Awesome Short Drama Skills + short-drama-skill-finderPartly tested🇨🇳 CN

Underleveled Builder's curated list of short-drama skills: 83 skills checked by hand, plus two published rounds of Opus / GPT blind tests against a no-skill control

Search GitHub for short-drama skills and you get dozens of repos with no way to tell which ones actually write shootable scripts. Underleveled Builder (@UnderleveledDev) maintains Awesome Short Drama Skills (~752★, CC0): 83 hand-checked short-drama / manju skills plus 16 related open-source platforms, sorted by type, production stage and video model. The list is itself a skill (short-drama-skill-finder)—install it and ask your agent which skill to use for novel-to-drama on Codex. Rarer still: the author published two bake-offs. Round 1 put 10 screenwriting skills plus a no-skill control on the same open brief, three runs each, judged blind by Opus and GPT. Round 2 ranked 11 premise skills the same way. Full logs, outputs and judge fingerprints live in evals/ and on shortdramaskills.com. We only read the README, the evals method and some output folders; we did not re-run the Claude Code harness.

Visit site Awesome Short Drama Skills + short-drama-skill-finder ↗
Pricing
Free
Region
Global
Where it's from
Chinese community
Added
Oct 25, 2026
Tested on
Oct 25, 2026
Version we used
awesome-short-drama-skills main (meta fetched_at 2026-10-23); evals Round 1 2026-10-18, Round 2 2026-10-20
Categories
VideoWriting
Original author
Underleveled Builder @UnderleveledDev ↗
Our test
Partly tested
Cost / time
The list and finder skill are free (CC0). Running short-drama skills or reproducing evals needs your own Claude Code, Codex or other skill-capable agent budget; one Round-1 contestant (manju-laoli) logged about $72.93 across three runs. Video generation (Seedance / Kling / MiniMax, etc.) is separate.
Source
X @UnderleveledDev(回复 @wlzh 贴出仓库) ↗X @GitHub_Daily(长文介绍清单与两轮实测) ↗X @wlzh(转发精选清单) ↗GitHub UnderleveledBuilder/awesome-short-drama-skills ↗shortdramaskills.com(清单站与实测页) ↗

The task

Pick short-drama / manju skills that actually write shootable scripts or usable premises from a crowded GitHub field, and compose installs by stage (novel adaptation, episode hooks, storyboards, Seedance / Kling / MiniMax prompts).

Steps

  1. Open the list: https://github.com/UnderleveledBuilder/awesome-short-drama-skills or https://shortdramaskills.com (bilingual).
  2. Start with the "Which skill to install for scripts" and "…for premises" sections: those rankings come from Claude Code + claude-opus-5-5 blind bake-offs, not from star counts.
  3. Browse the rest by production stage or video model. "Checked by hand" means the README and classification were reviewed, not that every entry was bake-off tested—only the ones in evals/ were.
  4. Let your agent choose: run `npx skills add UnderleveledBuilder/awesome-short-drama-skills` to install short-drama-skill-finder, then ask e.g. "I use Codex to adapt novels into short dramas—which skill should I install?" It fetches https://shortdramaskills.com/data.json (or falls back to the README).
  5. Install a recommended skill, e.g. `npx skills add zenstory-ai/drama-skills`. A few repos (like 0xsline/short-drama) lack a SKILL.md name/description header and need a git clone into the skills folder—see the list notes.
  6. Read evals/README.md and the Round 1 / Round 2 write-ups on the site for the limits (one writing model, reviewer disagreement on hooks, metadata tables in drafts that you must strip before production).
  7. Read each third-party SKILL.md before installing: the author notes skills may ship scripts.

Prompt

I use Codex to adapt novels into short dramas. Using the Awesome Short Drama Skills list, recommend which skills to install and give the install commands.

What the author says

The author and amplifiers like GitHubDaily say the list holds 83 skills + 16 platforms and picked up hundreds of stars in about 11 days. Round 1 (2026-10-18): 10 screenwriting skills vs a no-skill control, same open brief, 3 runs each—zenstory-ai/drama-skills, Shanyin screenwriting master, jtydhr88/screenwriting-skills and adoria111/emma-screenwriting-skills beat the control 3/3 on both scripts and character bibles per Opus and GPT; Hao0321/ai-short-drama swept character bibles. Round 2 (2026-10-20): 11 premise skills pairwise ranked—Tomsawyerhu/Chinese-WebNovel-Skill, lingfengQAQ/webnovel-writer and Hao0321/ai-short-drama were the shared top three and beat the control 3/3. Limits are stated plainly: Opus often preferred the control's episode-end hooks, only Opus wrote the drafts, mid-rank disagreement between judges was large, and three skills still finished below the control (9th) in Round 2.

Our test

We did not re-run the Claude Code harness (needs a logged-in claude CLI and paid usage). We did verify the README / README.en, evals/README.md, Round-1 review.md, skills/short-drama-skill-finder/SKILL.md, and that evals/runs contains per-skill output/transcript folders; the list was ~752★ when we checked on 2026-10-25. Win/loss tables are the author's published record, not recomputed by us.

What we found when we tried it:The public repo matches the write-up: the finder skill fetches data.json; evals ships run.py / judge.py plus raw outputs and blind packets; the Round-1 top four in review.md match the author's X / GitHubDaily summaries.

How to get in

GitHub and shortdramaskills.com are public; the detailed eval write-ups are mainly in Chinese (README.en covers the rankings). The finder and most skills use standard SKILL.md and install into Claude Code, Codex, OpenClaw and several Chinese agents.

Who it suits

People making AI short dramas / manju who have a novel or a one-line premise and need stage-based skill picks instead of starring whatever is trending; also useful as a reference for how to bake-off agent skills with a control and blind judges.

Pitfalls

  • "Checked by hand" ≠ "bake-off tested": only skills named in evals/ went through the blind runs; the rest are reliably classified, not ranked.
  • Drafts were written only with claude-opus-5-5; another model may rank differently. Judges are also AIs, and they disagreed on roughly a third of pairs.
  • Most skill outputs embed IDs, duration tables and checklists—strip them before feeding a video model.
  • Episode-end hooks and paywall beats are not where skills dominate: Opus often preferred the control's hooks in Round 1; paywall placement often favored the control too.
  • Model defaults are sticky (recurring character names, mother-in-law revenge tropes); swapping skills does not erase them—you still edit.
  • Some repos cannot use `npx skills add` (missing SKILL.md headers), or are 180+ days stale / archived—the list marks them.
  • Third-party skills may ship scripts: read SKILL.md before install; do not treat them as plain Markdown prompts.

More in this category