Common Mistakes

Common pitfalls when using Skills and how to avoid them.

Mistake 1: Wrong Skill for the Job

Problem: Using review for security audits, simplify for architecture refactoring.

# Wrong
/review Check the login endpoint for SQL injection

# Right
/security-review Audit the login endpoint security

Each Skill has a specific focus. See When to Use Which.

Mistake 2: Conflicting Skills Simultaneously

Problem: Activating caveman (ultra-compact) and grill-me (deep interrogation) together.

These have opposite communication modes — one minimizes, one maximizes. Both lose effectiveness.

Fix: Use one Skill at a time. Switch when done.

Mistake 3: TDD for Exploratory Work

Problem: Writing tests before understanding what to build.

# Requirements are unclear
/tdd Build a recommendation system

TDD needs clear requirements. Use prototype for exploration first.

Mistake 4: Skipping Diagnosis

Problem: Patching symptoms without finding the root cause.

# Adding retry logic without understanding the failure
/tdd Fix intermittent payment callback failures with retry logic

Use /diagnose first. Blind fixes may mask real issues.

Mistake 5: Overusing Skills

Problem: Using /review for a one-line comment change, /tdd for a config tweak.

Skills have startup cost (loading context). Simple tasks are faster without them.

Mistake 6: Ignoring "When NOT to Use"

Every Skill doc lists inappropriate scenarios. For example:

• simplify doesn't do architecture refactoring
• review doesn't do dedicated security audits
• prototype code isn't for production

Ignoring these boundaries degrades output quality.

Mistake 7: Prototype Code in Production

# Wrong flow
/prototype Build user system → prototype works → deploy directly

prototype emphasizes disposability. After validation, re-implement with tdd.

Mistake 8: Poor Context Management

Chaining many complex Skills quickly drains the context window.

Fix:

• Use caveman mode to reduce tokens
• Use handoff between phases
• Use loop for background periodic checks

Mistake 9: Blindly Trusting Skill Output

Skill output is AI-generated and may have errors. Especially:

• Security review false positives/negatives
• Code generation ignoring project conventions
• PRDs missing business context

Always review Skill output. Never accept blindly.

Summary

Mistake	Fix
Wrong Skill	Check selection guide
Conflicting Skills	One at a time
TDD for exploration	Prototype first
Skipping diagnosis	Diagnose first
Overusing Skills	Simple = direct
Ignoring boundaries	Read "When NOT to Use"
Prototype in prod	TDD after prototype
Context mismanagement	Caveman / handoff
Blind trust	Always review