Steering Claude Code: CLAUDE.md files, skills, hooks, rules, subagents and more

Summary

Claude Code offers seven distinct mechanisms for guiding its behavior in software projects: CLAUDE.md files, rules, skills, subagents, hooks, output styles, and direct system-prompt appending. These approaches differ primarily in when instructions enter context, how persistently they remain active, and the token cost they impose relative to their authority. Model selection and effort settings remain separate controls that determine baseline capability and diligence, independent of these steering methods.

CLAUDE.md files supply always-on project context by loading a markdown document from the repository root at the start of every session. The file remains resident throughout the session and suits stable information such as build commands, directory structure, coding conventions, and team norms. In shared repositories the file tends to accumulate contributions without pruning, so every engineer receives the full contents regardless of task relevance. This produces rising token consumption and weaker adherence to any single instruction as length grows.

To limit these effects, the guidance recommends keeping CLAUDE.md under roughly 200 lines, assigning an owner, and treating edits like code changes. In monorepos, teams can place scoped CLAUDE.md files inside their directories and rely on exclusion settings to prevent unrelated content from loading. Organization-wide standards such as security or compliance rules can instead be deployed centrally through device management so they cannot be bypassed locally. The file functions best as a high-level index that points Claude toward more targeted instructions stored elsewhere.

Why it matters

Highly actionable for Product Teams and Builders with concrete implementation guidance, context management strategies and workflow examples directly applicable to AI coding tools.