new-rules-of-context-engineering-claude-5listed

# New Rules of Context Engineering (Claude 5) Apply Anthropic's Claude 5 context engineering principles to a scope of agentic files. The philosophy shift: from "constrain the model to prevent failure" to "design context the model can intelligently navigate". Claude 5 models need fewer constraints — Anthropic cut over 80% of Claude Code's own system prompt with no performance loss. Read `references/principles.md` before editing anything. It contains the six shifts, the per-file-type guidance, and the audit checklist you apply to each file. Two modes: by default, audit **and** rewrite. If the user asks only whether there are improvements to make ("audit", "check", "review"), run steps 1–3 without editing and deliver the step 4 report as findings with proposed changes instead of applied ones. (Claude Code's built-in `/doctor` does a quick automatic rightsizing of skills and CLAUDE.md; this skill is the thorough, reviewable version — use it when the user wants an audit trail or a whole-repo pass.) ## Step 1: Resolve the scope The user gives a scope: a single file, a directory, or a whole repo. For a repo, sweep the full tree (`git ls-files` filtered to the names below — nested `subdir/CLAUDE.md` files count and are often the right pattern, not a smell). Find the agentic files — files written to instruct an AI agent, not humans: - `CLAUDE.md`, `AGENTS.md`, `.cursorrules`, `.github/copilot-instructions.md` — at any depth, not just the repo root - `SKILL.md` and skill refer