A film school your agent attends before it writes a single prompt

What it does

visual-skills packages two Claude Skills: video writes AI video prompts like a director/screenwriter/editor, image writes image prompts like an art director. Each picks the right model for the job, applies its exact syntax (Seedance, Kling, Veo, Nano Banana, GPT Image), and returns a copy-paste-ready prompt. The core is video/references/dramaturgy.md — film grammar compressed into rules an agent can execute on 5–30 second clips.

The interesting bit

The thesis is that model syntax is worthless without dramaturgy. The skill bans words like cinematic, epic, and he is sad outright — each is a placeholder for a detail the writer failed to invent. Instead it enforces seven laws: a scene formula (desire + obstacle + geometry + gaze + rhythm), three physical facts per shot, Walter Murch’s Rule of Six for cut priority, Fincher/Spielberg/Kurosawa staging rules, and beat maps for 15–90 second pieces. Output is gated twice — a six-point dramaturgy check and a three-detail audit per shot; fail either and the prompt doesn’t ship.

Key highlights

• Dedicated model files with exact syntax for Seedance (through 2.5), Kling (through 3.0 Turbo/Omni), Veo 3.1, Nano Banana 2, and GPT Image 2.5
• Runway, Luma, Pika, and Sora covered by a universal-rules layer
• Multiple output formats: single prompt, stitched multi-clip sequences with continuity blocks, storyboard tables, prompt audits, director treatments, Veo JSON
• Task modules for storyboards, animatic keyframes, race-and-speed grammar, genre patterns, and camera/lighting vocabulary
• Works with Claude Code, Claude Agent SDK, Cline, Cursor, OpenCode, Windsurf, and Hermes agents

Caveats

• It’s prompt-generation glue, not a generation tool — you still need the underlying models to render anything
• The README is unusually confident (“the strongest tool available for directing AI video”) — the craft rules are sound, but no benchmarks back the superlative
• Model files track fast-moving targets; Seedance 2.5’s reference is dated to ByteDance guides of July 2026, so freshness depends on the maintainer

Verdict

Anyone generating AI video or images through an agent and tired of adjective-soup prompts should look at this — the dramaturgy layer alone is a decent film-school crash course. If you just want a syntax cheat sheet, the reference files are still useful, but the value is the direction, not the dictionary.