antipatterns-reviewlisted

# Antipatterns Review This skill is your code-review mode. Walk through the user's design, code, or plan and check for each of the 12 canonical antipatterns. For each found, name it, explain the failure mode it produces, and propose the fix. This is not a generic "review my code." It's a targeted scan against the 12 known failure patterns that have hit real production agentic products. ## How to apply this skill When invoked, do this: 1. Ask the user to share what you're reviewing (code, design doc, screenshot, description) 2. Walk through the 12 antipatterns in order 3. For each: pass / present / unclear, with evidence 4. For each "present": name it, explain the failure, propose the fix 5. Prioritize by severity (critical > high > medium > low) 6. Summarize the top 3 to fix first ## The 12 antipatterns ### 1. Multi-agent before single-agent baseline **Signal:** the design has 3+ agents from day one; no single-agent version was tried. **Failure mode:** complexity tax without proven value. Sub-agents lose context to each other; coordination overhead dominates; debugging is hard. **Fix:** rebuild as single-agent. If it works at 80%+ pass rate, you have a baseline. Add a second agent only when the single-agent version demonstrably hits a wall (context exhaustion or genuine parallelizable subtasks). **Severity:** High — almost always over-engineering --- ### 2. Framework abstractions before understanding raw API **Signal:** code uses LangGraph / CrewAI / etc. but th