TechnicalJul 30, 2026~7 min · intermediate

The common mistakes with Claude Code

Claude Code multiplies a good dev and amplifies the bad habits of a rushed one. Here are the mistakes I see most, and the reflex that fixes each.

StackClaude CodeCLAUDE.mdgittestscontexte

01

The top mistakes

The mega-prompt

“Build the whole app” in one message. The agent drifts, mixes everything. Break into scoped tasks.

No CLAUDE.md

Without project memory, it re-guesses your stack each session and repeats choices against yours.

Accepting blindly

Approving diffs without reading them delegates responsibility for the code. You remain the pilot.

Never testing

Without having it run tests, the agent doesn’t know if it broke something — and neither do you.

02

Mismanaged context

The subtlest mistake: letting a session drag on until the context is saturated with abandoned old threads. The agent becomes confused, slow, contradictory.

bad reflex

one endless session.

·context polluted with finished tasks

·answers increasingly vague

good reflex

one session per task.

·start clean between big topics

·capture state in the CLAUDE.md

03

The fixing reflexes

·one task = one loop = one commit: everything reverts cleanly;

·have it describe its plan before editing, and fix the plan, not the code;

·have it run tests and build, and read its own error output;

·put repeated procedures in skills, permanent rules in CLAUDE.md.

Worth noting

Most of these mistakes vanish once the basics are set: the basics of Claude Code.

04

Sources & further reading

01Best practices for Claude Code — documentation officielleThe reference page: context window, /clear between tasks, plan mode, verification via tests, and a “common failure patterns” section listing the same mistakes.↗02How Claude remembers your project (CLAUDE.md & auto memory)Where to put a CLAUDE.md, what belongs in it, and why to target under 200 lines: beyond that, instruction adherence drops.↗03Extend Claude with skillsThe split rule: permanent facts in CLAUDE.md, repeated procedures in SKILL.md — loaded on demand, so they cost nothing until used.↗04Configure permissions — Claude CodePermission rules, allowlists and sandboxing: how to cut approval prompts without clicking “yes” blindly.↗05Common workflows — Claude CodeThe concrete recipes: tests, refactoring in small increments, PR creation, delegating research to subagents.↗06Effective context engineering for AI agents — AnthropicWhy a saturated context degrades the model (“context rot”), and the three remedies: compaction, structured notes, sub-agents.↗07Effective harnesses for long-running agents — AnthropicField notes on long-running agents: one feature at a time, git commits and a progress log to survive from one session to the next.↗

In short

None of these mistakes is technical: they're method shortcuts. Small loops, up-to-date CLAUDE.md, reviewed diffs, tests run. The agent doesn't replace rigor — it rewards it.

Claude Code docs ↗