Community生產力與協作github.com

Rajbharti06/sr-growth-loop

執行讓頻道隨著時間持續優化的學習循環:撰寫假設、改變單一變數、記錄每次貼文的結果,並透過秩和置換檢定判斷是否採用、繼續測試或放棄。

sr-growth-loop 是什麼?

這項技能建立了一個結構化的成長迴圈,幫助創作者或行銷人員透過嚴謹的 A/B 測試逐步改善內容表現。其核心流程為:從先前的稽核中選出一個問題、在發佈前撰寫假設、每次僅改變一個變數(如鉤子、第一幀、標題、縮圖、長度、字幕或發文時間)、每種變體至少發佈四篇貼文、記錄結果,最後透過無母數的秩和置換檢定(rank-sum permutation test)決定該變更是否值得採用、需要更多資料或應予放棄。此技能支援 Instagram、YouTube 等平台,並可直接在本地機器上運作,透過 `experiment_new`、`experiment_log`、`experiment_review` 等工具指令完成整個實驗流程。它還包含校準鉤子評分器的功能(`calibrate_run`、`write_weights`),以及將學習成果儲存至個人檔案(`profile_save`)的機制,讓後續計畫能參考過往經驗。此技能特別適合用於回答「該測試什麼?」、「這項變更有效嗎?」、「如何逐步改善?」等問題,或建立一套固定的成長 routine。

相容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/Rajbharti06/showrunner-agent-skills/tree/HEAD/skills/sr-growth-loop

Installed? Explore more 生產力與協作 skills: steipete/gemini, steipete/gh-issues, steipete/skill-creator · View all 6 →

在你喜歡的 AI 中提問

開啟一個已預先載入此 Agent Skill 的新對話。

說明文件

成長迴圈

成長是一個迴圈:猜測、測試、讀取、保留有效的。這項技能讓迴圈保持誠實。

迴圈步驟

  1. 從上次稽核中選一個問題(sr-audit)。
  2. 在發佈前寫下假設。「如果第一行提到數字,更多人會停留超過三秒。」
  3. 改變一個變數。 盡可能使用相同平台、相同格式、相同主題組合。其他一切保持不變。
  4. 每種變體至少發佈四篇。 少於此數無法進行測試,工具會明確告知。
  5. 記錄每次的結果。
  6. 審查。 採用它、繼續測試,或放棄它。
  7. 將學到的經驗儲存到個人檔案,以便後續計畫能參考。

執行方式

python scripts/sr.py exp new --hypothesis "A number in the first line improves 3-second hold" --variable "first line" --metric "3-second hold %" --platform instagram
python scripts/sr.py exp log <experiment-id> A 41 --ref "reel 12 May"
python scripts/sr.py exp log <experiment-id> B 58 --ref "reel 14 May"
python scripts/sr.py exp review <experiment-id>

作為工具:experiment_new、experiment_log、experiment_review。變體由使用者命名(A 和 B,或「數字」和「無數字」)。

判定結果的運作方式

審查會將所有記錄的貼文排序,並使用秩和置換檢定比較兩種變體,該檢定不假設觀看次數的分佈。白話來說:它會問,隨機打亂標籤多少次才會出現您所觀察到的那麼大差距。它還會報告兩種變體中位數的比率。

有四種結果:

  • 「X 看起來比較好……採用它並繼續記錄。」 p 值小於 0.1,且一種變體的中位數至少是另一種的 1.2 倍。
  • 「貼文數量不足。」 任一變體的貼文數少於四篇。
  • 「沒有明顯差異……延長測試,或測試更大的改變。」 差距很小或 p 值很高。
  • 「至少需要兩種變體。」 只有一種變體被記錄。

每種變體四篇貼文時,只有很大的差距才會顯現出來。請明確告知。在樣本數小時,「沒有明顯差異」的結果代表「我們無法判斷」,而非「這不重要」。0.1 和 1.2 的閾值是維護者的選擇,並非標準。變體會依名稱的字母順序比較,因此請命名為 A 和 B。

優先測試什麼

按改變能影響數字的程度排序:鉤子和第一幀,然後是標題或封面,接著是長度,再來是字幕第一行,最後是主題角度。發文時間排在最後,且僅限於使用使用者自己的分析工具。對於 YouTube,可在 Studio 中使用「測試與比較」功能測試縮圖,自 2025 年 7 月起也可測試標題(已有回報,youtube.test_and_compare;此功能可能並非每個頻道都有)。

校準評分器

在發佈 30 篇以上包含鉤子文字的貼文後,執行 calibrate_run。如果鉤子評分能將此帳號的熱門貼文與表現不佳的貼文區分開來,且區間超過 50%,則 write_weights 會將調整後的權重儲存到本地資料夾。如果不行,請更相信您自己的測試而非評分。

儲存經驗

當測試被採用時,在個人檔案的 lessons 清單中新增一行(使用 profile_save),例如「5 月的數字在第一行有助於提高 3 秒停留率(每組 n=5)」。寫下您的發現,並附上樣本數。儲存時請傳送整個清單,因為儲存會合併欄位並整體取代清單。

閘門

在開始(假設)和結束(採用或放棄)時需要 approve。

禁止事項

  • 一次改變兩件事。
  • 因為看起來不錯而提前結束測試。
  • 在沒有數字和樣本數的情況下報告「有效」。

Individual skills in this repo

This repo contains 9 individual skills — each has its own dedicated page.

