压缩任意来源,保留每一条论断、对冲、数值与归属。
文档
Brief
把你提供的素材整理为以读者下一步决策为中心的简报。
它能做什么
你提供素材并在需要时授权访问,简报负责结构与排版。覆盖执行简报、状态更新、会议预读、董事会/投资人通报、事件简报、交接文档与决策备忘录——每篇都以一句话结论、至多五条要点和一个明确请示开场。内容选取按决策相关性,而非按素材长度比例压缩,因此一页简报可以胜过四十页报告。配置与学到的格式偏好保存在 `~/Clawic/data/brief/`,且仅根据你每次对已交付简报的明确反馈更新。
什么时候用它
- 起草董事会或投资人更新,含指标对比计划与明确请示
- 准备会议前一天送达的预读材料
- 撰写事件通报,包含用户视角的影响、状态与下一次更新时间承诺
- 撰写决策备忘录,在多个选项之间给出推荐,并把"维持现状"列为真实可比项
技能文档
User configuration and learned preferences live in ~/Clawic/data/brief/ (see setup.md on first use, preferences-template.md for the learned-preferences file format). If you have data at an old location (~/brief/ or ~/clawic/brief/), move it to ~/Clawic/data/brief/.
When To Use
- User asks for an executive brief, TL;DR, or bottom-line summary of material they provide
- A recurring status or project update is due and the reader needs the delta, not a restate
- A meeting, board session, or review is coming up and the reader needs decisions-needed plus prep
- An incident is live and stakeholders need impact, status, and a committed next-update time
- A transition, offboarding, or vacation cover needs a handoff with gotchas and open questions
- The reader must choose between options and needs a recommendation with tradeoffs
- A long report, thread, transcript, or doc pile must be condensed to what drives a decision
Not for: compressing content when there is no decision or action to serve — use summarizer for shorter-content-only needs.
Quick Reference
| Situation | Play |
|---|---|
| Reader must choose between options | Decision brief: recommendation first, 2-3 options including do-nothing (templates.md) |
| Recurring status update | Project brief: lead with the delta since last brief, never a full restate (recurring.md) |
| Meeting coming up | Meeting brief: decisions needed + prep checklist; deliver the day before, not the morning of (delivery.md) |
| Board or investor update due | Board brief: metrics vs plan, lowlights before highlights, explicit asks (templates.md) |
| Something is on fire right now | Incident brief: impact in user terms, status, committed next-update time (templates.md) |
| Findings must become a decision | Research brief: answer with confidence level, evidence split from interpretation (templates.md) |
| Transition, offboarding, vacation cover | Handoff brief: gotchas and open questions outrank achievements (templates.md) |
| Source is huge or messy | Triage by decision-relevance, never proportionally to source length (sources.md) |
| Sources disagree or have gaps | Show the conflict and name the gap — never average or silently narrow (sources.md) |
| Audience unclear or mixed | Name the decider; their action sets depth and jargon (audiences.md) |
| Formal channel (exec email, external doc) | Same structure, strip emoji markers, plain headers (delivery.md) |
| Brief keeps coming back "too long" | Cutting passes, hedge blacklist, decision-grade rounding (writing.md) |
| User reacts to a delivered brief | Record it in ~/Clawic/data/brief/preferences.md; signal mapping in dimensions.md |
| Anything else | Executive structure: bottom line, 3 key points, explicit ask |
Depth on demand: templates.md section-by-section structures per type · sources.md raw-material triage, conflicts, gaps · audiences.md reader calibration · delivery.md channels, register, timing · writing.md line-level compression · recurring.md running a brief series · dimensions.md preference taxonomy · setup.md first-use loading.
Core Rules
1. User Specifies Sources
When the user requests a brief:
- User provides the information OR specifies where to get it
- If a source requires access, the user grants it explicitly
- Skill structures and formats the output
Example:
User: "Brief me on project X status"
Agent: "I'll need access to the project docs. Can you share
the status doc or grant access to the project folder?"
User: [shares doc or grants access]
→ Brief generated from user-provided source
2. A Brief Is Not a Summary
A summary compresses content; a brief serves the reader's next action. If you cannot name the decision or action this brief enables, ask "what will you do with this?" before writing. Competent people summarize; briefers select.
