Manul

Poklyk Ktulhu post-production tool

A service that turns a two-to-three-hour game recording into production data: transcript, characters, locations, key moments. AI does the grunt work, a person makes the calls.

A NocoDB table of Poklyk Ktulhu episode moments: timecodes, moment descriptions, comments and shot statuses.
Episode moments in NocoDB: timecode, description and shot status

The show

Poklyk Ktulhu is a Ukrainian actual-play show built on Chaosium's Call of Cthulhu tabletop RPG. An episode usually runs two to three hours, and every one is illustrated with stills and animation. Before anyone can make those, someone has to know what actually happened in the episode: who showed up, where, and which moments are worth a frame.

How it works

  1. Sound. Every player wears their own lavalier mic. The tool stitches the recorder's takes together by timecode and aligns them to the finished episode's timeline.
  2. Speakers. AssemblyAI transcribes each channel separately, and a script decides whose line it is: a word stays on the mic where it's loudest. When two channels are nearly level, the time offset between them and the recognition confidence break the tie.
  3. Episode breakdown. AI agents pull out characters, locations and key moments, around sixty per episode.
  4. Review. At every stage a person checks and edits the output in its own interface before moving on.
  5. On to production. Moments, characters and locations sync to NocoDB — the working table the stills and animation are made from.

Where automation stops

The tool doesn't write prompts for the shots. Each prompt depends on the previous shot, the light in the scene, who is in frame and in what state, so every one is a creative decision. That's why automation stops where its output can be checked: extracting data. I make the shots one at a time and review each one myself. There's a separate case about that.

My role

All of it: agent design, the server pipeline, audio processing, the interface for every stage.

Stack
  • Nuxt
  • TypeScript
  • Claude
  • Gemini
  • AssemblyAI
  • Server-Sent Events
  • NocoDB