Field notes from real-world AI-assisted programming. No theory, just patterns that work.
After 500+ hours programming with Claude Code, I kept running into the same problems and rediscovering the same solutions. This is the documentation I wish existed when I started.
What you'll find here:
- Configuration patterns that eliminate repetitive prompts
- Workflow habits that prevent context overflow
- Plugin ecosystem worth knowing about
- Git integration that actually saves time
What you won't find:
- Hype about AI replacing developers
- Generic productivity tips
- Untested recommendations
Claude forgets everything between sessions. You repeat yourself constantly.
Fix: Create CLAUDE.md files that load automatically:
~/.claude/CLAUDE.md → Global preferences
./CLAUDE.md → Project rules (commit this)
./CLAUDE.local.md → Personal overrides (.gitignore this)
# CLAUDE.md example
# Commands
npm run build
npm run typecheck
# Code style
ES modules only
Never use 'any' types
# Workflow
Typecheck after changes
# Never do
Uncomment disabled testsResult: Claude knows your stack from line one.
Writing the same scaffolding prompts over and over is a waste.
Fix: Install community plugins:
/plugin marketplace add edmundio/edmunds-claude-code
/plugin install edmunds-claude-codeWorth checking out:
| Repository | What it does |
|---|---|
| edmund-io/edmunds-claude-code | 14 slash commands + 11 agents. Next.js + Supabase |
| lodetomasi/agents-claude-code | 100 hyper-specialized agents with smart model selection |
| lodetomasi/Galadhrim-CodeSentinel | AI code review: 65+ issue types, security focus |
| obra/superpowers | 20+ battle-tested skills: TDD, debugging |
| hesreallyhim/awesome-claude-code | Curated list, good starting point |
Not every task needs the same cognitive effort. Simple fixes get over-engineered. Complex architecture gets under-analyzed.
Fix: Use explicit thinking keywords:
| Keyword | Use case |
|---|---|
think |
Quick tasks, simple fixes |
think hard |
Standard development |
think harder |
Complex logic, debugging |
ultrathink |
Architecture, system design |
Plan Mode: Hit Shift+Tab or type /plan. Claude analyzes before writing code.
Don't use "ultrathink" for everything. You pay in time and tokens.
Hit the 200K token limit mid-session and you lose everything. Permission dialogs interrupt your flow.
Fix:
| Command | When to use |
|---|---|
/clear |
After completing ANY task. Non-negotiable. |
/compact |
Mid-debugging to preserve context |
--dangerously-skip-permissions |
Eliminates permission dialogs (careful) |
Keyboard shortcuts:
Escape— Stop ClaudeEscape x2— Previous messagesCtrl+V— Paste imagesUp Arrow— Chat history
AI tools increase PR volume. Human reviewers focus on style nitpicks. Real bugs slip through.
Fix: Let Claude handle git operations:
/install-github-appThis auto-reviews every PR for logic errors and security issues.
What Claude handles well:
- Searching git history
- Identifying file ownership
- Writing commit messages
- Rebases and merge conflicts
- Branch management
Customize reviews in claude-code-review.yml:
direct_prompt: |
Review for bugs and security.
Only report actual problems.Four things that made the biggest difference:
- Always
/clearafter completing a task — Old context wastes tokens and confuses Claude - Match effort to complexity — "ultrathink" for architecture, regular prompts for CRUD
- CLAUDE.md is your onboarding doc — Time invested here pays off every session
- Use community agents — 100+ specialists already exist, no need to reinvent
"The biggest mindset shift was treating Claude Code not as a chat interface, but as a development environment that requires thoughtful setup and workflow design."
Found a pattern that works? Open a PR.
Something here is wrong or outdated? Open an issue.
MIT
claude code ai programming ai coding assistant anthropic llm development ai workflow developer productivity code generation ai agents prompt engineering
#ClaudeCode #AIProgramming #AIAssistant #DeveloperTools #CodingWithAI #Anthropic #LLM #DevProductivity #AIWorkflow #CodeAutomation