~/.claude/settings.json

~/.claude/settings.json config

@@ permissions.deny @@
	"deny": [
	"Read(*.env)",
+	"Read(/plans/archive/)",
	"Bash(rm:*)",
	"Bash(curl:*)"
	],
@@ ~/.claude/CLAUDE.md — new section @@
+	## Search Exclusions
+
+	- Never include `*/plans/archive` (any depth) in file
+	searches, glob patterns, or exploratory reads.

	## Working Style

+6 -0

layered search exclusions: instruction + permission deny

Excluding a directory from Claude searches isn't just a CLAUDE.md instruction — it's a two-layer problem.

CLAUDE.md tells Claude not to look there. But if the model ignores the instruction, or a subagent doesn't inherit it, files still get read.

The fix: pair the instruction with a `permissions.deny` rule in `settings.json`.

"deny": ["Read(*/plans/archive/*)"]

This blocks the `Read` tool at the permission layer — no instruction-following required. The `*/plans/archive/*` wildcard matches any depth, across any project.

The two layers are complementary:
— CLAUDE.md: prevents the call from being attempted
— `permissions.deny`: enforces it if the call happens anyway

For a completed-plans archive, or any "don't touch" path, this is the pattern that actually holds.

Follow for more Claude Code configuration patterns like this.

#ClaudeCode #AIEngineering #DeveloperTools #LLMOps

Twitter/X

Two layers beat one for Claude search exclusions.

CLAUDE.md tells it not to look. `permissions.deny` blocks the Read tool if it tries anyway.

"Read(*/plans/archive/*)" — enforced at the tool layer, no model cooperation needed.

Bluesky

CLAUDE.md instructions and `permissions.deny` rules aren't redundant — they cover different failure modes. Instructions prevent the call; deny rules block it if it happens. Layering both on `*/plans/archive/*` means the exclusion holds even across subagents.