screenwriting-skills: 26 Agent Skills for Screenwriting, TV and Dramaturgy
Professional agent skills for screenwriting, television writing and dramaturgy
At a glance
- What is it?
- A set of SKILL.md files for Claude Code and OpenAI Codex that encode craft knowledge from 47 books and 23 volumes of scripts. The install is simple; the trade-off is that the instructions themselves are written in Chinese.
- Who is it for?
- Adopt it if you already work inside Claude Code or Codex and want craft structure (premise, act outs, engine bibles, revision passes) available as invocable skills rather than pasted prompts. Do not adopt it if you need to read and edit the instruction files yourself and do not read Chinese, or if you need a skill the repository does not ship, such as prose fiction or game narrative.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 16, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What screenwriting-skills actually packages
This is not a screenwriting application. It is a directory of 26 SKILL.md files that an agent loads as context, plus the packaging needed to distribute them. The README describes them as "distilled from 47 craft books and 23 volumes of published scripts, scores and plays (Chinese, American, British, Japanese and Korean)."
The audience is narrow and specific: writers who already run Claude Code or OpenAI Codex and want the agent to reach for a named craft procedure instead of improvising one. A novelist with no interest in act structure will find nothing here. A staff writer breaking a season will find the series stage table (S0 through S7, from engine to submission) more useful than a general prompt.
The skills are organised in four layers. Layer 1 is medium-independent dramaturgy: premise, structure, character, dialogue, scene, format, Truby organic anatomy, genre anatomy, workflow. Layer 2 is the medium layer, and it is where the project stops being generic: series, episode and season structure, engine and bible, writers' room, half-hour comedy, stage, and Chinese opera with banqiang and qupai methods and full-script case libraries. Layer 3 covers tradition and trade (America, Japan, Korea and France, mainland China, the business). Layer 4 is master corpora: Chekhov, Ozu, Succession, television case studies.
The README is explicit that a feature project uses layers 1, 3 and 4, while a series or stage project adds layer 2, "because the medium layer replaces structure-for-film with the medium's own units." That is a real architectural claim, not decoration: the film structure skill is not the one you want when you are placing act outs.
How the skill files reach the agent
The mechanism is the open agentskills.io standard, which both Claude Code and Codex read. The README states the SKILL.md files "are shared by both agents: install once and it works in both." There is no server, no API key, no build step. The files sit in plugins/screenwriting/skills/ and get copied into an agent's skills directory, or installed through each agent's plugin marketplace.
Invocation differs by agent. In Claude Code the skills are namespaced as /screenwriting:<skill>, so dialogue work is /screenwriting:sw-dialogue. In Codex you invoke one explicitly with $sw-story-structure. The README also notes that skills "load automatically when the agent detects relevant context," which means the explicit command is a fallback, not the only path.
The interesting design decision is the language split. The skill bodies are written in Chinese, and the README gives the reason plainly: most of the sources are Chinese originals or Chinese translations. Output, however, follows the question. Ask in French and the answer comes back in French, with no flag and no separate install.
Terminology is handled by a mapping file rather than by translation. The README points out that the field's craft vocabulary is originally English (logline, act out, beat sheet, showrunner, staff writer), and that Chinese translations of those terms diverge between translators. The file sw-workflow/terms.md maps each concept back to its original term so the agent restores the word instead of inventing one. Terms with no equivalent keep their original form plus a gloss; the README gives 戏眼, 扣子, ト書き and 決定稿 as examples, rendered as `戏眼 (xìyǎn, the one-line core attraction of an episode)`.
One convention is worth noting because it affects day-to-day use: sw-workflow defines a story-bible.md file that keeps project state in one place across sessions. That is the project's answer to context loss, and it only works if you actually maintain the file.
Installing screenwriting-skills in Claude Code or Codex
The recommended path in both agents is the plugin marketplace. In Claude Code the README gives two commands, first adding the marketplace and then installing the plugin:
/plugin marketplace add jtydhr88/screenwriting-skills
/plugin install screenwriting@screenwriting-skillsAfter that, skills are invoked as /screenwriting:<skill>. The README's example is /screenwriting:sw-dialogue.
Codex uses a different command to register the marketplace, and the install itself happens in the app or CLI:
codex plugin marketplace add jtydhr88/screenwriting-skills
# Then, in the ChatGPT desktop app or Codex CLI:
# Plugins → select "Screenwriting Skills" → InstallThe README states the same 26 SKILL.md files ship through the plugin's skills/ directory "with no duplication and no rewriting."
If you prefer to manage files yourself, both agents accept a manual copy. For Claude Code, clone the repository and copy the skills into your personal directory so every project sees them:
mkdir -p .claude/skills
cp -r screenwriting-skills/plugins/screenwriting/skills/* .claude/skills/For Codex the target directory is ~/.agents/skills/ instead of ~/.claude/skills/. To scope the skills to one project, use .claude/skills/ or .agents/skills/ inside that project rather than in your home directory.
Verification is the same idea in both agents. Run /skills to list what loaded. In Codex you can also invoke a skill directly with $sw-story-structure.
