Claude Code is the developer-focused CLI version of Claude. Same brain as Cowork, different surface: a terminal, your codebase, your git history. The non-obvious part is that "writing code" is only about 30% of what it does. The other 70% is everything around the code: planning, reviewing, testing, deploying, remembering. That 70% is where most users plateau.
The list below is the unlock. Each row: the command or pattern, what it does, and the easy-to-miss tip that turns it from a feature into leverage.
The 10 patterns
CLAUDE.md
Project-specific rulebook Claude reads before every task. Mission, conventions, what to never do.
Pro tipRun
/init first to auto-generate it from your codebase. Edit afterwards to nail your conventions.Easy to missKeep it short. Long, vague rule files get ignored. Tight, specific instructions get followed.
/skills
Reusable markdown templates for standardised workflows. Code review, PR descriptions, deployment checklists.
Pro tipEncode your team's checklists. Get consistent output across every engineer for free.
Easy to missOne skill per workflow. Do not merge unrelated skills into a megadocument.
/mcp
Connects Claude to your databases, internal APIs, terminal logs and external services.
Pro tipWire MCP to your staging database for read-only queries during debugging. Logs flow straight into the chat.
Easy to missMCP replaces 90% of "let me copy-paste that into Claude". Stop copy-pasting.
subagents
Spawns parallel workers for isolated tasks. Each runs in its own context.
Pro tipSend a subagent to write unit tests while you focus on the feature logic. Parallel beats sequential.
Easy to missSubagents cannot see your main thread. Brief them properly or they guess.
/plan
Forces an architecture outline before any code is written. Read-only mode until you approve.
Pro tipCritique the plan first. Then say "implement it" only when you are happy.
Easy to missFor tasks longer than 10 minutes, plan mode is faster than going straight to code. Less rewriting later.
/compact
Compresses conversation history so the context window keeps working without losing key decisions.
Pro tipUse when Claude feels sluggish or starts forgetting earlier decisions.
Easy to missMega-sessions cause context pollution. Old, wrong context bleeds into new tasks.
autodream
A "sleep cycle" pattern that organises and refines project memory between sessions.
Pro tipRun it overnight after a heavy day. Wake up to a cleaner CLAUDE.md and tighter Skills.
Easy to missThis is a community pattern, not a built-in command. Worth borrowing.
/ralph-loop
Autonomous iteration loop for self-correction and testing. Useful for "fix until tests pass" tasks.
Pro tipPair with TDD. Always set safe iteration limits.
Easy to missWithout limits, ralph-loops can run themselves into the ground. Cap at 5 to 10 iterations.
hooks
Pre and post commands that run automatically before or after each tool use. Gates dangerous actions, runs linters, formats code.
Pro tipAdd a post-edit hook that runs your linter and formatter. Code style enforced for free.
Easy to missA pre-bash hook can block destructive commands like
rm -rf or git push --force. Belt and braces.CLAUDE.local.md
Personal overrides on top of the team's CLAUDE.md. Never committed.
Pro tipPut your personal preferences here (testing framework, debug verbosity, tone) without polluting the shared rules.
Easy to missAdd it to
.gitignore the same day you create it. Otherwise your private notes ship with the next commit.
Worth knowing: none of these patterns is hard to set up on its own. The leverage compounds when you stack them. CLAUDE.md plus Skills plus MCP plus Hooks turns Claude Code from "smart autocomplete" into a teammate that ships work overnight while you sleep.
Bottom line
Most people are still using AI like a chatbot. They prompt, they read, they copy-paste, they prompt again. The patterns above flip the relationship. You set up the system once. The system runs every time you open Claude Code. The work that used to take a morning runs in the background while you focus on the things only you can do.
Adapted from a Claude Code cheat sheet circulating in dev communities, with two additions (hooks and CLAUDE.local.md) and tips drawn from how engineers in the AIReady Leaders cohort actually use them.