Documented failure modes from working with Claude Code (Anthropic's AI coding agent). Each file is a post-mortem of a real incident where the agent made a mistake that cost time, destroyed work, or required human intervention to correct.
AI coding agents fail in predictable, repeated patterns. Documenting these failures concretely — with what happened, why, and what rule would have prevented it — is more useful than vague warnings about AI limitations. These are field notes, not benchmarks.
- Assuming instead of observing — generating plausible explanations without checking tool output
- Acting on questions — treating "why did you do X?" as "change X" instead of answering
- Hiding bugs instead of fixing them — rewriting tests, loosening assertions, adding retries to mask real failures
- Panic reverting — undoing verified work when an unrelated test fails, instead of investigating
- Skipping dependencies — declaring "ready" without reading linked issues, PRs, or referenced code
- Never committing — touching dozens of files without a single commit, then losing everything to
git reset - Completion drive — the urge to produce a "done" signal overriding the need to actually understand the problem
Each file is named YYYY-MM-DD-short-description.md and contains:
- What happened — concrete sequence of events
- The actual answer — what should have been done
- Root causes — why the failure occurred
- The rule — what principle, if followed, would have prevented it
If you work with AI coding agents and have documented failures, contributions are welcome. Same format: what happened, why, what prevents it.
These are facts about things that happened. Use them however you want.
