agentsclimarketplace

Firsthand deepdive

Skill yinjialu/ai-frontier-daily/skills/firsthand-deepdive

专题深度线——对一个值得深挖的功能点(如 Claude Tag)生成深度文章草稿。从内参索引列「值得深挖」候选→用户选→抓官方素材生成五段骨架 draft→交 article-harness 打磨。用户说「深挖 X / 做个 X 的专题 / 专题深度」时使用。From its SKILL.md

Install
npx -y skills add yinjialu/ai-frontier-daily --skill firsthand-deepdive

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
  • 3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.6 KB, ~1.1k tokens by cl100k_base, as published. Nobody here has run it

专题深度线

内参的第三条内容线:把官方一手的重要功能点(模型发布/重大功能)做成面向普通用户的深度文章。

流程

1. 选题初筛(AI 列候选,用户选)

python3 -c "import json,datetime; from scripts.firsthand.query import load_index; from scripts.firsthand.deepdive import candidates; \
today=datetime.datetime.now(datetime.timezone(datetime.timedelta(hours=8))).date().isoformat(); \
[print(i+1, '|', g[0]['title'], '|', ', '.join(a['source'] for a in g)) for i,g in enumerate(candidates(load_index('data/firsthand'), 7, today))]"

把候选清单(每个 = 一个事件,可能多源)报给用户,用户选一个。 也可用户直接指定(不在索引里时,手工构造 cluster:每篇 {title, source, resource, summary, detected})。

2. 生成骨架 draft

判断该功能当前能否体验(如 Claude Tag 暂不可体验 → feature_available=False):

python3 -c "from scripts.firsthand.query import load_index; from scripts.firsthand.deepdive import candidates, write_draft; \
import datetime; today=datetime.datetime.now(datetime.timezone(datetime.timedelta(hours=8))).date().isoformat(); \
cands=candidates(load_index('data/firsthand'), 7, today); \
p=write_draft('content', cands[<选中序号-1>], today, feature_available=False); print('draft:', p)"

产出 content/deepdive/<date>_<slug>/draft.md:五段骨架 + frontmatter sources(官方 URL)+ 官方原文素材。

3. article-harness 迭代润色

article-harness 已适配多宿主,同一个 Writer/Reviewer 协议按当前 Agent 环境选择执行方式:

Codex 内使用时(优先):

  • 读取并遵循 ~/.codex/skills/article-harness/SKILL.md
  • 不从 shell 直接跑 ~/.claude/.../harness.sh;由当前 Codex 主 Agent 按 article-harness 的 Codex subagent 流程编排 Writer/Reviewer 迭代。
  • 输入草稿为 content/deepdive/<date>_<slug>/draft.md,产物契约保持不变:content/deepdive/<date>_<slug>/.harness-workspace/finished.mdfeedback.md

Claude Code 内使用时(暂时保持原链路不变):

bash ~/.claude/skills/article-harness/harness.sh content/deepdive/<date>_<slug>/draft.md

Writer 整理润色 ↔ Reviewer 迭代,PASS 后自动进入步骤4。

4. 自动推送微信草稿箱

harness PASS 后,立即用 wechat-official-draft skill 将 <draft_dir>/.harness-workspace/finished.md 推送到微信公众号草稿箱,无需用户介入

用户在微信草稿箱完成最终验收。

5. 反馈沉淀(推送后主动询问)

草稿推送成功后,立即询问用户:"草稿已推送,有没有想沉淀的审稿反馈?" 用户说完后,将可泛化的反馈条目追加写入 ~/Documents/context-harness-data/article-harness/reviewer.md,并告知已写入。 没有反馈则跳过。

与其他线的关系

  • 选题来源:内参 data/firsthand/index.json(官方一手)。
  • 不重造 Writer-Reviewer(复用 article-harness);article-harness 已收敛进公开仓 yinjialu/bianliang-skills(skills.sh,npx skills add),全局安装路径按宿主区分:Codex 使用 ~/.codex/skills/article-harness/ 的 Codex 模式,Claude Code 暂继续使用 ~/.claude/skills/article-harness/ 的 CLI runner。

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Keep looking

Skills are one crate of 326,852. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.