Claude Code 101 Best Practices

These are the onboarding habits that get a new developer productive with Claude Code fastest, drawn from the fundamentals covered across this section.

How to Use This List

• Work through the groups roughly in order for a brand-new developer; each builds on the last.
• Treat the first-week items as non-negotiable; treat the ongoing habits as things to revisit periodically.
• Use this as an onboarding checklist for a team, not just an individual reference.

A - Get the Mental Model Right First

• Understand it's an agent, not autocomplete, before writing a single request. A new developer who treats Claude Code like line-completion will under-use it and get confused by its multi-step behavior.
• Learn the read, plan, act, verify loop before relying on it for real work. Knowing that one instruction can trigger multiple file edits and shell commands sets the right expectations from the start.
• Frame requests as outcomes, not step-by-step instructions. Describing the goal lets the agent plan the steps; over-specifying steps defeats the purpose of delegating the task.

B - Install and Connect Deliberately

• Verify the install with claude --version before doing anything else. A broken PATH after install is a common early snag that's easy to catch immediately.
• Always launch from inside the actual project directory. Claude Code reads the current working directory as the project root, so launching from the wrong place gives it the wrong context.
• Start a new project with an orientation request, not a complex one. Asking it to summarize the project first confirms it read things correctly before you hand off real work.

C - Choose Permission Mode With Intent

• Default new developers to ask-every-time until they've built trust in the tool. Seeing each proposed action before it happens is the fastest way to learn how Claude Code plans in a specific codebase.
• Move to auto-accept only for well-scoped, low-risk, or already-reviewed tasks. Auto-accept trades a per-step checkpoint for speed, which only pays off once the plan is already trustworthy.
• Never use auto-accept for destructive operations or high-stakes code paths. Deletions, migrations, and authentication or payment code deserve ask-every-time regardless of experience level.

D - Use Plan Mode Before Trusting Execution on Unfamiliar Code

• Research first in Plan Mode for anything unfamiliar or ambiguous. Plan Mode restricts the agent to read-only investigation, so no real edit can happen before you've reviewed the plan.
• Actually read the proposed plan before switching to execution mode. A plan you didn't read gives you none of the safety benefit Plan Mode is meant to provide.
• Reserve straight-to-execution for small, well-understood tasks. Skipping Plan Mode is reasonable once a task is clearly low-risk and you've done something similar before.

E - Set Up the Terminal for Speed Early

• Add a shell alias for launching Claude Code in your first week. A one-word alias like alias cc="claude" is a small setup cost with an immediate daily payoff.
• Confirm font and rendering work cleanly with the REPL's interactive UI. Fixing rendering issues early avoids a confusing experience that has nothing to do with the agent itself.
• Adopt one terminal tab per active project once you're juggling more than one. This keeps sessions cleanly separated instead of losing track of which project a session is connected to.

F - Know When to Reach for Extended Thinking

• Reserve Extended Thinking for genuinely hard, multi-step problems. Architecture decisions, subtle bugs, and complex refactors benefit from deeper deliberation; small fixes usually don't.
• Expect a latency and cost trade-off when using it. More reasoning time before a plan is formed is not free, so it's worth reserving for problems that actually need the depth.
• Match the model to the task's difficulty. Haiku 4.5 for fast, low-stakes work, Sonnet 5 as the balanced default, Opus 4.8 for harder reasoning, and Fable 5 when you want reasoning depth to adjust automatically.

G - Build Ongoing Habits, Not Just First-Week Steps

• Commit your work before any auto-accept or execution-mode session with real consequences. Version control is the practical safety net underneath every permission mode.
• Revisit permission mode and Plan Mode choices as trust builds in a specific codebase. Comfort in one project doesn't automatically transfer to an unfamiliar one; keep evaluating per project, not just per developer.
• Periodically clear or compact long sessions. A session that's grown very long can slow down or dilute context, so /compact or /clear is worth using as a routine habit, not just a fix for a stuck session.

FAQs

Related

• The Claude Code Mental Model: Agent, Not Autocomplete - the foundational concept every other habit in this list builds on.
• Permission Modes Explained: Auto-Accept vs Ask-Every-Time - the full comparison behind the section C practices.
• Switching Between Plan Mode and Execution Mode - the full workflow behind the section D practices.
• Terminal, Shell and Hotkey Setup for Claude Code - the full reference behind the section E practices.
• Extended Thinking Mode: When Claude Code Reasons Longer - the full explainer behind the section F practices.

Stack versions: Written against the Claude model lineup current as of ~June 2026 - Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 (the default), and Claude Haiku 4.5. Model names, pricing, and product features move quickly - verify current specifics at platform.claude.com/docs before relying on them.