A first real use, taken from the README's own examples: ask the agent to break a 12-episode series into acts and place the act outs. The README says the agent uses sw-series-structure for that. A second example covers the case where the problem is prose rather than architecture: "My dialogue is all on the nose, fix this scene" routes to sw-dialogue plus sw-scene-craft. Neither requires configuration beyond installation.
The Chinese source tree is the price of the install
The README is unusually candid about what the single-source-tree decision costs: "you cannot read the instruction file itself unless you read Chinese, only the agent's account of it." That is the main limitation, and it is not cosmetic. When a skill produces something you did not want, you cannot open the SKILL.md and see which instruction caused it. You are debugging through the agent's summary of its own instructions.
The README also records that an English edition existed. It was proposed in issue #3, built, merged, and later removed. The stated argument is that forking once removes any principled ground for refusing a second fork, and that five languages would mean 230 files drifting independently with nothing to indicate which is stale. Whatever you think of that reasoning, it is a deliberate choice with a visible cost, and the README names the cost rather than hiding it.
The READMEs themselves are translated into Chinese, Japanese, Korean and French, but the non-Chinese ones "deliberately stop at install, structure and this policy." The full skill tables and the source bibliography live only in this README and the Chinese one. So a French reader gets install instructions and a policy statement, not the skill inventory.
Two other boundaries are worth stating. The repository has no releases, so there is no versioned artifact to pin; you are tracking the main branch. And the skills cover screenwriting, television and dramaturgy. If your work is prose fiction, game narrative or copywriting, the layers here do not map onto it, and the medium layer in particular assumes episodic or stage units.
Where this differs from a general-purpose writing prompt
The obvious alternative is not another skill pack. It is the thing most writers already do: keep a document of craft prompts and paste the relevant one into the chat. The difference in approach is real. A pasted prompt is one block of instructions written by you, covering the one problem you anticipated. A skill is a file the agent selects, and the selection is driven by what you asked rather than by what you remembered to paste.
The project's own example makes the distinction concrete. Asking the agent to turn a 400,000-word novel into a 40-episode Chinese drama outline routes to sw-chinese-series-practice plus sw-series-engine-bible. That is two skills, and the pairing is not something a writer would necessarily think to assemble. The workflow skill exists precisely to name which skill to call at each stage, with per-stage deliverables and advisory exit checks.
The second difference is the terminology table. A general prompt drifts toward whatever synonym the model prefers. The terms.md mapping pins the vocabulary to the original English craft term, which matters when the same Chinese word has been translated three ways across three source books.
The third difference is the story-bible.md convention, which gives project state a fixed home rather than leaving it scattered across a conversation. That is a convention, not a feature, and it only holds if you keep the file current.
Maintenance status, licence and the upgrade question
The repository is not archived, and the last push was on 2026-09-16. There are no releases, which means there is no tagged version and no changelog to read before upgrading. If you install through the plugin marketplace, the practical upgrade path is whatever the marketplace does when the source repository changes. If you copied files manually, updating means re-running the copy command, and any local edits you made to a SKILL.md would be overwritten.
The licence is MIT, and the repository carries a separate NOTICE file at the top level alongside LICENSE. MIT is permissive, but the NOTICE file is worth opening before you redistribute anything, since it typically carries attribution requirements that sit outside the licence grant itself. That is a reading task, not a legal opinion.
One maintenance fact follows from the language decision. Because there is a single source tree, there is no translation drift to manage, and a fix to a skill lands everywhere at once. The corresponding risk is that the source tree is the only tree, so a change to a skill's instructions is not reviewable by anyone who does not read Chinese. The README treats this as the accepted trade.
Editorial conclusion
Adopt it if you already work inside Claude Code or Codex and want craft structure (premise, act outs, engine bibles, revision passes) available as invocable skills rather than pasted prompts. Do not adopt it if you need to read and edit the instruction files yourself and do not read Chinese, or if you need a skill the repository does not ship, such as prose fiction or game narrative. Before committing, open plugins/screenwriting/skills/sw-workflow/terms.md and confirm the terminology mapping matches the vocabulary you actually use, then run /skills to check the 26 entries load.
Frequently asked questions
What skills do screenwriters need?
The repository does not answer this as a general question about the profession. It answers it as an inventory: 26 skills across general dramaturgy (premise, structure, character, dialogue, scene, format, genre anatomy, workflow), a medium layer for series, stage and Chinese opera, a tradition and trade layer, and master corpora for Chekhov, Ozu and Succession.
What are the 5 basic writing skills?
The README does not group its skills into five basics. It organises them into four layers, and it states that a feature project uses layers 1, 3 and 4 while a series or stage project adds layer 2 because the medium layer replaces structure-for-film with the medium's own units.
What are the top 10 screenwriting tips?
The README does not contain a tips list. What it contains is procedural: sw-workflow defines a stage map from premise through structure, character, scenes, draft, revision and submission, with a separate S0 to S7 table for series, naming which skill to call at each stage.
Is screenwriting being replaced by AI?
The README takes no position on that question. It describes the skills as craft procedures the agent invokes, and it states that your script stays in your script's language: the conversation switches language, the draft does not.
Community notes