qwenchance

qwenchance is a behavioral-discipline skill for Claude Code that keeps long, multi-step agentic tasks from failing in the three ways they usually fail: looping on repeated actions, over-thinking without acting, and silently running out of context. It packages a single SKILL.md checklist the model is meant to run before each step, converting vague self-monitoring into concrete triggers and matching actions so the agent doesn't deliberate about whether it is stuck.

The guidance is organized into four parts. Loop detection lists explicit signals (re-reading an unchanged file, re-running a command with the same arguments, revisiting a dropped hypothesis) and a hard retry cap of three attempts before stopping to ask the user. A thinking budget caps reasoning at roughly 1000 words per step, pushing the model to decide, act, and observe rather than re-derive decisions. A context-budget section replaces gut-feel estimation with a count of concrete pressure signals (turn count, files read, plan steps remaining) mapped to CONTINUE or HAND OFF. The handoff routine lands durable artifacts first, then invokes a separate `handoff` skill to compact the conversation.

It targets Claude Code users and agent operators running long research, debugging, or refactoring sessions who want to avoid wasted cycles and mid-task context exhaustion. It is a lightweight, single-file prose skill with no scripts or executable code, so its value is purely as instructional guidance layered onto an existing agent workflow.