[REPLACE: SKILL_NAME] Skill
来自 RSS 订阅源和开放网络的 [REPLACE: TOPIC] 最有趣新帖摘要
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
${var} — Optional. Pass a different topic to override the default. If empty, digests [REPLACE: TOPIC].
Today is ${today}. Build a digest of the [REPLACE: MAX_ITEMS] most interesting new posts on [REPLACE: TOPIC].
Steps
- Read sources — pull the last 24h of entries from each feed:
[REPLACE: FEED_URLS]
(Comma- or newline-separated list of RSS/Atom URLs.)
Use WebFetch to retrieve each feed and parse the entries. If a feed 404s or returns malformed XML, log a single warning line and skip that feed for this run.
- Augment with web search — run a
WebSearchfor[REPLACE: TOPIC] latestand pick up to 5 fresh links published in the last 24h that aren't already in the feed results.
- Score and rank — for each candidate, score on:
- Recency — within the last 24h gets full marks.
- Source weight — feeds in the configured list outrank generic search results.
- Specificity — items mentioning concrete numbers, code, or named systems beat opinion pieces.
Drop anything obviously off-topic (the ${var} or [REPLACE: TOPIC] keyword should appear somewhere in title or summary).
- Pick the top [REPLACE: MAX_ITEMS] — write
output/articles/[REPLACE: SKILL_NAME]-${today}.mdwith one entry each:
### [Title](url)
*[Source · published date]*
2-3 sentences distilling the takeaway. No filler.
- Notify via
./notifywith:
*[REPLACE: TOPIC] digest — ${today}*
[N] picks. Top item: [shortened title].
Full digest: https://github.com/${GITHUB_REPOSITORY}/blob/main/output/articles/[REPLACE: SKILL_NAME]-${today}.md
- Log — append to
memory/logs/${today}.md:
## [REPLACE: SKILL_NAME]
- **Sources scanned**: N feeds + 1 web search
- **Items picked**: N (of M candidates)
- **Top source**: domain
- **Status**: DIGEST_OK | DIGEST_QUIET (no items) | DIGEST_DEGRADED (some feeds failed)
Network note
WebFetch and WebSearch are built-in Claude tools. There is no network sandbox — curl works too; use WebFetch as the fallback for a flaky public GET. For this research skill the reads are unauthenticated, so WebSearch + WebFetch are the simplest path.
Constraints
- Never repeat. Track which item URLs went out via
memory/topics/[REPLACE: SKILL_NAME]-seen.txt(append-only). Skip anything that's already in there. - No filler. If fewer than
[REPLACE: MAX_ITEMS]items meet the bar, send fewer items — never pad with low-signal content. - Quote, don't paraphrase the news. Say "Anthropic released X" not "AI labs are releasing things." Specificity beats hand-waving.