claude can read git history and project files to improve its reasoning.
• Best results come when Claude is given well-structured code and manageable context.
• Chris’s opinion: Claude performs much better when it “studies” the codebase first.

Questions

Are there any alternatives to Claude code: Google has a text base CLI like Claude code called Gemini.