Most developers spend their time configuring what AI agents produce - code style, naming conventions, architecture patterns. Almost nobody governs how those agents actually work. Travis Drake, an industrial/organizational psychologist (the field that studies human behavior in workplaces), noticed something familiar: AI coding agents fail in the same predictable ways that humans do.
So he built a free, open-source framework of six behavioral rules that you copy into your agent's configuration directory. Total cost: about 1,500 tokens of context. It works with Cursor, Claude Code, Windsurf, and anything else that reads system-level instructions.
The 11 Named Failure Modes
The framework catalogs 11 anti-patterns, each named after the real incident that inspired the rule. Naming matters here - telling an agent "don't cut corners" is vague and ignorable. Telling it to watch for "The 7% Read" (where the agent skims a file and edits based on the first 7% it processed) gives it a concrete behavior to avoid.
Some of the patterns that stood out to me:
- The 7% Read - Agent skims a file instead of reading it fully before editing. The fix: a rule called
never-give-up-readingthat forces complete file reads. - The Trailing Off - Quality degrades when an agent handles a sequence of tasks. First task gets full attention, fifth task gets sloppy shortcuts. Rule:
never-give-up-planning. - The Spot Check - Agent runs one test instead of the full verification suite and calls it done. Rule:
never-give-up-checklist. - The Confident Declaration - Code that looks syntactically correct but doesn't actually work, and the agent ships it anyway because it "looks right." Rule:
never-trust-syntax. - The Pass-Through - Agent delegates to a subagent, gets back a failure, and accepts it without question. Rule:
never-trust-agents.
Does This Actually Work?
The psychology behind it is sound. I/O psychology has decades of research showing that named, specific behavioral interventions outperform vague directives for humans. The same principle applies to LLMs - concrete examples of what not to do are more effective than abstract instructions.
Each rule follows a consistent template: a one-sentence mandate, specific examples of the anti-pattern, corrective behaviors, and the origin story of why the rule exists. At roughly 150 tokens per rule, the overhead is minimal.
The project is early (4 commits, CC BY 4.0 license), but the approach is practical. If you're running any AI coding agent and noticing sloppy behavior, dropping these six files into your config directory is a zero-risk experiment. The worst case is you lose 1,500 tokens of context window. The best case is your agent stops pretending it read the whole file.