Claude Code AGENTS.md: What Actually Reaches Context
Four arms, one codeword each, every reading tool denied. Only CLAUDE.md loaded on its own; AGENTS.md needed one line to get in.
Sep 1611 min read
CATEGORY · 65 ARTICLES
AI Coding Assistants
Articles
AI Coding Assistants
Claude Code Compaction Explained: Two Real Events
Two auto-compactions cut roughly 98% of active context, took 52 and 147 seconds, and left structural evidence of what the CLI preserved.
Sep 1511 min read
Claude Code Custom Commands: 30 Files, One Dead Field
A custom command is a markdown file with five documented fields. Measured against 30 shipped commands, the field usage and the reference have come apart in both directions.
Sep 1512 min read
Claude Code Doctor Command: Exit 0 Can Mislead
claude doctor and /doctor are different surfaces. One prints read-only installation diagnostics; the other can investigate configuration and propose fixes.
Sep 159 min read
Claude Code Memory Files: 22 Sessions, 11 Memories
The CLAUDE.md hierarchy and the auto-memory directory are separate systems with separate costs. A census of both across 470 sessions on one machine.
Sep 1512 min read
Claude Code Sandboxes: There Is No Sandbox Flag
No flag creates an isolated Claude. What exists is three separate layers — one that removes your config, one that removes capability, and one with no flag at all.
Sep 1511 min read
Claude Code Verify Command: It Is a Skill
/verify is a bundled, user-invoked skill—not a shell subcommand—and it needs enough permission to launch and inspect the real application.
Sep 159 min read
Claude Code Agent Teams: 33 Tools Down to Six
One environment variable takes the tool list from 33 to 6. The lead keeps Agent and SendMessage, loses Bash and Read, and the same one-line answer costs 2.5x.
Sep 1311 min read