AI news intelligence and daily briefing powered by CellCog. News digests, competitive intelligence, market updates, trend monitoring, industry reports, current events research. Multi-source synthesis for accurate, comprehensive briefs.
设计与多媒体
Digest
围绕你关注的话题、人物和行业,做周期性新闻摘要:搜集、筛选、排序并送达。
它能做什么
围绕你关心的话题、关注的公司和行业,输出周期性新闻摘要。每期按「搜集 → 筛选 → 排序 → 排版 → 送达 → 学习」六步走,自动去重、核实单源消息,并用条目上限保持精炼。每条都附带标题、一行摘要、对你的意义和来源署名。格式、时间、深度会随你的反馈逐步贴合。
什么时候用它
- 搭一个每日早间行业简报
- 每周追踪竞争对手动态
- 假期或断更后一次性补齐近期要闻
- 收紧一份总漏新闻或越读越长的旧摘要
技能文档
User preferences, learned signals, and the sent-item log live in ~/Clawic/data/digest/ (see setup.md on first use, preferences-template.md for file formats). If you have data at an old location (~/digest/ or ~/clawic/digest/), move it to ~/Clawic/data/digest/.
Configuration
User-dependent variables. Defaults apply until the user states a preference; store them in ~/Clawic/data/digest/config.yaml.
| Variable | Type | Default | Effect |
|---|---|---|---|
| item_cap | number (3-15) | 8 | Hard ceiling on items per digest (rule 5); Highlights stay ≤3 at any cap |
| single_source_policy | hold | hedge | hold | Market-moving or reputational single-source claims: hold one cycle vs ship marked "one source, unverified" (verification.md) |
| channel | text | channel where the digest was first requested | Where digests deliver; drives the per-channel format rules in delivery.md |
| digest_language | text | the language the user writes in | Digest prose language; items keep original-language names and quotes with translated summaries |
| timezone | text | ~/Clawic/profile.yaml, else confirm before first scheduled send | Anchors every scheduled slot and recency window |
Preference areas — learned state lives in ~/Clawic/data/digest/preferences.md and climbs the ladder in learning.md:
- content scope — topics, industries, competitors, people, regions, exclusions; drives Filter
- sourcing — trusted and blocked sources, weighting, recency windows; drives Source
- format & structure — medium, structure, length, tone, visuals; drives Format
- timing & cadence — schedule, frequency, weekday/weekend split, urgent-signal rules; drives Deliver
- prioritization & depth — what leads, what buries, per-topic depth; drives Prioritize
When To Use
- The user wants a recurring daily or weekly briefing on topics they care about
- A news roundup filtered to one person's interests, not a generic feed
- Tracking competitors, an industry, people, or markets with ongoing updates on their moves
- A scheduled topic update is due, or the source list needs building or tuning
- An existing digest feels off: too long, repetitive, missing stories, wrong timing
- Not for: synthesizing documents the user hands you (→
synthesize), internal business reporting (→brief), or full competitor dossiers (→competitor-monitoring)
The test of a real digest: every item carries a why-this-matters-to-you line. A digest that could be sent to anyone is a feed, not a digest.
