使用方法
- 复制安装命令，在自己的智能体环境安装前先检查源码包。
- 从原始文档核对所需工具、凭据和支持的宿主。
- 说明输入材料与期望交付物，并对照来源检查输出。
npx skills add https://github.com/daveebbelaar/bama --skill research开始前需了解
- 安装不包含模型订阅、API 额度或访问文件的权限。
- 仓库星数和安装量反映活跃情况，不等于独立质量或安全测试。
按源仓库内容呈现，保留标题、案例、代码、表格、链接以及原文引用的演示图片。
Research
Research answers three questions before code exists: what the best apps in the category do, what users love and hate about them, and how macOS lets you build it. The output feeds app-brief and user-stories.
1. Frame
Write the question in one line, and the decision it informs ("which features go in version one", "Carbon hotkeys or an event tap"). If the user gave only a category, the question is "what makes the best <category> app on the Mac".
2. Fan out
Start one subagent per track, in parallel, with the brief below. Pick the tracks that fit. Four is typical. Market and Reviews always run. Reviews covers many apps: give each subagent 3 or 4 apps, or the 600 words run out.
| Track | Finds |
|---|---|
| Market | The 5 to 8 best apps in the category, their price, platform, and headline features |
| Reviews | Per app: what people like and what they dislike, from App Store and Mac App Store reviews, Reddit, Hacker News, forums, and "switched from X" posts |
| Features | The full feature list of the 2 or 3 leaders, from docs, changelogs, and help pages |
| Design | Screens, interaction details, and keyboard shortcuts of the leaders |
| Platform | The macOS APIs, permissions, and limits each feature needs, and what other apps on this Mac already use (global shortcuts, URL schemes, bundle ids) |
Subagent brief (fill in the brackets):
Research [track] for a native macOS app that [one-line idea].
Question: [the question]. Decision it informs: [decision].
Search the web. Prefer primary sources: official docs, changelogs, Apple
developer docs, WWDC sessions, the app's own help pages. Use reviews and
forums for opinions only. If a source is blocked, say so and use another
(Hacker News, App Store reviews, Mac forums). Do not write any files.
Tag every claim: [V] verified in a source (give the URL), [K] from your
own knowledge, [U] unverified, must be tested.
Return at most 600 words: a findings list, a table where it helps, and
the 3 things that matter most for the decision. No intro, no summary.3. Merge
Write one file per question: docs/research/<topic>.md.
# <Topic>
Research for <AppName>. Done <YYYY-MM-DD>.
[V] verified in a source, [K] from knowledge, [U] unverified, test first.
## Findings
- ... [V](url)
## Likes and dislikes
| App | People like | People dislike | Sources |
| --- | --- | --- | --- |
| App A | <3 to 5 points, most mentioned first> | <3 to 5 points> | <urls> |
Patterns: <likes that repeat across apps (must have), dislikes that repeat
(our chance to do better)>
## Feature matrix
| Feature | App A | App B | App C | Version one? |
## Decisions
1. <decision>, because <finding>. Planned test: <test, for a [U] finding>.
## Open questions
- <[U] claim and how to test it>Keep the feature matrix: user-stories turns its "Version one?" column into stories. Every "People like" point maps to a decision or to an "accepted gap" line in the brief, so nothing loved is dropped by accident. Check whether a system store (EventKit for reminders and calendars, CloudKit) gives sync and phone capture for free. Turn each [U] that a decision depends on into a test before you trust it (a probe script, a self-test step, or a preview field in the reference app).
4. Hand off
Call the Skill tool with clone if one app stands out as the reference and you can open it. If it is not installed, or computer use and a browser are not available, link its public help pages and screenshots in the research file instead. Then call the Skill tool with app-brief.
Done when every track returned, the file has a likes and dislikes table for every app in the market list, a feature matrix, and a decision list, and every decision cites a [V] or a planned test.
