coding-pitfalls

Require the model to state assumptions explicitly before implementing and to present alternatives when multiple interpretations exist. This counters anchoring bias, where an initial framing persists despite contradictory evidence.

The most common are silent assumptions, overcomplexity and abstraction bloat, unscoped side-effect edits, N+1 query patterns, and security vulnerabilities. AI-generated code shows roughly 2.7x higher vulnerability density than human code.

LLM effectiveness drops 60-80% after two or three fix attempts on the same bug, a pattern called debugging decay. The fix is to stop, revert to the last known good state, re-read the error with fresh context, and reason about root cause before editing.

Not on their own. Models test what they think the code does rather than what the spec requires, systematically under-testing edge cases. Write acceptance criteria from the specification first, then generate tests against those criteria.

Define explicit input/output contracts at every agent boundary and give each subagent a single stated done-condition. Have the orchestrator verify the final output against the original specification rather than trusting the last subagent's output.

Never rely on it unverified for unfamiliar APIs, since models hallucinate deprecated methods and removed parameters from stale training data. Pin to current official documentation and verify method signatures before use.