3. Write in Reverse Reading Order
Compose ask → bottom line → key points → context; the reader consumes the opposite order. Writing front-to-back is how ledes get buried: you discover the point last and leave it there.
4. Bottom Line Commits
One sentence, no hedge. If it contains "and", you have two takeaways — keep the one that changes what the reader does, demote the other to key points. "Mostly on track, but..." is not a bottom line; it is a decision you refused to make.
5. Three Key Points, Five Hard Max
A point qualifies only if removing it would change the reader's decision. Six "key" points means you haven't decided what matters — cut, or tier the rest under a "more detail" note. This selection rule beats source length: one decision-relevant page outweighs forty pages of background.
6. Every Number Carries a Comparator
"$1.2M revenue" is decoration; "$1.2M, 8% under plan" is a brief line. Valid comparators: target, prior period, benchmark. None available → write "no baseline yet" or drop the number.
7. Status Words Have Definitions
- On track — schedule/scope buffer intact
- At risk — buffer being consumed; the first slip counts, no grace period
- Blocked — cannot proceed without a named external action
Declaring at-risk early is cheap. A green-to-red flip with no warning destroys trust in every future brief you send. (status_scheme: rag in config maps these 1:1 to Green/Amber/Red.)
8. Learn Only From Explicit Feedback
"Too detailed" / "missing X" / "perfect" → one line in ~/Clawic/data/brief/preferences.md with a level (pattern/confirmed/locked, promotion rules in dimensions.md). Check the file before writing any brief — a well-built brief in the wrong learned format still misses.
Scope
This skill:
- ✅ Structures information the user provides into briefs (act-as writer: it drafts; the user sends)
- ✅ Learns format preferences from explicit feedback
- ✅ Stores configuration and preferences in
~/Clawic/data/brief/
User-driven model: the user specifies WHAT information to include and grants access to any needed sources; the skill handles STRUCTURE and FORMAT.
This skill does NOT:
- ❌ Access files, email, or calendar without user request
- ❌ Pull data from sources the user hasn't specified
- ❌ Store brief content (only configuration and format preferences)
Output Gates
Run before delivering any brief:
- Bottom line survives alone — if the reader stops there, they still have the takeaway
- Every metric has a comparator or an explicit "no baseline yet"
- The ask names an owner and a date, or states "no action needed" — never an implied ask inside an FYI
- Bad news sits above the fold, next to its mitigation, not in section four
- Data that can go stale carries source and as-of date
- Conflicting sources are shown as a conflict, never averaged into one number
- Jargon calibrated to the least-technical decider, not the most technical reader
- Format matches
~/Clawic/data/brief/config.yamlandpreferences.md— checked, not remembered
Configuration
User-dependent variables. Defaults apply until the user states a preference; store them in ~/Clawic/data/brief/config.yaml.
| Variable | Type | Default | Effect |
|---|---|---|---|
| default_length | one-screen | one-page | two-page | one-page | Sizing target: one-page = 450-500 words (templates.md); one-screen = fits without scrolling (chat cap, delivery.md); two-page only when the reader asked for depth |
| status_scheme | words | rag | words | Status labels everywhere a status renders: On track/At risk/Blocked vs Green/Amber/Red (definitions in Rule 7 apply to both) |
| emoji_markers | bool | true | Section markers (⚡📊🎯) on informal channels; false = plain headers everywhere (formal channels strip them regardless, delivery.md) |
| default_channel | chat | email | doc | chat | Assumed delivery surface; sets register and length caps per delivery.md |
| locale | text (BCP-47, e.g. en-US) | en-US | Date order (Jul 21 vs 21 Jul), decimal/thousands separators, and currency symbol + placement in every rendered number and date (writing.md); falls back to ~/Clawic/profile.yaml before the default |
Preference areas — customizable dimensions; a stated preference gets recorded in config.yaml and applied:
- Audience mix: who reads most briefs (boss, board, clients, team) — shifts the default template and jargon level (
audiences.md) - Conventions: section names, bullet style, status-word extensions — shifts the structures in
templates.md - Timing: lead times, cadence, delivery day — shifts the scheduling guidance in
delivery.md - Exclusions: topics or metrics never to include (confidential, legal) — screens extraction in
sources.md - Voice: bad-news directness, person (I/we/team), formality floor — shifts the register rules in
writing.md - Localization: beyond the
localetag — timezone for timestamps and deadlines, and multi-currency policy (show native vs converted, which symbol) when amounts span currencies — affects number lines inwriting.mdand the incident/board timestamps intemplates.md
Config vs learned preferences: config.yaml holds what the user declared; preferences.md holds what feedback revealed (levels in dimensions.md). A stated preference that names a config variable goes straight to config.yaml. Precedence: config > confirmed/locked learned preferences > defaults. Universal variables (locale, timezone) additionally fall back to ~/Clawic/profile.yaml before their table default: config > profile.yaml > default.
