Produce 8-12 complete draft candidates written in the "warm fellow traveler" style for thesocial observation / life philosophyvoice (hereafter voice C, theme configurable), save them to a queue file and display them in the session. You only do final review: pick 1-3, revise the flesh and blood, and manually copy-paste to publish.
Scope boundary: this pipeline only produces voice C. Any voice that depends on your real actions and numbers (such as tech field tests / going global to make money and other topics requiring first-hand results) cannot be quality-guaranteed by machine and must be produced by you personally;not within the scope of this skill. Voice C is pure philosophy / observation,does not need real evidence, and can be generated directly into complete drafts based on the topic library + knowledge base — this is exactly why it can be mechanized.
Fixed assets (must-read every time)
These assets are placed under your content pipeline directory; if the corresponding files exist in your environment, they are must-reads; if not, build them on the spot as described below.
- Style card:
style/STYLE.md— voice single focus, four motifs, four genres, hard metrics, tone, red lines - Topic library：
style/topic-bank.md— several topics × four motifs + hook templates + real seeds; this batch prioritizes drawing from here - Few-shot library:
style/fewshot-bank.md— take samples of the same genre - Benchmark account data (optional): scraped samples from a high-like philosophy account — map of high-like angles (borrow the form, not the content)
- Historical posts:
data/tweets-history.jsonl— basis for deduplication + your real stance - Past batches:
queue/*.md— batches from the last 7 days must be deduplicated
Process
1. Topic selection (draw from the library, no trending-topic collection)
- Main source：
topic-bank.md, take several from each of the four motifs (long-termism · delayed gratification / disenchantment · seeing as equals / small wins · stable inner core / altruism · generosity in dealing with the world), with balanced coverage. - Extension: when the library's topics are exhausted or you want new angles, refer to the benchmark account's topic map + the model knowledge base, and generate new topics by motif (still pure philosophy, no pretext needed).
- Inspiration parameter: the text the user includes after /ray-tweet is the day's observation,highest priority, and should be made into candidates first.
- Altruism motifhas historically been written about the least (it is a topic gap); deliberately ensure ≥2 per batch.
2. Generation (8-12 complete drafts)
- Read the few-shot / topic seeds before starting to write.
- Four-genre rotation: assertive aphorism style (mainstay, two-beat close) / first-principles deconstruction style (signature genre, "the first principle of X is A, not B") / definitional progressive style ("the essence of X is Y") / observation → insight style.
- Two length tiers: ≤50-character short assertions, or 150-280-character structured medium-long; medium-long proportion ≥1/3.
- Voice C needs no pretext— pure philosophy / observation can be written directly, no need to open with "saw this / someone sent me". Only when a topic is tied to your real experience (
topic-bank.mdinternal annotation[需要你的真实案例]), only then leave a placeholder for you to fill in; the machine does not make things up. - Each entry is labeled: motif / genre / length band.
Red lines (violators are discarded)：
- Do not fabricate first-person experiences, numbers, income, or costs. Pure philosophy is abstract assertions and general observations,not disguised as your specific experiences; "I actually tested / I earned Xw" belongs to a voice that depends on real operations, and voice C does not include it.
- Temperature red line: a warm fellow traveler, eye-level, giving confidence; not condescending, not harsh. Coldness to the point of criticism is another cold voice line, not this one. End on encouragement / an actionable next step; stacking two golden lines is forbidden.
- "Not… but…" for the whole batch, each entry ≤1 time; no "baobao / jia ren men" Taobao tone; write 𝕏, not Twitter / 推特.
- Check for duplication against historical tweets + the past 7 days' batches: the same assertion must not appear repeatedly.
3. Quality check
- Write the candidates as JSON
[{id, text, voice:"C-体裁"}]. If your pipeline has a lint script (such asscripts/lint.js), runnode scripts/lint.js <file>for mechanical error checking; issues must be cleared to zero (warns depend on the situation). - Self-review the model twice: ① voice match score(does it resemble the "warm fellow traveler" voice, 1-10) ② AI-flavor score(10 = no AI flavor). If either is <7, rewrite or discard. Reviewersignore
[需要你的真实案例]placeholders。 - If multi-agent fan-out is available in the current session, you may dispatch an independent reviewer agent to do adversarial quality checking instead of self-review.
4. Delivery
- Write to disk
你的内容管线目录/queue/YYYY-MM-DD.md, format as below. - In the session, display as "Today's Top 3 recommendations → the rest grouped by motif"; give one line of recommendation rationale for each.
- End with a reminder: after publishing, for those that perform outstandingly / flop, manually record them for the playbook (before data feedback is connected).
Queue file format
# 声部 C 候选批次 YYYY-MM-DD
## ⭐ 今日推荐
### 1. [母题·体裁|长度带] 一行推荐理由
(全文)
...
## 按母题分组
### n. [母题·体裁|长度带]
(全文)
Iteration discipline
For each piece of feedback you send back, judge whether it is a one-off or a pattern: pattern → immediately change STYLE.md / topic-bank.md / this file, not just the draft. Goal: calibrate the style card to a batch pass rate ≥ 1/3.