Skip to content

Exercise 2: Context Management

Learn to control what Claude reads, and what it forgets, so your sessions stay fast and focused.

Lab: Lab 01 overview

Theory: Context Window & @ References

What it is. Claude operates within a context window, a fixed amount of text it can "see" at once. Every file read, every command output, every message takes up space.

Why it matters. A sloppy prompt that reads 15 files eats 10x more context than a precise prompt with @ references pointing to 2 files.

Three tools:

  1. @ references, focus Claude on specific files: 
    Add validation to @src/routes/quizzes.py using @src/schemas/quiz.py
  2. /context, shows token usage breakdown
  3. /compact, compresses conversation history to reclaim space

Gotcha. /compact summarizes, it does not delete. Use /rewind to revisit specific earlier decisions.

Further reading: Common workflows docs

Hands-on

Part 1: Without @ References

  1. Fresh session — run /clear now to start clean. Ask: 
    Add input validation to POST /quizzes, the title must be at least 3 characters
and the topic must be non-empty.
  2. Watch which files Claude reads.
  3. Run /context and note the token usage.

Part 2: With @ References

  1. Fresh session (/clear or restart).
  2. Same task with references: 
    Add input validation to @src/routes/quizzes.py, the title must be at least 3
characters and the topic must be non-empty. The schema is in @src/schemas/quiz.py.
  3. Run /context, compare. Should be noticeably smaller.

Part 3: /compact During a Long Task

  1. Give a bigger task: 
    Validate all endpoints: title >= 3 chars, topic non-empty on POST /quizzes;
exactly 4 options and correct_index 0-3 on POST /quizzes/{id}/questions.
Also write tests for the new validation and update docs/api.md with the
validation rules.
  2. Wait for Claude to finish, then run /compact. Run /context and verify the history is compressed before telling Claude to continue.
  3. Tell Claude to continue. Verify it picks up where it left off.

Checkpoint

  • [x] Validation is in place
  • [x] You ran /context twice and observed the difference
  • [x] You used /compact and Claude continued successfully

← Exercise 1 · Wrap-up: theme, verify, recap →