PRODUCER
PRODUCER
Diving In
These settings control the main conversation where you and the AI develop your script together. The System Prompt tells the AI how to behave as your producer throughout the conversation. The Model is the OpenAI model used for drafting.
In single-stage mode (default), the System Prompt above handles both producer work and widget formatting. In two-stage mode, a producer-only prompt does the thinking/research, then a lightweight formatter model injects PICK / CHOOSE / ASK markers in a second pass. Useful if the combined prompt causes the producer to skimp on research or over-use widgets.
After the draft script is written (by OpenAI), it goes through a two-pass rewrite using Claude to make it sound authentically like you. Pass 1 (Rewrite) does a full voice transformation using your Voice Guide and Rewrite Instructions, plus the conversation history and research docs for context. Pass 2 (Voice Polish) then reviews the rewrite line by line, catching any remaining AI-sounding language and checking for your signature moves.
Punch It Up runs three passes: a joy pass that casts a wide net for moments of delight, a tighten pass that hunts for flab to cut, and a judge pass that culls the joy candidates so only the strongest reach you. You can run the joy pass through Anthropic, OpenAI, or both; the tighten and judge passes follow the primary provider.
Fact Check extracts every checkable claim from the script, researches each one in parallel with web search, and presents them as cards you can confirm, flag, edit, or dig deeper on. The Extraction Instructions tell Claude what counts as a claim worth checking. Research runs on OpenAI with web_search_preview.
The Final Edit phase auto-generates YouTube title candidates when you arrive, and offers an optional learning-log reflection on the way out. Reflections are appended to learning-log.md and folded into the system prompt for future sessions, so the producer compounds insight over time.
Every time you save a Final Edit, the before and after versions are stored as a training pair. This corpus can be used to fine-tune future voice models. You can browse, inspect, and download entries here.