Quick Reference
| Situation | Play |
|---|---|
| First digest, no preferences | Derive topics from the request; if none derivable, ask the one blocked question (topics) — everything else runs on defaults |
| Story covered by 3+ independent sources | Promote toward Highlights — convergence is an importance signal (rule 4) |
| Surprising claim from a single source | Hold one cycle or ship hedged per single_source_policy; independence test in verification.md |
| Major global news outside user's topics | One line in Worth Noting, never a Highlight — relevance beats magnitude (rule 1) |
| User says "too long" | Cut item count before per-item depth; log the signal (learning.md) |
| Story already shipped in a prior digest | Resurface only as "Update:" + what changed — check the sent log (rule 7) |
| Urgent item mid-cycle | Interrupt only on a [confirmed] or [locked] urgent-signal rule; otherwise hold (delivery.md) |
| No reaction for 3 consecutive digests | Ask one concrete question instead of guessing (learning.md) |
| Slow news day in their topics | Send the short honest version ("quiet day, 2 items") — never pad |
| Big-event day floods one theme | Cluster the theme into one block that counts as one item (triage.md) |
| Competitor activity to report | Moves, not mentions — the move catalog is in competitors.md |
| Gap since last digest (vacation, downtime) | One consolidated catch-up covering the period, top items only (delivery.md) |
| Anything else | Run Source → Filter → Prioritize → Format → Deliver → Learn with current preferences |
Depth on demand: sources.md building and vetting the source list · triage.md dedupe, ranking, slow and heavy days · verification.md claims, rumors, corrections · writing.md item craft and summaries · delivery.md channels, timing, interrupts · learning.md signals and the preference ladder · competitors.md tracking companies and people · dimensions.md dimension catalog with defaults · preferences-template.md data-file formats.
Core Rules
- Relevance beats magnitude. Rank by interest match first, freshness second, source weight third. A tracked competitor's pricing change outranks a world-news headline outside their topics.
- Every item is attributed. No source you can name = item does not ship. Never blend your own inference into a summary unmarked — prefix it: "Likely:" or "My read:".
- Exclusions are binary, interests are weighted. One explicit "don't care about Y" beats any number of inferred interest signals for Y. An excluded topic appears only when it intersects a confirmed interest, flagged as such ("normally excluded; included because it hits X").
- Dedupe before ranking. Same story from N outlets = one item; N itself is rank input — 3+ independent sources covering one story promotes it toward Highlights. Independence = separate original evidence, not separate URLs (
verification.md). - Fixed budget forces curation.
item_cap(default 8) items, hard max 3 Highlights. Adding one over cap means deleting one. Worked example: 12 candidates → 3 Highlights (top interest-match), 5 body items, remaining 4 dropped or compressed into one Worth Noting line. - Preference changes climb a ladder — pattern → confirmed → locked, mechanics in
learning.md. Never jump from a single signal to locked. - The digest remembers what it sent. Log every shipped item to
~/Clawic/data/digest/sent-log.md; repeats resurface only as "Update:" + delta. A digest without memory re-reports yesterday and reads as a broken filter.
Protocol
Source → Filter → Prioritize → Format → Deliver → Learn
- Source — pull from configured feeds and the per-entity source kits (
sources.md). Default weighting: primary > original reporting > aggregators > social; social is discovery, never citation — trace it back before it ships. - Filter — topic match includes; exclusion match drops (rule 3); recency windows per content type (
dimensions.md→ Sources); single-source surprising claims go toverification.md. - Prioritize — cluster duplicates, then rank per rule 1; bury borderline items into Worth Noting, don't delete (
triage.md). - Format — per-item shape: headline → 1-line summary → why it matters to this user → (source). Craft rules in
writing.md; length pressure cuts item count before per-item depth. - Deliver — channel and timing per preferences; interrupts, catch-ups, and per-channel formats in
delivery.md. - Learn — capture reactions into
preferences.mdafter every delivery; signal counting and ladder inlearning.md.
Output Format (Default)
📰 [DIGEST TYPE] — [DATE/TIME]
🔥 HIGHLIGHTS
• [Item — 1-line summary + why it matters to you (source)]
• [Second item]
📋 FULL DIGEST
[Items in per-item shape, ordered per weighting profile]
💡 WORTH NOTING
[Borderline items, one line each]
---
Sources: [count] | Next digest: [time]
Adapt entirely to learned preferences once they exist; the template above is only the cold-start default.
Output Gates
Before sending, verify:
- Every item has a source and a why-it-matters line?
- ≤3 Highlights, each strong enough that the user could read only those?
- Zero exclusion-list items, unless flagged per rule 3?
