Working smart at scale

Subagents, anti-patterns, and the complete context toolkit

⏱ Est. ~7 min

01 · Read

You now have all the building blocks — CLAUDE.md for permanent rules, memory for learned patterns, sessions for continuity, context management for efficiency, specs for handoffs.

This lesson is about combining them effectively. The biggest productivity gains come from knowing what tool to use in what situation — and which anti-patterns to avoid.

Key points

Each context tool solves a specific problem
The key skill is picking the right tool for the situation
Anti-patterns waste context and make Claude less effective
Small workflow habits compound into big productivity gains

02 · Read

When you ask Claude to investigate something — "how does authentication work in this codebase?" — it reads files, follows imports, looks at patterns. All that investigation fills your context window.

Subagents solve this. A subagent runs in its own separate context window, does the investigation, and reports back a summary. Your main conversation stays clean and focused on implementation.

Think of it as sending an assistant to research and bring back a one-page summary, instead of reading every source yourself.

Key points

Subagents run in a separate context window
Investigation-heavy tasks are a good fit for subagents
Only the summary comes back to your main conversation
Keeps your working context clean for implementation

03 · Code example

CLAUDE.md is always loaded. Skills load on demand. Choose based on how often you need the instructions.

Put in CLAUDE.md (always loaded, every session)

- Build commands: npm run dev, npm test
- Code style: TypeScript strict, 2-space indent
- Architecture: where handlers, schemas, components live
- Project quirks: "purple is actually orange"

Put in a Skill (only loaded when needed)

- Detailed deployment procedure (10 steps)
- Database migration playbook
- API documentation for a specific service
- Complex debugging checklist

If you use it every session, put it in CLAUDE.md. If it's specialized knowledge for specific tasks, make it a skill. Skills cost zero context until activated — perfect for detailed procedures you only need occasionally.

04 · Read

Some common habits waste context and make Claude less effective.

Kitchen-sink session — implement a feature, then ask unrelated questions, then go back to the feature. Use /clear between unrelated tasks.

500-line CLAUDE.md — trying to document everything overwhelms Claude's attention. Keep it under 200 lines; move details into skills or rules files.

Correction loop — Claude gets it wrong, you correct, it's still wrong. After two corrections, /clear and rewrite the prompt based on what you learned.

Endless exploration — "investigate this codebase" with no scope. Use a subagent with a specific question instead.

Key points

Kitchen-sink session: /clear between unrelated tasks
Overly long CLAUDE.md: keep under 200 lines, move details to skills
Correction loop: after two failures, /clear and rewrite the prompt
Endless exploration: use a subagent with a specific question

05 · Read

Here's your toolkit, all in one place.

CLAUDE.md — permanent rules, loaded every session, survives compaction. Auto memory — notes Claude saves from experience. Sessions — name, resume, fork. Context window — monitor with /context, compact with /compact, clear with /clear. Specs — handoff documents for complex work. Checkpoints — Esc+Esc to rewind any change. Subagents — isolated context for investigation. Skills — on-demand knowledge without wasting context.

The difference between a beginner and an expert Claude Code user isn't typing speed — it's context management.

Key points

CLAUDE.md: permanent rules (project + user level)
Memory: patterns Claude learns
Sessions: name, resume, fork
Context: monitor, compact, clear
Specs: handoff documents for complex features
Checkpoints: rewind any change
Subagents: isolated investigation
The key meta skill is context management

06 · Quiz

You have a detailed 50-step deployment procedure. Where should it live?

CLAUDE.md, always loaded
A Claude Skill, only loaded during deployment
In conversation — tell Claude every time
As code comments

07 · Read

You've completed the entire Claude Code track. From Level 8 basic commands, to Level 9 skills, Level 10 MCP, Level 11 context management.

You don't just know how to use Claude Code — you know how to use it well. Manage context, write specs, structure work for clean handoffs, avoid the traps that slow most people down.

The next level shifts from Claude Code specifics to general junior engineer mode — debug, deploy, work like a pro. The context management skills you learned here will make every lesson after this more effective.

Key points

You've completed the full Claude Code track (Levels 8-11)
You understand the tools and the meta skill
Context management makes everything else more effective
Next up: junior engineer mode — debug, deploy, work like a pro

08 · Quiz

You're in a long Claude Code session. You just finished implementing a feature, and now you want to ask an unrelated database design question. What should you do first?

Just ask — Claude can handle multiple topics
Close the terminal and open a new one
Use /clear to reset context before starting the unrelated topic
Delete your CLAUDE.md file

09 · Fill in the blank

When you need Claude to investigate a large codebase without filling your main context window, delegate the work to a _____.

Other lessons in this chapter

⚠ The full interactive experience needs JavaScript. Please enable it and reload.

※ This is an independent Traditional Chinese teaching project — not an official Anthropic product. Claude™ is a trademark of Anthropic, PBC.