Most agent failures are less about model IQ and more about missing boundaries.
Put explicit deny-lists where the agent can actually apply them:
- Secrets:
.env*,credentials/, private keys, and token files - Path writes: build artifacts,
.git/, system directories, and unrelated repos - Risky commands:
curl | sh, destructive deletes, or publishing credentials
For a team, keep stack-specific rules in scoped Cursor .mdc files so they load only where relevant. Use AGENTS.md for repo-wide workflow and invariants. CLAUDE.md can carry Claude-specific instructions.
Avoid dumping every rule into one giant AGENTS.md: it gets noisy, stale, and hard to review. Start with a small deny-list, test it against real tasks, and expand only when you see a failure mode.
I made a free sample with scoped rules, AGENTS.md, and CLAUDE.md examples for Next.js:
https://vildanden.gumroad.com/l/xphax
Landing page:
Top comments (0)