[REPLACE: SKILL_NAME]

${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

1. 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.

2. Augment with web search — run a WebSearch for [REPLACE: TOPIC] latest and pick up to 5 fresh links published in the last 24h that aren't already in the feed results.

3. 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).

4. Pick the top [REPLACE: MAX_ITEMS] — write output/articles/[REPLACE: SKILL_NAME]-${today}.md with one entry each:

   ### [Title](url)
   *[Source · published date]*
   2-3 sentences distilling the takeaway. No filler.
   

5. Notify via ./notify with:

   *[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
   

6. 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.