从必答问题出发
明确范围、读者和深度，按证据缺口选来源，再检查每个问题是否落实到实际答复。来源、主张与不确定性分别记录。
开源研究技能
研究工具箱
把问题、证据、分析与审查连起来。用于产业、市场、公司、产品和技术等深度研究，让报告回答真正的问题。
审阅实际正文和证据
按任务组合意图、证据、对抗性、结构深度、读者用途、去过程化、自然表达与领域检查。重要争议独立裁决，通过的内容也纳入抽查。
每一路评审都有效完成
超时、截断和泛泛通过不能完成评审。失败先恢复，有依据的负面结果保留。最终检查绑定当前交付版本。
用缺陷改进方法
公开数据集包含 23 组成对语义诊断、9 条合成来源和 2 个合成任务。保留原始问题及评审分歧；开发材料不构成质量金标准或效果证明。
中文文档
将仓库链接交给 agent，先读 SKILL.md，按当前任务需要读取参考文件。检索、文件访问和模型调用由所用工具提供。工具箱不依赖特定模型或固定四路评审。
OPEN-SOURCE RESEARCH SKILL
Research Toolkit
Connect research questions, evidence, analysis and review. A reusable skill for substantial industry, market, company, product and technology reports.
Question-led evidence
Define the reader, scope and depth. Choose sources from evidence gaps, then map each required question to its actual answer. Distinguish sources, claims and uncertainty.
Review the actual text
Combine intent, evidence, adversarial reasoning, depth, reader usefulness, clean prose and relevant domain checks. Adjudicate important disputes independently and sample passed material.
Complete every required review
Timeouts, truncated replies and generic PASS claims do not count. Recover failed calls and retain valid negative results. Bind delivery checks to the current artifact.
Learn from preserved defects
The dataset contains 23 synthetic diagnostic pairs, 9 source records and 2 tasks. These development materials preserve disagreements and are not validated quality labels or effect estimates.
Documentation
Give your agent the repository link and ask it to read SKILL.md first. Load references as needed. Search, files and model execution come from the host tool; no particular model or four-provider panel is required.