Rajbharti06/showrunner

Conductor for growing a channel on Instagram, YouTube, TikTok, Facebook, LinkedIn, X or Threads. Use when someone wants to start, grow, fix or plan social content, asks what to post or where to post, wants a content strategy, calendar or weekly routine, or mentions views, followers, reach, Reels, Shorts, carousels or building a creator business. First asks what they want to make, for whom and where, recommends a hub platform and spokes sized to their real weekly hours, then runs playbooks that chain the sr-* skills with approval gates. It writes and plans. It never posts.

Rajbharti06/sr-edit-plan

Turns a video and its transcript into an edit plan. Reads an SRT or VTT file to report speaking pace, dead air to cut and filler words with timestamps, re-chunks subtitles into short readable lines for vertical video, and writes a cut list, b-roll and on-screen text notes, and export settings. Use when someone asks how to edit a video, wants captions or subtitles, wants to remove pauses and ums, or says their edit feels slow. Produces instructions and a caption file to apply in CapCut, Premiere, DaVinci, Descript or any editor. Nothing edits the video itself.

Rajbharti06/sr-engage

Handles the replies, comments, DMs and outreach that happen after a post goes up. Drafts replies in the creator's voice for them to post, sorts comments into questions, praise, critique and spam, turns repeat questions into content ideas, plans a small daily engagement block, and handles criticism and trolls calmly. Use when someone asks how to reply to comments, grow through engagement, handle a hater, build community, or what to do in the first hour after posting. Drafts only. Nothing is sent or posted by Showrunner.

Rajbharti06/sr-hook

Writes and scores hooks, the first line, first three seconds or first frame that decides whether anyone keeps watching or reading. Drafts ten options in different shapes from 32 common structures, scores them on specificity, front-loading, who it speaks to, curiosity, stakes and brevity, fixes the weakest part, and returns the best three. Use for Reels, TikToks, Shorts, YouTube openers, LinkedIn or X first lines, and when someone says their videos get swiped away or asks "is this a good hook".

Rajbharti06/sr-platform-guide

A sourced, dated field guide to YouTube, Instagram, TikTok, Facebook, LinkedIn, X and Threads. For each platform it gives formats and limits, how people find content, what it reports to creators, what the creator controls, and what nobody outside the company knows, with every fact labelled official, reported, inferred or study and dated. Use when someone asks how a platform works, what the algorithm rewards, which format to use there, or whether a rule they heard is true. Checks the facts database instead of repeating folklore.

Rajbharti06/sr-repurpose

Turns one long video, podcast, talk or article into a week of native pieces. Scores a transcript for the best 20 to 60 second windows on opening hook, self-containment, density and payoff, writes ready-to-run ffmpeg commands for 9:16 cuts with matching caption files, then adapts each piece for each platform with a new first frame, platform captions and no foreign watermarks, plus text and carousel spin-offs. Use when someone has long content and asks for clips, Shorts, Reels, a content week or "how do I get more out of this video".

Rajbharti06/sr-script

Writes and times scripts for short vertical video and long YouTube videos. Structures the hook, the promise, the beats with open loops, the payoff and the loop back, writes for the ear in the creator's voice, then times it and flags a hook that runs long, beats that sit still, filler runs, an early call to action and a missing loop. Use when someone needs a script, a talking-points outline, a voiceover, a faceless video, or asks "how long will this run" or "why do people leave my videos".

Rajbharti06/sr-seo

Finds the phrases people actually search for and puts them in the right places on YouTube, TikTok, Instagram and Google. Builds a keyword table with search intent and placement (title, first line, spoken opening, on-screen text, description, chapters), checks titles and captions against the target phrase, and uses a keyword connector such as vidIQ when one is connected. Use when someone asks about SEO, keywords, search, discoverability, hashtags, or why a good video is not being found. Showrunner has no search-volume data of its own and says so.

Rajbharti06/sr-text-post

Writes text posts and threads for LinkedIn, X, Threads and Facebook. Builds a first line that survives the preview cut, one idea per post, a structure that fits each platform, and sensible link and tag placement. Use when someone asks for a LinkedIn post, a thread, a tweet, a Threads post, a build-in-public update, or wants to turn a video or article into written posts. Keeps the creator's voice and flags what a script cannot fix.

相關技能