Tech • AI • Robotics • Game
Claude Code: the 14 tools to master (complete course)
7/10TL;DR
Claude Code is most effective when used as an agentic coding system with strict permissions, controlled context, verification steps and specialized extensions rather than as a simple chat interface.
KEY POINTS
From chat tool to agentic loop
Claude Code works through an agentic loop: it gathers context, edits files, runs checks and iterates until a task is resolved. In practice, a request such as fixing a broken login button or failing cart tests can trigger file searches, code changes and test execution automatically. The model is only one part of the system; performance also depends on the surrounding harness, including context, instructions, tools, permissions, verification and the loop itself.
Permissions are the first safety layer
Durable restrictions should be set in .cloudsettings.json, not only in CLAUDE.md. A written instruction such as “never read.env” can be ignored, but a permission rule can block access deterministically. The hierarchy matters: deny overrides allow and ask, while ask can still force confirmation before actions like git commit.
Execution modes vary sharply in risk
Manual mode allows reads but requires validation for most other actions. Accepted edits permits file modifications, plan explores without changing files, and auto relies on a classifier to screen actions. Even then, roughly 1 in 6 dangerous actions may still pass, making bypass mode especially risky outside tightly controlled environments.
Git remains the real rollback system
Built-in checkpoints can rewind file edits within a session through commands such as /rewind or double Escape, but they do not replace Git. They may fail to recover destructive shell commands like rm -rf or work done by sub-agents. Permanent, named and shareable history still depends on commits and branches.
Context can help or poison results
Context is the model’s working memory, and once it gets too full accuracy can drop. A practical threshold is around 50% to 70% usage, well before near-saturation levels such as 95.5% of a 1 million token window. Commands like /clear and /compact are essential because old instructions can bleed into new work, such as a prior request for “large blue cards” influencing the design of an unrelated invoices page.
Session rules belong in CLAUDE.md or AGENTS.md
Files such as CLAUDE.md or AGENTS.md preload guidance at the start of each session, making them useful for habits like always using pnpm instead of npm. But they are not deterministic enforcement tools, and overloading them reduces usefulness. The recommended approach is to keep them short, ideally under 200 lines, focused on stable project conventions rather than vague advice.
Verification must produce evidence
Passing compilation is not enough to prove that a feature works. Effective verification includes automated tests, browser checks, screenshots and visible before-and-after results. Built-in verification flows can launch the app, click through actions and identify failures such as an “add to cart” button that leaves the counter unchanged.
Model choice and effort should match the task
Light models such as Haiku 5.5 suit repetitive or simple work, while Sonnet covers most day-to-day coding. Opus and Fable 5.1 are better reserved for architecture, multi-step reasoning and ambiguous long problems because they are slower and more expensive. The same logic applies to reasoning effort: a typo fix does not justify maximum effort on a top-tier model.
Skills, MCP and hooks extend capabilities
Skills package repeatable know-how, such as drafting product pages or running a standard verification workflow, and are loaded only when relevant. MCP acts like a universal connector, linking the coding agent to tools such as GitHub, Slack, databases or Figma. Hooks differ from guidance files because they are deterministic: if a formatter must run after every edit, a post-tool-use hook can enforce it every time.
Sub-agents and parallel sessions help scale work
Sub-agents isolate heavy investigation in their own context and return only a summary, keeping the main session lighter. Parallel sessions can speed delivery across tasks like contact, FAQ and legal pages, but they also increase consumption and the risk of conflicts. Using separate Git worktrees gives each session its own branch and avoids multiple agents overwriting the same files.
CONCLUSION
The main challenge with Claude Code is not prompting harder but building the right operating framework around it. Teams that control permissions, manage context, verify outputs and use extensions selectively can gain speed without surrendering safety.
Ask a question