Quest 1 · First Steps in the Forest
Install the agent, pair with it for the first time, and learn to read what it does instead of writing it yourself.
A terminal-native agent that reads files, edits them, runs tests, and explains itself — Anthropic's flagship CLI agent.
Install the CLI, authenticate, and pair on a real task in five steps.
The single skill that separates pretend vibe-coders from real ones. Read diffs faster than you can type code.
Plan, act, observe, adjust. Memorise the rhythm.
Quest 2 · Project Memory
Teach the agent your codebase's conventions once, and it will stop forgetting them every session.
Without a CLAUDE.md, you'll correct the agent on the same thing every conversation forever.
The five sections every solid CLAUDE.md has, with examples.
Author a real CLAUDE.md for a hypothetical project. The grader checks for the five required sections.
Watch the agent plan a multi-file change before touching anything. The plan is the artifact — review it like a PR description.
Quest 3 · Hooks & Smart Routing
Wire validation, logging, and cost-aware routing into the agent so it behaves the same way every session.
Custom slash commands turn your most-used 5-line prompts into one-word reflexes. Build three and you'll never look back.
Hooks let you intercept the agent before it acts or right after — for validation, logging, or hard-stopping unsafe operations.
Not every task needs Opus. Routing the boring 80% to a fast model cuts cost and latency without losing quality.
Author a SKILL.md following the agentskills.io standard. The grader checks the YAML front matter for name and description.
Quest 4 · The Enterprise Forest
Real teams shipping with Claude Code. Numbers, scope, and the capstone that proves you can do the same.
1,370 engineers, a 10,000-line Scala-to-Java migration, four days. What it took, what it didn't.
Investigating prod incidents got 80% faster. The pattern that worked at Ramp generalizes to any oncall rotation.
Multi-step quest. Pick a repo you own. Use Claude Code to ship a non-trivial, reviewable refactor end-to-end. Mark it complete when the PR is merged.