You Found Your Level. Here Is Everything Below It.

The rest of the guide

4. Context windowThe total amount of text a model can hold in its head at once, counted in tokens.
5. TokensThe chunks of text a model reads and bills by, roughly four characters each.
6. System promptThe standing instructions a model reads before it ever sees your message.
7. API keyA secret string that proves a request is yours and bills it to your account.
8. The CLIRunning the model in your terminal, inside a real folder, where it can read files, write files and run commands.
9. MCP serversA shared standard that lets any model use any tool.
10. CLAUDE.mdA plain markdown file in your project that the agent reads at the start of every session.
11. Slash commandsSaved prompts you can run by name.
12. Model routingSending each task to the cheapest model that can actually do it, instead of paying top rates for everything.
13. SubagentsA separate agent with its own fresh context that goes off, does a big messy job, and returns only the conclusion.
14. Prompt cachingTelling the provider to reuse context it has already processed, so you are not paying full price to send the same 50,000 tokens on every single call.
15. Git worktreesSeveral working copies of the same repository, in separate folders, on separate branches, at the same time.
16. SkillsA packaged capability: instructions plus any scripts, in a folder, that an agent loads only when a job matches it.
17. HooksYour own code, fired automatically by the agent's actions rather than by its judgement.
18. Harness engineeringDesigning everything around the model: which tools it gets, what limits it runs under, how it remembers, what it is allowed to touch, and how its loop terminates.
19. Background agentsAgents that run detached, carrying on with a long job while you get on with something else, and tell you when they are finished.
20. Headless runsRunning an agent with no chat interface at all.
21. Parallel agentsMany agents working simultaneously on independent slices of the same job, then merging what they produced.
22. Progressive disclosureLoading instructions and capabilities only at the moment they become relevant, instead of putting everything in front of the model up front.
23. Eval-driven loopsThe agent's output is graded against a fixed set of test cases on every run, and it keeps working until it passes rather than until it feels finished.
24. Orchestrator-workerOne agent plans and splits the job, several workers execute the pieces, and the planner merges and checks what came back.
25. UltrathinkExplicitly telling the model to spend far more reasoning effort before it answers, trading time and cost for depth.
26. Memory consolidationTurning raw session history into a small set of durable facts worth keeping, rather than storing everything or nothing.
27. Self-improving loopsThe agent updates its own instructions based on what went wrong, so the same mistake cannot happen a third time.
28. Tool poisoningHidden instructions buried inside a tool's description that your agent reads and obeys as if you had written them.
29. Self healing agent testsThe agent runs your test suite, reads the failures, fixes the code, and runs them again, looping until everything is green.
30. Agent run CI/CDAgents owning the pipeline itself: opening the pull request, watching the build, fixing what breaks, and merging when it is green.
31. Multi-repo orchestrationOne agent coordinating a single change across several repositories at once, keeping them consistent.
32. Agents managing agentsAn agent that plans work, dispatches other agents to do it, reviews what comes back against the acceptance criteria, and sends the failures back out.
33. Where To StartDo not try to climb this list in order.