If you hand work to an AI agent (Cursor, Claude Code, Codex, custom agents), most failures trace back to the brief, not the model. Here are 10 anti-patterns I keep seeing, each with a one-line fix. Copy freely.
Smell these in your briefs. Each wastes tokens, time, or trust.
1. The vibes brief
“Make it better.” / “Clean this up.” / “Explore options.”
Fix: Name the artifact and the acceptance test.
2. Hidden prod access
Tools say “repo write: Y” with no path limits.
Fix: Allowlist directories; forbid .env, secrets, infra by default.
3. Success = activity
“Research the market” with no table, memo, or decision.
Fix: Definition of Done = files a human will open in a meeting.
4. Constraint collision
“Be exhaustive” + “max 10 minutes” + “cite everything.”
Fix: Rank constraints; mark Hard vs Soft.
5. Secret smuggling
API keys in the brief “for convenience.”
Fix: Point to a secret manager / env name; never paste values.
6. Infinite autonomy
“Whatever you need to get it done.”
Fix: Escalation table; stop conditions.
7. Unowned brief
No human owner → agent invents product policy.
Fix: Owner field required; acceptance is a human step.
8. Example-free taste tasks
“Write in our voice” with no voice doc or samples.
Fix: Link 2 good examples or a 10-line voice sheet.
9. Fake precision
“Confidence must be 95%.”
Fix: Ask for source grades and assumption tags instead.
10. Scope amnesia
Out of scope empty → agent expands forever.
Fix: At least two explicit non-goals.
Quick pre-flight checklist
- [ ] Named artifact + acceptance test
- [ ] Tool/path allowlist; secrets referenced by env name only
- [ ] Hard vs Soft constraints ranked
- [ ] Stop conditions + escalation owner
- [ ] At least two explicit non-goals
- [ ] 1–2 examples for any "taste" task
This is a free excerpt from a small set of briefing templates I sell. If you want the full fill-in-the-blank packs (briefing, eval, digest, research, incident), they're at https://briefstack.surge.sh/ — the list above is useful on its own either way.
Top comments (0)