Claude Code - 47 PRO TIPS in 9 minutes

Summary of technological concepts & product pro tips (Claude Code)

CLI-based workflow (terminal-first)

• Claude Code is a CLI that works like other bash-based tools.
• Supports command-line arguments on startup.
• Can run headless mode (via -p).
• Designed to be chained/piped with other CLI tools.
• Can run multiple instances concurrently.
• Can spawn sub-agents / sub-tasks—described as effectively “launching instances of Claude Code.”

Images: manual and automated feedback loops

• On macOS, you can drag images into the terminal.
• Screenshot workflow:
  • Use Shift + Command + Control + 4
  • Then paste with Ctrl+V (explicitly not Command+V)
• Two main image-use patterns:
  1. Mockups: paste a UI mockup and ask Claude to build the interface.
  2. Iteration loop: generate something → open it → screenshot → feed back to Claude → iterate effectively.
• Automation via Puppeteer + MCP server
  • Set up Puppeteer MCP locally so Claude can open an app, take screenshots, and save them to a local directory.

MCP (Model Context Protocol) integration

• Claude Code can act as both an MCP server and an MCP client.
• It can be turned into an MCP server for other agents.
• Examples of MCP use cases:
  • Postgres MCP to connect directly to a database.
  • API-wrapper MCP servers for up-to-date documentation (e.g., mentioned Cloudflare).
• Even without MCP docs, Claude Code can fetch URLs:
  • Paste links, and Claude can retrieve content to inform code.
• Example “fetch rules from the web”:
  • For a game, instead of relying on training data, the host referenced Uno rules from unorules.com, and Claude implemented the game logic based on them.

claude.md (persistent prompt/config loaded every turn)

• A claude.md prompt is loaded with every request.
• It can include project-specific instructions such as:
  • bash commands to use
  • style/linting/testing guidelines
  • repo etiquette and workflows
• /init: creates claude.md by scanning and summarizing directory structure.
• Instructions can be updated by:
  • using # to add/append instruction content
  • setting a global claude.md in ~/.cloud/ (loaded across projects)
  • adding claude.md in subdirectories
• Recommendation: refactor the claude.md files often to prevent bloating, since it’s loaded every turn.
• Mentions Anthropic’s prompt optimizer tool to improve claude.md quality.

Slash commands as prompt templates

• Slash commands can be defined in a cloud/commands folder.
• They function as prompt templates and can interpolate command-line arguments at runtime.
• Examples mentioned conceptually include templates for:
  • GitHub issue solving
  • refactoring
  • linting
  • PR review

UI/productivity tips

• Use tab completion for files/directories (Claude Code performs better with more specific targets).
• Press Escape often:
  • If Claude goes off-track, stop it and ask Claude to undo its previous work.

Failure mode: “overbuilding” leading to breaking changes

• Biggest failure mode described: Claude Code gets ambitious, introduces breaking changes, and rollback is hard.
• Mitigations:
  • Use version control (e.g., Git).
  • Ask Claude to commit after major changes.
  • Let Claude generate commit messages.
  • Revert/rewind frequently:
    • clear conversation history
    • revert to a previous save point
    • retry with more specific instructions.

GitHub integration

• Install GitHub CLI for GitHub interactions.
• Alternatively, use a GitHub MCP server.
• Claude Code can:
  • create PRs
  • perform code reviews on PRs you generate.

Context management and cost controls

• Monitor the auto-compacting indicator and compact at sensible breakpoints (e.g., after committing).
• Consider clearing rather than compacting to keep memory fresh.
• Use external memory strategies to reduce context/token costs:
  • Scratchpads for planning
  • GitHub issues as an alternative planning/external memory mechanism
• For robust team cost tracking:
  • use OpenTelemetry support + DataDog dashboards (reference to Martin Amp’s blog post).

Pricing note / motivation

• Mentions a common complaint: high expense.
• Highlights Claude Max plans bundling Claude Code:
  • the host says they’re on the $100 plan
  • claims spending about $150 worth of tokens over ~3 days

Where to read more

• Encourages checking the original detailed post by Boris Churnney and other linked resources in the video description.

Main speakers / sources

• Speaker: Greg (developer, “my name is Greg”)
• Primary source referenced: Boris Churnney (creator of Claude Code at Anthropic)