- Every single-source claim hedged or held per
single_source_policy? - Nothing repeated unchanged from the sent log?
- Channel and time match current preferences; signals from the previous delivery recorded?
Traps
| Trap | Why it fails | Do instead |
|---|---|---|
| Leading with globally important news | User already saw it everywhere; the digest adds zero | Lead with niche relevance; big news gets one line (rule 1) |
| Padding slow days | Trains skimming; erodes trust in Highlights | Short honest digest ("quiet day") |
| Counting syndicated copies as corroboration | Ten outlets rewriting one press release or tweet are one source | Independence test in verification.md before promoting |
| Trusting aggregator timestamps | Aggregators resurface old stories with fresh dates | Check the original dateline against the event date (sources.md) |
| Locking a preference from one comment | One signal may be mood, not preference | Climb the ladder in learning.md (rule 6) |
| Reading silence as satisfaction | Ignored digests look identical to loved ones | No reaction for 3 consecutive digests → ask one question |
| Unmarked opinion inside summaries | User can't separate reporting from your read | Attribute claims; prefix inference (rule 2) |
| Resurfacing yesterday's story unchanged | Repetition reads as a broken filter | Only resurface as "Update:" + what changed (rule 7) |
| Interrupting off-schedule on your own judgment | "Urgent" without a learned rule is noise at a bad hour | Interrupt only per [confirmed]+ urgent-signal rule (delivery.md) |
Where Experts Disagree
- Curation vs completeness. Tight curation is the default — the budget is the product. Some readers want everything as FOMO insurance; the boundary is an explicit request, and the answer is a longer Worth Noting tail, never a bigger Highlights section.
- Relevance-first vs breaking-first. Default: highest interest-match leads. Breaking-first is legitimate for users who act on time-sensitivity (markets, on-call, PR); it must be learned or stated, never assumed from topic choice.
- Neutral summarizer vs opinionated curator. The why-it-matters line is opinion, and it is the digest's value. The boundary: reporting is attributed, inference is prefixed — the reader always knows which is which (rule 2).
Related Skills
More Clawic skills, get them at https://clawic.com/skills/digest (install if the user confirms):
brief— internal business information summarized rather than external news curatedsynthesize— the user provides the documents; you distill instead of sourcing the outside worldcompetitor-monitoring— full competitor dossiers, pricing alerts, and positioning analysis beyond digest itemsnewsletter— the user wants to publish to an audience, not receive a personal digest
Feedback
- If useful, star it: https://clawic.com/skills/digest
- Latest version: https://clawic.com/skills/digest
Part of Clawic, the verified skill library. Get this skill: https://clawic.com/skills/digest.
常见问题
- 每期摘要包含多少条?
- 默认最多 8 条(区间 3–15),其中 Highlights 不超过 3 条。多加一条就得删一条,预算本身就在驱动筛选。
- 只有一家媒体报道的消息怎么处理?
- 按你的「单源策略」执行:要么压一期再发,要么标注「单一来源、未经核实」发出;社交媒体线索一定会追溯到原始来源才会发布。
- 摘要会学习我的偏好吗?
- 会。你的反应会沉淀进话题、信源、时间、深度等偏好;信号按「模式 → 确认 → 锁定」阶梯上升,单次评论不会直接锁定。
相关技能
Turn any industry into a daily intelligence briefing. An AI agent searches, filters, writes, and delivers structured daily briefs to 9 channels — with machine-checked formatting and a business review gate. Ships with a Data+AI profile out of the box; switch to any domain via config.
Produce polished, production-ready daily news briefings, morning reports, market snapshots, executive digests, and founder updates with verified sourcing, cl...
汇总过去 30 天 Reddit、X、YouTube 和网页上关于某个话题的真实讨论。
通过托管 OAuth 连接,管理 beehiiv 出版物、订阅者、帖子、自定义字段、用户分群、订阅分层与自动化流程。
压缩任意来源,保留每一条论断、对冲、数值与归属。