Use this skill when the user wants a daily briefing generated directly from Syft CLI plus the agent's own model.
-
Check Syft availability.
Run syft status.
-
Read following topics with ids.
Run syft following.
Preserve both:
- the topic title
- the topic id
Topic ids matter because they are the bridge from a user's declared interest list to syft top --topic-id.
-
Pull a first-pass global pool.
Start with:
syft top --days <N> --limit <M> --rich
-
Do pool triage before writing anything.
Split results into:
keep: structurally important or clearly high-matchmaybe: potentially useful but not yet cleardiscard: promo, coupon, generic shopping, shallow affiliate content, obvious noise
-
Normalize multilingual titles before selection.
If titles are in multiple languages, first translate or paraphrase them into one working language for your own reasoning.
Do not let language fragmentation hide a strong event cluster.
-
Build an event-level pool, not an article pile.
Merge duplicates when they describe the same event, actor move, or causal development.
Preserve factual anchors during the merge.
-
Build an interest coverage map before final selection.
Use following_topics.md plus profile_summary.md to list:
- main structural axes
- identity-defining hobby / culture axes
- thin but still important side interests
For each axis, ask:
- is there a strong chainable storyline
- is there at least one high-salience item even if it does not chain well
- is the pool empty, weak, or noisy
-
Repair obvious coverage gaps with topic top before broad search.
Use this retrieval ladder by default:
- global top
- topic top
- targeted search
If the first-pass global top pool is weak on a clearly important axis, do not jump straight to search.
First use the topic id from syft following and pull:
syft top --topic-id <TOPIC_ID> --days <N> --limit 20 --rich
Topic top is usually the most efficient second-pass repair because it already carries topic-aware recall.
Start with limit 20.
In most topics, the top 20 items already carry enough signal for editorial recovery.
Only expand beyond 20 if:
- the line is still clearly thin
- the returned 20 are dominated by duplicates or low-value items
- the topic is known to have a dense event cadence and still needs more chronology
Use it for:
- explicitly followed niche interests
- hobby lines that global top often misses
- narrower structural lines that matter to this user but not to the whole platform
Only after global top plus topic top still leave a real gap should you add 2 to 5 targeted syft search queries.
Search is the third layer, not the default second layer.
-
Select for editorial value with profile awareness.
In Syft-only environments, use these layers in priority order:
- following topics as the canonical declared-interest source
- profile summary as the high-level interpretive lens
- confirmed guidance rules
- optional heuristic
raw_interest_block.txt
- optional explicit aversion notes, only if the user later supplied them
Do not assume you have user-authored scores. Most of the time you will not.
-
Judge importance relative to the interest domain, not only by universal news value.
For macro, finance, geopolitics, and war:
- importance usually comes from structural consequence, causality, repricing, or regime stress
For hobby, fandom, collecting, and travel:
- importance may come from local cadence inside that interest world
- official collab drops
- merch or peripheral releases
- exhibition or event announcements
- creator statements
- store / pop-up / cafe / campaign cycles
- seasonal ryokan / onsen / local experience openings or promotions
Example:
- a new EVA collab may be globally minor but highly important inside EVA fandom
- a Fate merch release can be a real event if the line being watched is official content cadence
-
Match prose voice to the lane.
Do not write every line in the same analytical register.
Use:
- macro / market / geopolitics:
restrained, causal, analytical, briefing-like
- ACG / fandom:
more scene-aware and community-native
allowed to sound warmer, more release- and cadence-sensitive, less like a policy memo
- photography:
enthusiast-facing, concrete, product- and collector-aware
- travel / onsen:
texture-rich, experience-aware, less boardroom-analytical
Example:
- weak:
GU 联名说明 EVA 仍在向大流通服饰端扩散
- better:
EVA 这轮已经不只是零星周边，而是连 GU 和伊势丹线都在接力，30 周年的商业联动热度明显还在往上走
The goal is not roleplay or slang overload.
The goal is that each lane sounds like it was written by someone who actually knows that audience.
-
Keep titles disciplined.
Headlines should look like publishable titles or section names, not like the model talking to itself.
Avoid:
- overlong explanation jammed into the title
- sentence fragments that sound like live note-taking
- staged phrasing that sounds like commentary rather than editing
-
Write natural Chinese.
The body must read like edited Chinese prose.
Avoid:
- translationese
- awkward oral filler
- shorthand note fragments
Example:
- weak:
《序》CGI 制作秘话再被翻出来讲
- better:
《序》的CG制作内幕又被周年节目重新带热
-
Separate the edition into:
今日新闻池概览主线副刊 / 兴趣侧翼兴趣雷达 / 未成线但值得看本轮新闻池覆盖不足
-
Keep bodies publishable.
The article body must read like news briefing prose, not instructions to the user.
Do not write things like 对你来说 or 你需要知道.
Use them as scaffolding, but adapt the section logic to the specific user's profile.
