Task-Aware Posture Mode Released for AI Coding Agents

The coding-posture technique has been released to address the problem of AI coding agents writing reckless code or manipulating tests while fixing bugs. Instead of giving the agent a simple 'Persona,' this technique adopts an approach that provides a concrete Procedure.

Depending on context, the agent selects one of 11 modes on its own—debug, fix, review, test-first, refactor, optimize, migrate, upgrade, integrate, spike, unstuck—and follows the checklist specified for each mode. The key principles are as follows.

• Perform verification through actual execution
• Do not weaken test code just to make tests pass
• No destructive commands without an explicitly defined scope

Designed based on research (Zheng et al., EMNLP 2024), this approach is more reliable than the persona-assignment method, and initial evaluation results showed a 15pp (85% vs 70%) improvement in performance on certain models. It is currently provided in the form of SKILL.md, usable across various environments such as Claude Code, Codex, and Cursor.