Claude Code guides compare `.claude/` commands, agents, and global rules

• Two widely shared Claude Code guides are converging on the same workflow model: use CLAUDE.md and the .claude/ folder as a lightweight control surface for project instructions, custom commands, agents, skills, and permissions rather than a giant prompt blob, according to the cheat sheet and the folder guide.
• The practical split is getting clearer: the folder guide frames CLAUDE.md as the primary instruction file, while Hacker News commenters cited in the discussion describe .claude/commands/ for repeatable tasks, .claude/agents/ for narrow roles, and .claude/skills/ for broader behavior patterns.
• Practitioners are favoring smaller rule sets over heavier prompt stacks: in fresh discussion, commenters say to keep instructions “short and simple,” and examples in the thread discussion focus on a few durable guardrails like “don’t push to main” and “always read CONTRIBUTING.md.”
• The guides also double as change-tracking for a fast-moving CLI: the cheat sheet includes recent additions such as transcript search and a PowerShell tool, while the HN discussion says the sheet auto-tags new features with a daily changelog check.

The new overlap is less about introducing a single feature than about documenting the operating surface Claude Code users now have to manage. Martin Baláž’s cheat sheet compresses shortcuts, slash commands like /clear, /model, and /effort, MCP server setup, memory files, and newer additions such as transcript search and a PowerShell tool into one reference. In the companion folder guide, the .claude/ directory is presented as the project-level control plane: CLAUDE.md for shared instructions, CLAUDE.local.md for personal overrides, and a global ~/.claude/ area for preferences and reusable setup.

That structure is what engineers appear to be standardizing around. The guide says /init can scaffold the starting point, then the rest of the folder becomes a place to separate stable instructions from task-specific automation. In the discussion around the cheat sheet, the author says the page was built because they “kept forgetting commands,” and a daily cron job now checks the changelog and marks additions with a “NEW” badge, which underscores how quickly the CLI surface is changing.

The strongest practitioner signal is about scope discipline. In the fresh Hacker News discussion, several commenters converge on keeping instructions “short and simple,” with one saying “more instructions does not mean better results” in the latest thread. Another commenter, quoted in the discussion, draws a practical boundary: use .claude/commands/ for “well defined repeatable” tasks, .claude/agents/ for narrow-scoped helpers, and .claude/skills/ for “job description like behaviors.”

The rule examples are similarly concrete. One user’s global Claude.md, cited in the thread discussion, encodes defaults such as “always use uv and venvs,” “don’t push to main ever,” and “always read CONTRIBUTING.md.” Another says their setup always runs task build before Claude can claim work is done, a simple check surfaced in the core discussion to catch false completions. The pushback is also part of the story: today’s delta notes skepticism that the folder should become a bloated framework, so the emerging consensus is not “add more structure,” but add only the smallest repeatable constraints that prevent common mistakes.