Anti-Pattern Rules

Every suggestion must pass this gate before being shown. If a suggestion violates any rule, discard it and generate a replacement.

The 11 Rules

1. No Restating the Obvious

Violation: "Would you like to know more about JWT tokens?" (when you just explained JWT tokens) Fix: "Implement token refresh with sliding window expiry" (advances the work)

The user already knows what you just told them. Move forward, don't circle back.

2. No Generic Filler

Violation: "What are the pros and cons?" / "Can you explain further?" / "Tell me more about this" Fix: "Compare HMAC vs RSA for your token signing — you're using HS256 but your team size suggests RS256"

Generic questions could apply to any topic. Every suggestion must be specific to THIS context.

3. No Already-Answered Topics

Violation: Suggesting the user explore something the conversation already covered in depth. Fix: Suggest the NEXT step beyond what was covered, or a different angle entirely.

Scan the conversation. If it was discussed, it's done. Move on.

4. No Overly Broad Suggestions

Violation: "Learn about security best practices" / "Tell me about database optimization" Fix: "Add CSRF protection to your three form endpoints" / "Add an index on users.email — your login query does a full table scan"

Narrow to the user's specific situation. Name files, functions, or concrete actions.

5. No Sycophantic Suggestions

Violation: "Would you like me to do anything else?" / "Is there anything else I can help with?" Fix: Remove entirely. This is never a valid next step.

The agent serves the user by suggesting useful actions, not by asking permission to exist.

6. No Hallucinated Context

Violation: "Continue working on your Kubernetes deployment" (when the user never mentioned Kubernetes) Fix: Only reference things explicitly mentioned in the conversation or found in project files.

If the user didn't mention it and it's not in their project, don't suggest it.

7. No Repetitive Structure

Violation: Three consecutive activations all starting items with "Consider..." or "Implement..." Fix: Vary phrasing: "Add...", "Deep-dive into...", "Resume...", "Quick win:...", "What if..."

Check your last few activations. If the structure feels repetitive, rewrite.

8. No Trivially Lookupable Items

Violation: "What is a JWT token?" / "What does async/await do?" Fix: "Implement async error boundaries for your unhandled rejections" (assumes knowledge, advances work)

If a web search answers it in 3 seconds, it's not a next step. Suggest actions, not definitions.

9. No Scope Mismatches

Violation: "Redesign your entire database schema" (when user asked a quick CSS question) Fix: Match suggestion effort to session scope. Quick fix session → quick win suggestions.

Read the room. A user fixing a typo doesn't want an architecture proposal.

10. No Format Repetition

Violation: All 5 suggestions starting with "Implement" or all being questions. Fix: Mix verbs: one "Implement", one "Explore", one "Resume", one "Consider", one "Quick win".

Visual monotony signals low effort. Vary the opening words and item structures.

11. No Low-Value Clarifications

Violation: "Would you like me to clarify the difference between X and Y?" (when it doesn't matter) Fix: Remove entirely, or replace with an actionable task.

Only clarify things that would change the user's approach. Most clarifications are filler.

Quality Bar

Every suggestion that passes the anti-pattern gate must ALSO satisfy ALL of these criteria:

• Specific: Names concrete files, functions, APIs, or actions — not vague categories
• Actionable: The user can act on it immediately without further research
• Non-obvious: The user wouldn't have thought of it on their own (or needs the reminder)
• Contextually grounded: Directly connected to the current conversation or project state
• Scope-appropriate: Effort level matches the session scope (quick fix vs. deep work)
• Differently framed: Each suggestion in the set uses different wording and structure

Self-Review Procedure

After generating all candidates:

1. Read each suggestion against all 11 rules above
2. Check each against all 6 quality criteria
3. Remove any violators
4. If removed count puts you below display-count, generate replacements
5. Verify the final set has category diversity (not all the same type)
6. Verify varied sentence structure across the set
7. Present the final set