# Oct 16, 14:40

## Overview

Chipp and David meet at David's office to review David's AI-made short ad for "the roster" ("everyone has AI, not everyone has taste") and work through his pipeline. They look at the music, the warm, filmic color grade that Claude applies in DaVinci Resolve, and a distorted face on the boss in clip 1. Chipp helps David set up a Claude Code status line, merge terminal windows into tabs, and run each session from a project folder while keeping content in a separate folder. Claude then plans a higher-resolution LTX v2 re-render of clip 1, plus a Topaz version for comparison. The pair also agree to swap skills: Voiceify, a cinematic/filmic prompt skill, vocal removal, Buttercut video analysis and the html questions skill.

## Key points

### The roster video and grade
- The music sits in the background, subtle "like a department store," and starts and stops with the cuts.
- The Omni reference storyboards are graded warm and get warmer shot by shot, matched to the look of an Apple ad. Chipp found them washed out; David says the milky, filmic look is deliberate, like old Cooke lenses.
- David grades in DaVinci Resolve by telling Claude what to match. Resolve Pro (about $200, one-time) is needed for the full API.
- The grade isn't finished; some shots are darker than David wants.

### Fixing the boss's face in clip 1
- The boss's face warps and gets "mushy" as the camera pushes in. The start frame itself is clean.
- LTX v2 rendered at roughly 720p, and the DaVinci timeline is 1080p, so playback is interpolated upscaling, not AI upscaling. They proposed a 720p duplicate of the project, but the timeline points to ProRes upscales, not the native LTX files.
- Claude advised against upscaling the existing take. Resolve Super Scale or Topaz would only patch it. The best route is a re-render at a higher target resolution using LTX's two-stage latent upscaling, estimated at 20–40 minutes.
- David asked for a parallel Topaz (Starlight) version through its API/MCP, to compare all three versions side by side.
- Open questions about LTX: how many seconds it can render at higher resolution, whether clips can be stitched into a 15-second shot, and whether it can edit just one region, such as the faces.
- David is trying Draw Things, a new Apple app for running local models such as LTX.

### Claude Code setup
- Claude built a status line showing the model, context used (as a progress bar) and the session folder, with the folder in an accent color.
- At first the status line didn't show because the session ran in a folder David hadn't trusted. Starting from a trusted project folder fixed it, and the context bar appears once the script can work out a percentage.
- Chipp runs every session with `claude --dangerously-skip-permissions`.
- Terminal windows can merge into tabs ("Merge All Windows").
- Chipp's folder rule: keep code, skills and apps in one repo folder (e.g. Mulligan, "AI projects") so the dev brain can walk them. Keep content (e.g. the roster) on another drive, and tell the session where the content lives.

### Skill sharing and review workflow
- The html questions skill renders several versions (images, videos, audio) with checkboxes and notes. You copy your picks back into Claude, and it can also serve as review gates between script, music, storyboards and video.
- David auditioned each Voiceify line about four times and judged them by feel.
- David showed his other projects (the UGV tutorial, "build versus buy") and an open-source tool that strips music out of audio.
- David's ElevenLabs credits ran out, and he needs the Gmail account to log in.

## Action items

- Re-render clip 1 of the roster at a higher resolution in LTX v2 with the two-stage pipeline.
- Render a Topaz version of clip 1 and compare it side by side with the other versions.
- Decide whether to duplicate the DaVinci Resolve project at 720p and relink it to the native LTX files.
- Finish the color grade on the roster video.
- Answer the open LTX questions: length limits at higher resolution, stitching clips, and editing only the face region.
- Email David a copy of Chipp's status line code from CW-AI.
- Package the Voiceify skill and put it in the shared pCloud folder.
- Add the cinematic/filmic skill to the pCloud folder as an .md, adapted for Midjourney, Nano Banana Pro and Nano Banana 2.1 instead of Adobe Firefly.
- Package the vocal-removal skill (demucs, "vocal gate") and the Buttercut analyze-video skill for David.
- Find the html questions skill and send it to David.
- Log into the Gmail account to get back into ElevenLabs.
