Show Notes
Claude Code has levels, and this video breaks down how to use them effectively—from core CLI tools to advanced agents and subagents. Tailor your setup, don’t just copy-paste, and build workflows that actually fit your project.
Levels and philosophy
- There are layers to Claude Code and Claude Code. Start with the prompt and customize from there.
- Don’t rely on “Claude awesome lists” or copied agents. The closer you are to the prompt, the better the outcome.
- Use prompts and wrappers as a base, then tailor them to your exact use case and stack.
Core tools and environment
- Claude Code can reach out to any command-line tool you use (yt-dlp, ffmpeg, Whisper, etc.). This unlocks powerful automation.
- Hotkeys and a streamlined setup speed up common tasks.
Key concepts you’ll work with:
- Root dot folder (~/.cloud or similar) for your dotfiles
- cloud.md and MCP settings at the user level
- A base Claude configuration that can be reused across projects
- A minimal, consistent default setup that supports real-world tooling decisions (e.g., npm vs bun)
Quick commands and aliases
- Common aliases (examples):
- CC: default Claude command (bypasses some prompts/permissions)
- CCI: quick access to a project/root workflow
- CCD: run tests and start a headless Claude for fixes
- CCT: tests + headless Claude output
- Headless mode: useful for quick, small tasks; otherwise keep an eye on outputs
- ZSH aliases tied to cloud commands for fast access (e.g., CC, CCI, CCD, CCT)
Documentation and local docs mirror
- Use a local docs repository to keep Claude docs up to date:
- Docs mirror updated every 3 hours (faster than live fetch, works offline)
- Example: local clone of Claude Docs; you can run for/docs to fetch the latest docs
- Benefits: always know if docs changed and quickly adapt your prompts/agents
Projects, agents, and MCPs
- User level vs. project level configurations
- A dotcloud/.cloud setup lets you add domain-specific agents (e.g., a Convex MCP agent)
- Tailor agents by editing YAML definitions locally before deploying
- Don’t copy-paste agents wholesale; edit to reflect your stack and MCP tools
Tailoring prompts and agent templates
- Start from a solid base prompt, then customize for your project
- Extract useful bits from public cloud code prompts, then adapt to your needs
- Example workflow: create a TypeScript engineer agent, adjust for your ESLint/Convex setup, and test
- Editing YAML prompts locally often yields better results than using off-the-shelf ones
Sub-agents vs slash commands
- Sub-agents: act as teammates with a specific role; great for recurring responsibilities (e.g., code reviews, planning)
- Slash commands: quick, reusable actions for frequent tasks
- Background agents: best after you have solid core processes in place; use for autonomous, headless tasks
- Planning mode (Shift+Tab) helps formalize work into phases before coding
Practical workflows and examples
- A TypeScript engineer agent example helped align ESLint/configs with Convex and generated types
- Emphasize planning before coding; use 03 for planning and Opus for execution when appropriate
- Use tooling like GitHub Actions for automation (auto review workflows beat the Cursor bugbot in cost and effectiveness)
Best practices and guardrails
- Enforce guardrails with build steps, linting, and tests before outputs exit your machine
- ESLint generally outperforms Biome for LLM-driven code workflows
- When enabling dangerous skip-permissions mode, do so only after trust is established
- Use aliases and resume functionality to manage conversation context and reuse previous threads
Resources and prompts to check out
- Claude Code prompts and workflows (Kieran’s prompts and related writeups)
- Integration ideas: terminal, VS Code extension, and MCP toolsets
- Prominent voices and docs:
- Eric Bu's tweets for Claude Code documentation updates
- Peter Steinberger’s take on “Claude Code in no prompt mode”
- joinvi.com/blog for Claude Code workflows and prompts
- Various community prompts and workbooks for velocity coding and planning with Opus/03
Final takeaway
Claude Code is more than just a coding assistant. It’s an adaptable agent framework that can manage code, CLI tools, notes, and even project planning. Start with a strong baseline, tailor everything to your stack, and gradually add subagents and MCP integrations as your maturity grows.
Links
- Claude Code Documentation (official docs reference)
- Convex MCP Server (Convex ecosystem docs)
- Every.to Claude Code Camp (Claude Code prompts and workflows)
- Peter Steinberger's Claude Code Blog (Claude Code in no prompt mode)
- Eric Buess on X/Twitter (Claude Code docs and updates)
- GitHub Actions (auto review and other workflows)
- yt-dlp, FFmpeg, OpenAI Whisper (essential CLI tools)
- Vibe with AI Discord (practical engineering discussion and collaboration)
If you have workflows I didn’t cover or tips you’re using, drop a comment. For more hands-on collaboration, check out the VI community and subscribe for future deep-dives.