I've used Fish, Cartesia, Panda, Starfish, Mockingbird, ElevenLabs, and ElevenLabs V3 for client work. ElevenLabs V3 was my default for a while because it handled full sentences and emotion well. On several older projects, though, clients approved the Fish version in the first round more often. If the voice mangled a client's name, I generated another version. A bad pause or sentence ending meant another generation. Once several projects were running, I started losing track of which recording and script had produced the approved version. The workflow below keeps the source files and approval history in one folder. It can use Fish or another AI audio API without changing the client review step. Three steps

Codex or Claude Code handles step two. One prompt creates and runs five subagents inside the project. I don't write the agent config by hand. If something breaks, I paste the error into the same chat. Before the first run, I add Fish's official skill and set the API key:

1. Add the three input files

2. Paste one prompt into Codex or Claude Code

3. Send the review pack to the client The main agent returns two folders:

In the HeyGen tests I've run so far, Fish S2.1 Pro gave me the best result. My next run uses the same authorized recording and the same test-script.md in this workflow. For each run, I keep the reference recording, input script, first generation, and the version the client picked. On the next revision, I use that folder as the project's long-term memory instead of rebuilding the context from old chats. Hope this was useful. Vox ❤️