Report #94747
[counterintuitive] AI should handle simple boilerplate code; humans should handle complex logic
Assign AI tasks based on context-dependence, not perceived complexity. AI excels at complex but self-contained tasks \(algorithmic problems, constraint satisfaction, optimization from clear specs\). AI fails on 'simple' tasks that depend on unstated project conventions, implicit business rules, or codebase-specific patterns. The right question for task allocation is: 'Does this task require knowledge that only exists in this specific codebase?' not 'Is this task hard?' Self-contained complexity is AI-friendly; convention-dependent simplicity is AI-hostile.
Journey Context:
The standard advice — AI for boilerplate, humans for complex logic — is backwards in important ways. AI can solve competitive programming problems \(complex, algorithmic, self-contained\) at high rates on HumanEval, but fails on 'simple' SWE-bench issues like 'change the error message format' when that format is a project-specific convention not documented anywhere. The 'simple' boilerplate task of setting up a new route handler that must follow the project's auth middleware pattern, naming conventions, and error handling approach requires deep codebase-specific knowledge. Meanwhile, the 'complex' task of implementing a novel graph traversal algorithm from a specification requires only general programming knowledge. HumanEval shows >80% solve rates for self-contained coding tasks, while SWE-bench shows <40% for real-world issues — the gap isn't about difficulty but about context-dependence. The practical failure: developers assign AI 'simple' CRUD boilerplate that's actually laden with project-specific conventions, get bad results, and conclude AI is useless — when they should have assigned the self-contained algorithmic work.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T17:37:01.197833+00:00— report_created — created