waylean/siduchishui-skill

This project, called maneuver-decision, is a skill for AI agents that gives them a structured way to make decisions under uncertainty, aimed at situations like technical diagnosis, production incidents, and risky code changes. The methodology is drawn from the Four Crossings of the Chishui River military campaign, combined with ideas from Mao's Selected Works around investigation, contradiction analysis, testing ideas through practice, staged judgment, and combining a single sample with the wider picture. The README is explicit that it does not retell the history itself or use historical comparisons as evidence for a current problem, it only extracts the transferable decision-making approach. When an AI agent uses this skill on a complex bug, an outage, a performance regression, a legacy system change, a large pull request, or a high risk fix, it is prompted to answer a set of questions first: whether the known facts are sufficient, what the core conflict is between the goal and the main constraint, which stage the situation is currently in, which direction could change things at the lowest cost, which action would verify a judgment fastest, what signal would confirm it, what signal would prove it wrong, and how to roll back or change direction if it turns out to be wrong. The README includes evaluation results comparing this skill against a plain structured baseline prompt across different scenario types. It reports a small overall improvement, with the clearest gains in technical diagnosis scenarios, but slightly worse results in scenarios like product decisions, organizational firefighting, and architecture management, where the author notes the skill still lacks business and organizational context. The README is candid that most of the benefit comes from having any structured format at all, and that this particular skill should not be treated as a universal framework. To use it, you copy the maneuver-decision folder into a Codex skills directory. The package includes a SKILL.md file, an agent configuration file, and reference documents covering the historical case, the Mao's Selected Works methodology, and a validation guide. You then trigger it in a prompt, for example asking the agent to use it to diagnose a performance regression or review a large pull request, and it responds with a structured output covering the main constraint, current stage, known facts, missing evidence, and a rollback plan.