Commentary#Coding agent#Content quality
Stop writing talking-head scripts from scratch: 5 open-source skills that cover hooks, storyboards and retros
Five talking-head script skills verified by writing stage: script checks, blind retros, verbatim drafts, storyboards, hook polish. Stars and licenses checked.
Talking-head scripts stall in the same three places: the first-3-seconds hook, the pain-point resonance in the middle, and the call to action at the end. A Chinese WeChat article aimed at creators (“自媒体不会写口播稿？用大博主的 skill！！”, October 29, 2026) argues you shouldn’t fight through them alone — install the methods top creators have already distilled as agent skills, and let the Agent write against them. It names five GitHub repos. We verified all five: star counts re-checked via the GitHub API on November 4, 2026, licenses confirmed by opening each LICENSE file or README, and — more useful than either — which stage of script writing each one actually covers. They are not five copies of the same tool; they split the work: script checking, topic retros, verbatim drafts, storyboards, and hook polishing.
Key points
- All five install into Claude Code, Codex and similar Agents, but their licenses differ sharply: one standard MIT, two CC BY-NC 4.0 (no commercial use), one all-rights-reserved, and one whose README badge says MIT while the repo ships no LICENSE file.
- Scale runs from 9 skills (rayskills) to 61 top-level skill directories (rnskill); dbskill’s method base comes from the author’s 16,152 public tweets, distilled into 4,176 knowledge atoms.
- The stages compose: cheat-on-content owns topic selection and post-publish retros, rayskills or creator-buddy owns drafts and storyboards, and rnskill’s ra-hook specializes in openings.
| Repo | Author | Stars | License | Best for |
|---|---|---|---|---|
| dbskill | dontbesilent | 11,147 | All rights reserved | Full-script checks and content diagnosis |
| cheat-on-content | 蜗牛学长 | 7,456 | MIT | Blind predictions and publish retros |
| rayskills | Ray Wang | 382 | CC BY-NC 4.0 | Verbatim drafts and a full publish pipeline |
| creator-buddy | 空格的键盘 | 1,800 | See below | WeChat/ Xiaohongshu / video, three platforms |
| rnskill | 雪踏乌云 | 1,713 | CC BY-NC 4.0 | Opening hooks and delivery polish |
dbskill: check the whole script
dbskill is the broadest of the five: 34 Chinese skills that install into Doubao, WorkBuddy, Claude Code and Codex, all reachable through a single /dbs entry point. The script-relevant abilities: turn a topic into a content plan, check which lines fail to land, write video titles, cover text and openings, and run a compliance check before publishing. The “check the finished script” job the article emphasizes is fully covered — wordiness, broken transitions, lines that read awkwardly aloud are all on its checklist. Its differentiator is the method base: the author dontbesilent distilled his own 16,152 public tweets into 4,176 knowledge atoms, and the skills’ judgments come from that corpus. Current version is v3.2.3; 11,147 stars as of November 4, 2026 (the article said 10,928). Watch the license: the README states that as of this update the collection is all-rights-reserved — free to install and use for your own business and client work, but no republishing, bundling, selling, or AI training use.
cheat-on-content: the retro loop
cheat-on-content (MIT) owns the post-publish stage. Its core loop: score and blind-predict a piece before publishing → publish → retro against real data at T+3 days → fold the conclusion back into your scoring rubric. The README puts the problem plainly: a creator with 200 posts is barely sharper than one with a single post, because they never kept books after each round. It suits opinion and personal-observation pieces — the article positions it as pulling topics and first drafts from your own experience, then letting your data, not someone else’s viral template, evolve the judgment. Current version v0.1.0; 7,456 stars as of November 4 (7,385 in the article), and the only standard-MIT repo of the five.
rayskills: 9 skills, one pipeline
rayskills is Ray’s set of 9 Chinese skills (with a /ray router) covering the full path from writing to publishing: Chinese long-form, translation pieces, talking-head scripts and science-narration drafts; covers; explainer videos; and final layout saved as drafts to WeChat Official Accounts and X Articles — it stops at the draft box, never auto-publishes. The /ray router distinguishes three situations: pick only the next step when you’re unsure, run the whole pipeline when stages are clear, and stop to confirm before any destructive or publishing action. Script writing goes through /ray-kb: it retrieves material as needed, generates a verbatim script in the author’s voice, and lists uncertain claims in a separate “to confirm” section instead of blending them into the text. At 382 stars as of November 4 it’s the smallest of the five, but the most complete stage by stage; CC BY-NC 4.0 means check the terms before commercial use.
