给 Agent 用的两个研究 Skill。它们负责搜索，也会检查来源、时效和结论是否站得住。
Two research Skills for Agents. They search, check the evidence, and explain how much each conclusion can be trusted.
diving-in-research 把一次研究任务拆成几个简单步骤：确定范围、搜索、去重、核验、评级和总结。结果会注明来源、时间、适用条件和证据缺口。
项目包含两个 Skill：
| Skill | 用途 | 主要来源 |
|---|---|---|
diving-in-social |
研究中文舆情、真实体验、消费口碑和跨平台观点 | 微博、抖音、小红书、B站、知乎；按需补充快手和微信公众号公开文章 |
diving-in-tech |
排查报错、核对版本、比较方案和验证社区经验 | 官方文档、源码、GitHub、Stack Overflow、Gitee、V2EX 和精选中文技术社区 |
- 评级对象是具体结论，不是平台或作者。点赞、采纳、stars 和搜索排名只用来发现内容。
- 社交内容区分实测、亲历、转述、推断和机构声明。技术内容分别评价来源质量与解法适用性。
- 重复转载只算一个来源。旧内容、版本不符、营销软文和无依据断言会被降级或排除。
- 检索有明确上限，只保留短证据卡，避免把整页正文和长评论区塞进上下文。
多平台搜索已经有很多方案。这个项目主要处理搜索之后的问题：资料是不是独立的、有没有过期、能不能复核，以及最终结论有多可信。
从 GitHub 安装：
npx skills add weizesunmd-byte/diving-in-research --skill diving-in-social -g
npx skills add weizesunmd-byte/diving-in-research --skill diving-in-tech -g从本地仓库安装：
npx skills add . --skill diving-in-social -g
npx skills add . --skill diving-in-tech -g安装后新建一个 Agent 会话，让宿主重新发现 Skill。
不登录也能使用公开网页和搜索索引，但覆盖会少一些。要使用完整能力，请提前准备：
- 在 Agent 使用的浏览器里亲自登录抖音、小红书、快手，以及需要登录的 B站或技术社区。
- 安装
zhihuSkill，并按它的初始化流程配置知乎开放平台 Access Secret。 - 为技术研究配置 GitHub Connector/MCP，或运行
gh auth login；需要 Gitee 时再配置对应的官方访问能力。
密码、验证码、Cookie、Token、SSH key 和 Access Secret 不要写进普通研究提示、仓库、Issue 或日志。认证缺失时，Skill 会改用公开来源，并在结果中说明覆盖缺口。
使用 $diving-in-social 研究最近三个月大家对某产品的真实评价，审查营销和模板化灌水，并给出 A–E 评级。
使用 $diving-in-tech 排查这个报错，优先核对官方文档、源码、版本和可复现性，并给出 A–E 评级。
每个关键结论使用六个维度评分，总分 100：
| 等级 | 分数 | 含义 |
|---|---|---|
| A | 85–100 | 证据充分，可以高置信采用 |
| B | 70–84 | 证据较完整，但仍有次要限制 |
| C | 55–69 | 只能带条件使用，需要继续验证 |
| D | 40–54 | 只作线索 |
| E | 0–39 | 不采用 |
评分脚本只负责统一算术和强制封顶，不能代替事实核验。每个分数仍需对应原始链接、时间、版本、复现结果或反证。
这是低量、只读的研究工具，不是批量爬虫。它不会自动点赞、评论、关注、发帖或下载媒体，也不提供隐藏接口、请求签名、代理轮换、验证码绕过和反检测功能。遇到验证码、限流或账号警告时，该平台会立即停止。
如果你实际使用过，请告诉我们具体场景和结果。有效反馈不需要写很长，说明使用了哪个 Skill、想解决什么问题、哪里不准确，以及你认为应该怎样改就够了。
反馈中不要包含账号凭据、Cookie、Token、私有代码、内部网址或未经处理的个人信息。
当前版本是 Preview。项目包含自动测试、16 个行为评测场景和两次实际烟雾测试。GitHub CI 会在 Python 3.11 和 3.13 上运行发布验证。跨宿主行为复测还在继续。
diving-in-research turns a research request into a bounded workflow: define the scope, search, deduplicate, verify, grade, and summarize. The result keeps links, dates, applicability, and evidence gaps visible.
The repository contains two Skills:
| Skill | Use it for | Main sources |
|---|---|---|
diving-in-social |
Chinese public opinion, firsthand experience, product feedback, and cross-platform discussion | Weibo, Douyin, Xiaohongshu, Bilibili, and Zhihu, with Kuaishou or public WeChat articles when useful |
diving-in-tech |
Debugging, version checks, solution comparison, and validation of community advice | Official documentation, source code, GitHub, Stack Overflow, Gitee, V2EX, and selected Chinese developer communities |
- Grades apply to specific claims, not to platforms or authors. Popularity and search rank help discovery but do not prove correctness.
- Social evidence is labeled as measurement, firsthand account, relay, inference, or official statement. Technical research grades source quality separately from solution fitness.
- Copied posts count as one source. Stale content, version mismatches, marketing copy, and unsupported claims are downgraded or excluded.
- Search and context budgets are explicit. The Skills keep short evidence cards instead of copying full pages and long comment threads.
Multi-platform search already exists. This project focuses on what happens after retrieval: provenance, independence, recency, applicability, and calibrated confidence.
From GitHub:
npx skills add weizesunmd-byte/diving-in-research --skill diving-in-social -g
npx skills add weizesunmd-byte/diving-in-research --skill diving-in-tech -gFrom a local clone:
npx skills add . --skill diving-in-social -g
npx skills add . --skill diving-in-tech -gStart a new Agent session after installation so the host can discover the Skills.
Public pages and search indexes work as a reduced mode. For fuller coverage:
- Sign in manually in the Agent-controlled browser for Douyin, Xiaohongshu, Kuaishou, selected Bilibili pages, and developer communities when needed.
- Install the
zhihuSkill and configure its Access Secret through its current initialization flow. - Configure the GitHub Connector/MCP or run
gh auth login. Add official Gitee access only when the target project is hosted there.
Do not put passwords, verification codes, cookies, tokens, SSH keys, or Access Secrets in ordinary research prompts, repositories, issues, or logs. When authentication is missing, the Skills fall back to public sources and report the loss of coverage.
Use $diving-in-social to research recent real-world opinions about this product, remove copied marketing content, and assign A–E confidence grades.
Use $diving-in-tech to investigate this error using official documentation, source code, version checks, and reproducible evidence.
Real usage reports are useful. A short report is enough: name the Skill, explain the task, describe what was wrong or missing, and suggest the expected behavior.
Do not include credentials, cookies, tokens, private code, internal URLs, or unredacted personal data in public feedback.
This is a bounded, read-only research tool, not a bulk crawler. It does not automate engagement, media downloads, hidden APIs, request signing, proxy rotation, CAPTCHA bypass, or anti-detection behavior. A CAPTCHA, rate limit, or account warning stops that platform.
The project is currently a Preview. It includes automated checks, 16 behavior cases, and two documented smoke tests. GitHub CI validates the repository on Python 3.11 and 3.13. Cross-host behavior testing is still in progress.