growth-distill

name	growth-distill
description	成长轨迹蒸馏 / Growth-trajectory distillation. Reads the windowed historical posts scraped by collect-history.sh (one dir per half-year window), and distills the blogger's path from cold-start to now — stage breakdown + turning points + a "same-stage imitation checklist". Trigger when the user says "distill growth trajectory {handle}" / "蒸馏成长轨迹 {handle}".

Growth-trajectory distillation / 成长轨迹蒸馏

Division of labor 分工: grok/scripts = scraping (collect-history.sh grabs historical posts into collected/{handle}/{window}/); Claude (this skill) = distillation 蒸馏. Core stance 核心立场: to learn from a blogger, learn what they did when they were at your follower level, not their post-fame behavior. A big account's throwaway line today is backed by a huge following you don't have. The "same-stage imitation checklist" is this skill's core deliverable. 学一个博主，要学他在你这个量级时的做法，不是学他功成名就后的现状。

Trigger 触发

User says "distill growth trajectory {handle}" / "蒸馏成长轨迹 {handle}" / "growth distill {handle}". handle is the directory name under collected/ (set by collect-history.sh, e.g. dotey).

Input corpus is scraped by collect-history.sh into collected/{handle}/{window}/ (one dir per half-year window, verbatim post files). If that dir doesn't exist or is empty, tell the user to run the scraper first — do NOT pad with recent posts.

Steps 执行流程 (skip none)

① Explore the corpus 探索语料

List all window dirs under collected/{handle}/ (in time order), then read EVERY post in each window. Record: number of windows, N posts actually read per window. If corpus is large, parallelize (one agent per window, then merge).

② Per-window profile 按时间窗归纳 (five items, each backed by real quotes ≤2 lines + source)

1. Topic distribution 主题分布: what they mainly wrote, rough share per topic.
2. Voice 声音特征: tone / sentence shape / self-reference / signature moves — quote real lines.
3. Content form 内容形式: short posts / threads / long-form / with-image / data posts.
4. Engagement play 互动打法: who they reply to, how, do they ride trends, ask vs assert.
5. Rough follower level 粉丝量级: ONLY from in-post signals (they mention a follower count, a milestone, "just passed X"). No reliable signal → mark "unknown", never invent 绝不编.

③ Break into stages 划分成长阶段

Cluster windows into stages by shifts in content & engagement (loose framing: start / growth / mature — don't force it; two stages is fine, add a plateau if there is one). Per stage: window range, defining traits, difference from the previous stage (with contrasting real quotes).

④ Turning points 找关键转折点

Where does the tone change — topic swap? form swap? voice swap? engagement-strategy swap? Each turning point MUST have before/after evidence (one quote before + one after). No contrast quotes → write "insufficient evidence", not a turning point.

⑤ Same-stage imitation checklist 同阶段模仿清单 (core deliverable)

1. Fix the reader's cold-start level as the baseline (single digits ~ a few thousand followers).
2. Locate the blogger's same-level window/stage (from the level signals in ②; if all "unknown", align by early/late stage and note it's inferred).
3. List their concrete moves in that period, each executable:
   • What to post 发什么: topic types, topic mix, cadence (only state cadence if evidenced).
   • How to write 怎么写: length / form / opener / tone, each with a real example line.
   • How to engage 怎么互动: reply strategy, what to ride, who to interact with.
4. Every move backed by a real quote; a move with no evidence does NOT enter the checklist.

Output 输出

Write collected/{handle}/growth-distill.md with four parts:

1. Stage table 阶段划分表 (stage | window range | follower level or "unknown" | defining traits)
2. Per-stage profiles 每阶段画像 (the five items from ②, with real-quote evidence)
3. Turning points 转折点 (④, with before/after evidence)
4. Same-stage imitation checklist 同阶段模仿清单 (⑤, baseline = reader cold-start level)

Closing report 收尾汇报: how many windows read, how many posts total, which file written, whether level signals were evidenced or unknown.

Hard rules 硬规则

• Every conclusion backed by a real quote 所有归纳必须引真帖佐证; can't back it → "insufficient evidence". Distillation extracts what's there, it doesn't hallucinate a trajectory.
• No reliable in-post follower signal → "unknown", never invent 绝不编 a number.
• Don't copy long passages 不复制大段原帖: quotes ≤2 lines, shortest slice that proves the point.
• Keep quoted lines in their original language.
• Don't write any publishable copy — this skill only produces trajectory analysis.