screenwriting-skills Review 2026 — 26 Agent Skills That Turn 47 Craft Books and 23 Script Collections Into a Working Writers' Room

What screenwriting-skills Is

Ask a general-purpose coding agent to fix a scene and you get fluent, competent, generic prose: the beat lands, the characters talk, and nothing about it could only have been written by someone who had read Syd Field and John Truby and forty-five other books. The problem is not that the model cannot write. It is that it has no named method to hold itself to, so it optimises for plausible sentences instead of dramatic structure — no controlling idea, no act out, no moral argument, no scene as value turn. screenwriting-skills, created September 6, 2026 by jtydhr88, is a set of 26 Agent Skills that closes that gap by making the craft explicit and loadable.

The repository is blunt about its ambition: 26 skills “distilled from 47 craft books and 23 volumes of published scripts, scores and plays.” The books are the canonical ones — Field’s Screenplay, Snyder’s Save the Cat, McKee’s Story and Dialogue, both of Truby’s Anatomy volumes, Egri, Hoxter, Hicks, Cron, Walter, Henson — alongside a substantial Chinese craft literature (Mei Feng’s Screenwriting Self-Study, Lu Jun’s Screenwriting Theory and Technique, and Liu Dapeng’s edited Story Craft Masterclass, International Volume). The collected works range from Chekhov’s complete plays to all four seasons of Succession’s shooting scripts. At the time of writing the repository sits at 1,081 stars and 122 forks in nine days, which for a writing-craft repo is a strong signal.

Four Layers, and a Rule About What Earns a Skill

The skills are organised in four layers rather than a flat pile. Layer 1 — general dramaturgy is medium-independent: sw-workflow orchestrates a project and names which skill to call at each stage; sw-story-structure carries the paradigm and the Save the Cat board and McKee’s event/scene/sequence/act; sw-premise-theme handles controlling idea and the “third rail”; sw-character-conflict, sw-dialogue and sw-scene-craft cover the working parts; sw-format-adaptation holds spec-format hard rules and a Fountain output contract.

Layer 2 — the medium layer is what a feature does not teach: sw-series-structure covers teasers and cold opens, broadcast four/five/six-act grids with page anchors, three tests for locating the invisible acts in streaming scripts, and Oberg’s four information tools (mystery, surprise, dramatic irony, suspense). sw-series-engine-bible, sw-writers-room and sw-sitcom-comedy round it out, and the stage side gets its own serious treatment.

The design rule matters more than the list. A genre gets its own skill only if it makes the model write with different structure, format or language rules. Subject matter (spy thriller, costume drama) and directorial style do not qualify; they live as case studies inside existing skills, unless a complete script corpus exists — as with Ozu. This is why the project can state that adding the series layer (13 → 20 skills) and the opera slot left sw-story-structure unchanged, and even shortened sw-dialogue. That is a maintenance argument a lot of skill collections never make.

The Multilingual Decision — and Its Cost

The most interesting engineering choice is the one the project fought with itself over. The skills are written once, in Chinese, and the agent answers in whatever language you asked in. There was an English tree — a parallel screenwriting-en plugin, proposed, built and merged — and it was then deleted. The argument, documented in the README, is that if a reader who cannot read Chinese deserves a translated tree, so does a Japanese reader, then Korean, then French; five languages is 230 files that drift independently with nothing to say which is stale.

The runtime does the rest. Output language follows your question. Terminology is anchored back to its original term through sw-workflow/terms.md, so a concept that different Chinese translators rendered differently is restored to logline, act out, beat sheet rather than reinvented. Concepts with no equivalent come through as the original term plus a gloss — the one-line core attraction of an episode is carried as xiyan, with its meaning spelled out beside it.

It is a defensible call, and the project names its cost honestly: you cannot read the instruction file itself unless you read Chinese, only the agent’s account of it. That is exactly what the deleted English edition bought, and the maintainers decided it was not worth a permanent five-way maintenance burden.

What Using It Looks Like

Install is standard for the two majors. Claude Code: /plugin marketplace add jtydhr88/screenwriting-skills, then /plugin install screenwriting@screenwriting-skills, giving skills names like /screenwriting:sw-dialogue. Codex: codex plugin marketplace add jtydhr88/screenwriting-skills, then invoke with $sw-story-structure. The same SKILL.md files serve both because they follow the agentskills.io format and name no agent-specific syntax — each plugin carries only a thin .claude-plugin/plugin.json and .codex-plugin/plugin.json over one shared skills/ directory.

The README’s own examples show the intended granularity: “My dialogue is all on the nose, fix this scene” pulls sw-dialogue plus sw-scene-craft; “Turn this 400,000-word novel into a 40-episode Chinese drama outline” pulls sw-chinese-series-practice plus sw-series-engine-bible. You install once into ~/.claude/skills/, .claude/skills/, ~/.agents/skills/ or .agents/skills/ and the skills load when the agent detects relevant context.

The Opera Layer Is the Surprise

Where the project goes well beyond anything comparable is Chinese opera. It treats it as two vocal systems, not one genre: the qupai system (Yuan zaju, Ming-Qing chuanqi, kunqu) where a lyric is filled into a fixed tune pattern with set line counts, lengths and tones, and the banqiang system (Peking opera, Henan, Yue and Qin opera) where a lyric is built from paired seven- or ten-character lines varied by metre. Each system gets a method skill and a full-script case library, taken from public-domain editions and checked against the originals; where a scan could not be verified (some Wu Mei and Legend of the White Snake passages) the notes say so and the skills refuse to build rules on them. Four skills, 25 sources, and an explicit statement of what is not planned — short-form vertical drama and AI-generated comic drama, “whose logic is distribution, with no dramaturgy to distil.”

The Honest Limits

The constraints are as clearly documented as the features, which is why the rating lands where it does.

The license governs everything. The repository declares no SPDX license; the README says “for personal study use,” and quotations remain the property of their authors and translators. For an individual writer exploring craft with an agent, that is fine. For shipping inside a product or a paid client deliverable, it is a blocker you must clear first.

Auditability is traded away. The instruction bodies are Chinese-only. A non-Chinese-reading team adopts the project on trust — it can read the README and the agent’s output, not the principles and tables underneath.

It is young and single-maintainer. Two weeks old, and the issue history is nearly all one person’s own follow-ups: a 17-sentence README rewrite, a “13 scale fixes” pass, an opera-layer addition. The one genuinely external signal is issue #24, “Feedback: Testing screenwriting-skills in AutoClaw.” There are no tests, no schemas, no CI — quality rests on source discipline, not machine-checked guarantees.

Activation is heuristic. A 26-skill install can add context overhead, and the project ships no token-cost estimate and no scoping tool to keep only the skills a project needs.

Bottom Line

screenwriting-skills is the most carefully sourced creative-writing skill pack we have seen: named methods, primary texts, a maintenance rule that survives its own growth, and a multilingual decision the maintainers were willing to argue for and against in public. Set against a personal-study license, Chinese-only instruction files and a two-week-old single-maintainer project, it earns 8.2/10 — a genuinely strong tool for individual screenwriters and TV writers working in Claude Code or Codex, and a research reference for anyone building craft skill packs. If you write for a living and your agent keeps producing plausible, structureless drafts, install it and ask for a beat sheet; the difference in the first revision will tell you whether the source base is worth its constraints.