creator-buddy: three platforms
creator-buddy, by “空格的键盘” (SpaceZephyr), splits into three boards: gzh-Skills (WeChat Official Accounts, 11 skills from positioning to layout), xhs-Skills (Xiaohongshu, a 10-step account-building track plus a controller), and video-Skills (a space-video controller dispatching topic, script, editing, B-roll, subtitles, voiceover/music and cover). The two the article names are script-facing: space-video-topic writes topics and body copy, and space-video-script polishes spoken lines and arranges storyboards — built for tool intros and screen-recorded tutorials. Its positioning line is “from one topic to a publishable piece”, pulling real platform data back and keeping the content yours. 1,800 stars as of November 4. One thing to nail down: the README badge says MIT, but the repo currently ships no LICENSE file (the GitHub API returns none either), so confirm with the author before commercial use. It follows the Agent Skills standard and works with skills.sh, installing into Claude Code, Codex, Cursor, OpenClaw and Hermes.
rnskill: 61 skills of ammunition
rnskill, maintained by “雪踏乌云” (Pluviobyte), collects 61 top-level skill directories (compatibility aliases included) in two broad groups — content creation and coding. The creation side spans topic planning, script rewriting, AI voiceover, digital humans, video editing, subtitles, covers, production scheduling and commercial diagnosis. The two the article names are ra-hook, which picks and writes opening hooks, and ra-人话, which rewrites stiff lines into natural spoken Chinese — the ones to install when you only want to fix openings and delivery. The author’s README claims that combining these skills with Codex, HyperFrames, HeyGen and IndexTTS2 took under 10 hours a month, gained 2,000 Douyin followers, and landed a first sponsorship deal (a single-author claim, take it as anecdotal). The repo carries a sponsor section, so expect promotional content around the install instructions. Licensed CC BY-NC 4.0; 1,713 stars as of November 4.
How to pick
Pick by the stage you’re stuck at. If the draft is done but you don’t trust it, dbskill checks wordiness and broken transitions — accepting its all-rights-reserved terms. If you publish without any sense of the numbers, cheat-on-content’s blind prediction plus T+3 retro is the only bookkeeping loop among the five, and it’s standard MIT. If you want one pipeline from draft to draft-box, rayskills is the most complete and the smallest to read. If you publish across WeChat, Xiaohongshu and video, creator-buddy’s controllers save the most coordination, pending a license confirmation from the author. If only the opening and delivery need work, rnskill’s ra-hook and ra-人话 are enough on their own. All five repos saw pushes between November 2 and 3, 2026 and are actively maintained; stars measure attention, while license fit and stage fit decide the choice.
Two costs come with combining them. All five are written for Chinese-language content creation: prompts, sample scripts and check rules target Chinese spoken-word delivery, so English-speaking creators will get limited value — filter on this first. Install style matters too: dbskill’s 34-at-once install fills an Agent’s skill list, and the more common practice is installing per task — rayskills for writing weeks, cheat-on-content for retro weeks — rather than keeping the full bundle resident; with many skills loaded, the Agent has to pick from a pile of SKILL.md files every run, and a wrong pick costs more than one fewer install.
Getting started
dbskill and rnskill print runnable install commands in their READMEs:
# dbskill: install all 34 skills at once
npx -y skills add dontbesilent2025/dbskill -g --all
# rnskill: install one skill at a time (ra-hook owns opening hooks)
npx -y skills add Pluviobyte/rnskill --skill novel-distill
Install routes for cheat-on-content, rayskills and creator-buddy follow their READMEs (rayskills also mirrors on Gitee for mainland China). The shortest path for a talking-head script: blind-score the topic with cheat-on-content, run the finished draft through dbskill’s checks, then polish the opening with rnskill’s ra-hook. Adjacent stages on this site: video-shotcraft covers storyboards and finished edits, and startup-skill covers validating the business behind the content — writing, filming and commerce each have their own installable skill.