Four Claude Code settings I change before letting it touch a project

At a glance:

• Add a CLAUDE.md file to give Claude Code permanent project context.
• Pre‑configure .claude/settings.json to skip repetitive permission prompts.
• Choose the appropriate model (Opus, Sonnet, Haiku) and mode (plan, acceptEdits, auto) before writing code.

Why a CLAUDE.md file matters

When you drop Claude Code into a fresh repository without any guidance, it behaves like a curious intern—making educated guesses based on the few files it sees. The author recommends creating a CLAUDE.md file in the project root. This plain‑text markdown file is read at the start of every session and can contain the tech stack, folder layout, naming conventions, and any directories the model should avoid. By doing this, Claude Code carries the same brief you would give a human teammate, eliminating the need to repeat context in each prompt.

Claude Code even ships with an /init command that scans the repository and generates a starter CLAUDE.md. The autogenerated version is a good baseline, but you should edit it whenever the model makes a mistake you don’t want to repeat. The extra five minutes spent writing or tweaking this file pays off in reduced back‑and‑forth and fewer mis‑generated snippets.

Configuring permissions up front

By default Claude Code pauses before every potentially destructive action—npm installs, git status checks, file edits, web searches, etc. This safety net can become a nuisance when you find yourself clicking “yes” dozens of times in a single session. The tool stores allow/ask/deny rules in a .claude/settings.json file, which you can edit directly or manage through the /permissions command.

The author’s workflow is to review the typical actions for the project (e.g., allowing npm install, permitting git status, denying arbitrary rm -rf commands) and set those preferences before the first prompt. This avoids the temptation to flip the risky --dangerously-skip-permissions flag, which would hand Claude unrestricted access to the host machine. A balanced permissions file keeps the safety net while removing repetitive confirmations.

Starting in plan mode

The first step after context and permissions is to switch Claude Code into plan mode (toggle with Shift+Tab). In this mode the model can read files and run read‑only shell commands, but it cannot modify code. The author says this habit has saved her from many “bad ideas” early in a project because she can review the plan before any changes are made.

Once the plan looks good, she shifts back to the default mode or to acceptEdits, which automatically approves file edits and common filesystem commands inside the working directory. Two newer modes are also mentioned: auto (research preview that auto‑approves after safety checks) and bypassPermissions (runs everything without prompts, recommended only inside isolated containers or VMs). The choice of mode shapes the entire session, so picking the right one at the outset is crucial.

Choosing the right model and effort level

Claude Code offers three model tiers: Opus, Sonnet, and Haiku. Most users set a model once and forget about it, but the author matches the model to the task. Heavy reasoning or debugging work gets Opus; simple scripts and re‑organization tasks get Haiku; everything in between stays on Sonnet. She frequently uses the opusplan shortcut (/model opusplan) which runs Opus during plan mode for deep thinking and automatically switches to Sonnet for implementation.

An additional /effort command controls the “thinking budget” for the session. The default is sufficient for most tasks, but she bumps it to high or max when tackling problems that demand extensive reasoning. Together, model selection and effort level ensure you’re not over‑paying for cheap work or under‑powering complex challenges.

The four settings you should never skip

Summarising the article, the author stresses that the four settings—CLAUDE.md context, pre‑configured permissions, starting in plan mode, and model/effort selection—take roughly five minutes to configure but deliver a measurable productivity boost throughout the entire coding session. By treating Claude Code like a new teammate and giving it the same onboarding routine, developers can harness its power without the typical trial‑and‑error friction.