Agent Skills
› mohitagw15856/pm-claude-skills
› newsletter-digest-brief
newsletter-digest-brief
GitHub将杂乱邮件订阅整合为兴趣过滤简报，提取关键信息并附来源，合并重复内容。提供被忽略项统计以建立信任，支持按节奏生成周报或日报，取代日常碎片化阅读。
触发场景
digest my newsletters
summarize what my subscriptions said this week
what did I miss that I actually care about
make my reading pile useful
安装
npx skills add mohitagw15856/pm-claude-skills --skill newsletter-digest-brief -g -y
SKILL.md
Frontmatter
{
"name": "newsletter-digest-brief",
"description": "Turn a pile of newsletters and subscriptions into one skimmable brief — the items that matter to YOUR interests extracted with sources, the noise dropped with a count, on a cadence that replaces daily trickle-reading. Use when asked digest my newsletters, summarize what my subscriptions said this week, what did I miss that I actually care about, or make my reading pile useful. Produces the interest-filtered brief with per-item sources, the dropped-with-reasons ledger, and the cadence that makes trickle-reading obsolete."
}
Newsletter Digest Brief Skill
Newsletters are individually reasonable and collectively unreadable — twelve arrivals a week, each 70% padding, each interrupting on its own schedule. The digest move: batch them (the inbox-unsubscribe-purge filter routes them to a label), then process the pile into one brief filtered by the reader's actual interests — items extracted with their source and why-it-matters, everything else dropped with a visible count so the reader trusts the filter instead of re-reading behind it.
What This Skill Produces
- The brief — the items that matter to this reader's stated interests: one line of what + one of so-what + the source link
- The dropped ledger — "skipped 31 items: 12 promos, 9 repeats of the X story, 10 off-interest" — the trust mechanism
- The repeat-collapse — one story covered by four newsletters appears once, best-source-first
- The cadence setting — weekly for most; daily only for genuinely fast-moving beats
Required Inputs
Ask for these if not provided:
- The pile — the newsletters' content (pasted/forwarded), or the label they collect under
- The interest profile — the beats that actually matter to this reader ("AI tooling, pricing strategy, my competitors, nothing about funding rounds") — the filter IS the product, and it needs edges
- The action bias — reading-for-awareness vs. hunting-for-actions; the brief leads with actionable items when the latter
Framework: The Digest Rules
- Filter by declared interest, ruthlessly: the brief includes what matches the profile and drops the rest — a digest that summarizes everything is the pile again, shorter. Edge cases get one "borderline" line at most, not inclusion.
- What + so-what + source, one item = two lines: "Vercel shipped edge-side A/B testing (what). Relevant: replaces the tool you're paying $400/mo for (so-what). [source]" — the so-what is written for this reader, which is what separates a digest from a table of contents.
- Collapse the echo: the same announcement in four newsletters is one item, attributed to the best treatment ("deepest take: X's issue"). The echo-count is itself signal ("covered everywhere = the week's consensus story") and gets a line when heavy.
- The dropped ledger builds the trust: the counted skip-list ("12 promos, 9 echoes, 10 off-interest") is what lets the reader not re-read the pile — a filter with invisible drops never earns retirement of the source reading.
- Cadence follows the beat's clock: weekly suits almost everything; daily digest is for beats where a day's lag costs (markets, security advisories). The cadence is chosen once and defended — an "urgent" exception lane exists only for the reader's pre-named topics.
Output Format
Digest: [period] — [N] sources, [M] items kept, [K] dropped
Actionable (if any)
[Items implying a move for this reader, first]
Worth Knowing
[Per item: what — so-what for you. ([best source])] [Echo notes where heavy: "covered by 4 of your sources"]
Borderline (one-liners)
[…]
Dropped: [K]
[The counted categories — promos / echoes / off-interest]
Quality Checks
- Every kept item has a reader-specific so-what, not a generic summary
- Repeats are collapsed with the best source named
- The dropped ledger is counted and categorized
- Actionable items lead when the reader's bias is action
- The interest profile's edges were honored — no scope creep toward "interesting generally"
Anti-Patterns
- Do not summarize everything — an unfiltered digest is the pile with extra steps
- Do not write so-whats that would fit any reader — the profile is the point
- Do not hide the drops — invisible filtering keeps the reader re-reading the sources
- Do not let one loud newsletter dominate — items compete on interest-match, not sender volume
- Do not run daily by default — cadence inflation recreates the interruption problem the digest exists to solve
版本历史
- 471c606 当前 2026-07-22 14:04