Traps
| Trap | Why it fails | Do instead |
|---|---|---|
| Burying the lede | Readers triage; the point placed last is the point never read | Bottom line first — and write it first (Rule 3) |
| Hedged bottom line | Reader cannot act on a hedge | Commit to one takeaway; nuance goes in key points |
| Watermelon status (green outside, red inside) | The eventual flip costs more trust than early honesty ever would | First buffer slip → At risk (Rule 7) |
| Summarizing proportionally to source length | Background volume drowns the decision material | Select by decision-relevance (Rule 5, sources.md) |
| Averaging conflicting sources | Reader inherits a number nobody actually reported | Show both figures with provenance; the conflict may be a key point (sources.md) |
| Sandwiching bad news | In a brief, softening reads as hiding | Bad news first, mitigation beside it |
| Options without a recommendation | "Neutral" presentation transfers your analysis work to the reader | Recommend one; keep tradeoffs honest and visible |
| Exactly two options | Frames a false binary and hides the status quo | Price do-nothing as an explicit option (templates.md → Decision) |
| One brief for decider and spectators | Too deep for one, too shallow for the other | Write for the decider; spectators get the same brief, not a longer one (audiences.md) |
| Restating unchanged items every edition | Trains readers to skim; they miss the week something changes | Delta only, unchanged sections collapse to one line (recurring.md) |
| Precision theater ($1,234,567.89) | False confidence and a slower read | Decision-grade rounding: 2-3 significant figures (writing.md) |
| Emoji markers on formal channels | Reads unserious to exec/external audiences | Same structure, plain headers (delivery.md) |
Where Experts Disagree
- Bullets vs narrative memos. The Amazon six-page school argues prose forces complete thinking; the BLUF school argues readers triage and bullets respect that. Boundary: match the reading ritual — narrative where the room reads silently before discussing; bullets where the brief replaces the meeting. Default here: bullets, with a prose-memo variant in
templates.md. - Recommendation-first vs neutral staffing. Staff tradition separates information briefs from decision briefs and warns against advocacy; operators say a briefer who won't recommend transfers the analysis work. Boundary: recommend by default; go neutral only when brokering between peer factions — and say you're doing it.
- How much context. "Context is respect for the reader" vs "context is where ledes go to die." Boundary: context earns lines only when it changes how the reader acts on the points; SCQA-test it (
templates.md), otherwise cut.
Related Skills
More Clawic skills, get them at https://clawic.com/skills/brief (install if the user confirms):
summarizer- compression when there is no decision to serve, just shorter contentdigest- recurring curated updates pulled from external sources on a schedulereport- recurring configured reports with fixed data sourcesmeetings- the full meeting system (notes, agendas, follow-ups); brief covers only the pre-read
Feedback
- If useful, star it: https://clawic.com/skills/brief
- Latest version: https://clawic.com/skills/brief
Part of Clawic, the verified skill library. Get this skill: https://clawic.com/skills/brief.
相关技能
把 YouTube 视频整理成带章节、时间戳和要点的 Markdown 摘要
为单篇或一批 ArXiv 论文下载源文件与 PDF,再通读全文并按指定语言生成结构化的 summary.md。
用你自己的语气起草、修改、重写邮件、帖子、备忘录等各类文稿。
把内容转成适合阅读和分享的幻灯片图片,提供多套样式预设,可合并为 PPTX 或 PDF。
Generates a batch arXiv report by combining model-written templates with brief conclusions and abs URLs extracted from per-paper summary.md files.