What It Is
DirectorSkills (author: geegl) distills 8 Hollywood filmmaking textbooks into 98 skill files callable by AI agents — 630K characters of Chinese skill text covering the chain from screenplay to final cut. Each skill directory contains a SKILL.md and a companion test. SKILL.md follows a six-part structure (RIA-TV++):
| Section | Content | |---|---| | R | Original quote, ≤150 characters, with chapter/page citations | | I | Rewritten methodology skeleton | | A1 | ≥2 in-book case examples | | A2 | Trigger scenarios + "language signals" | | E | Steps + stopping conditions | | B | Boundaries + failure modes + author blind spots |
Book yield: Save the Cat (8), Mackendrick's On Film-making (15), Kenworthy's Shot by Shot volumes (26), Katz's Film Directing Shot by Shot (8), Blain Brown's Cinematography (7), Van Hurkman's Color Correction Handbook (11), Bays' Master Shots (9), Sonnenschein's Sound Design (13) — 97, plus 1 cross-book shot-anchoring library, totaling 98.
Spot-checking the highest-scored skill ("think about the ending before writing the opening," from Mackendrick), even edges like "exploratory writing doesn't fit this principle" are documented. The B-section "author blind spots" — teaching the AI where a methodology fails — is rare granularity among prompt libraries.
The Pipeline
Commit history records the full flow: PDF page-slicing → visual model transcription → whole-book comprehension → five parallel extraction passes → triple validation (rejected candidates kept with reasons) → RIA-TV++ drafting → Zettelkasten interlinking → trigger/anti-trigger testing → 98 skills. Filtering was real: from 72 story-design candidates, 31 passed validation and only 8 survived. Non-generalizable or idiosyncratic techniques were explicitly rejected.
The "100% Pass Rate" Problem
Two conflicting reports exist in the repo. An older one shows 659 tests at 56.9% overall, with trigger recall at only 33% (sound design: 2%). The root JSON shows 659/659, which the README cites. The gap was closed by rewriting test cases ("de-templating") and adding "book-domain isolation" to the router: a skill "passes" if it ranks top-3 *within its own book*. The author re-ran the full suite locally: 670/670 — reproducible, not faked. But a failed query from the old report (dialogue drowned by sound effects in a mix) ranked the correct sound-design skill third globally, behind two irrelevant skills. The router — keyword matching plus TF-IDF — didn't get smarter; the threshold moved. Similarly, "Darwin audit average 8.9" reflects optimizing toward scoring rubrics (Goodhart's law). Real usage needs agent-side skill discovery (e.g., Claude Code).
Copyright Discrepancies
The README claims the repo contains no original book content. Three mismatches: a 193KB transcript remnant of the Color Correction Handbook's Chinese edition with original section numbers; extensive verbatim excerpts in the candidates directory; and a historical config file containing 7 Google Drive links to source PDFs (files deleted, but links remain in public git history). The code is MIT-licensed; the books are not.
Upstream and Follow-up
The project credits two upstream tools: the six-part format from cangjie-skill (9,618 stars) and quality auditing from darwin-skill (5,893 stars). It's a strongly executed downstream application rather than an invention. The author's follow-up repo, directormind, shifts toward evidence-driven rules that must pass forward testing — implicitly a critique of the first project's unverified distillation.
Worth and Not Worth
Worth: the granularity of individual SKILL.md files (language signals, stopping conditions, author blind spots); the "non-generalizable means rejected" validation taste; and the most complete public demonstration of a book → agent-skill pipeline, with artifacts, rejection rates, and reproducibility at every step.
Not worth: the routing layer is a toy; scores were gamed; none of the 98 skills has been validated on real productions. Valuable as a methodology reference; add your own retrieval and verification before wiring it into an agent.
Caveats
Per-skill Darwin scores were not committed, so the "average 8.9" cannot be independently verified; the Chinese source of the quotes (translation vs. model-translated) is undocumented; and the repo is internally inconsistent about whether the rubric is 9- or 10-dimensional. Data snapshot: 2026-09-07. Routing tests re-run locally; other figures from repo files and the GitHub API. Star counts are dynamic.
---
References: geegl/directorskills · kangarooking/cangjie-skill · alchaincyf/darwin-skill · geegl/directormind