Close-UpMonologueRun

1. Scaffold the working files first

Before any other action, suggest a workspace folder named `~/Documents/close-up-monologue-run-workflow/` and ask me to confirm or replace that path. After I confirm, create these three files in that folder as your first act:

- `plan.md` — write the complete runbook for this close-up podcast monologue workflow. Include the goal, the operator override fields, the local screenshot reference rule, the normal Nova Ops stage handoff sequence, the validation checks, and the review/reroll criteria. Re-read `plan.md` whenever context is compacted or a new session starts.
- `.env` — create this credentials file even though this specific run needs no new external credential. Put only placeholders and comments, no secret values. Use this exact content unless the existing pipeline environment names a specific provider variable I should preserve:
  `# No new credentials are required for this operator-override run.`
  `# Keep the existing Nova Ops image/video provider keys already configured in the pipeline environment.`
  `EXISTING_PROVIDER_KEYS_ALREADY_CONFIGURED=`
  Prompt me to leave this blank as a confirmation flag or to fill it with `yes`; never copy provider secrets into chat, `plan.md`, or `progress.md`.
- `progress.md` — track the run as a timestamped checklist. After each completed step, append a line in this format: `## 2026-05-11 14:23 — completed: <step>` followed by `## next: <step>`. Re-read `progress.md` before every new action so you know what is done and what remains.

Never store any API key, provider key, token, login, or webhook secret in `plan.md` or `progress.md`. If a secret is ever needed, keep it in `.env` only and ask me to enter it directly.

2. What the operator connects up-front

Use the existing Nova Ops pipeline and the normal image/video generation providers already configured for podcast video runs. Do not ask me to install a new AI workspace, create a new external app, add a new connector, or grant a new OAuth permission for this specific run. No MCP server, CLI installation, SDK package, or direct third-party connector is added here; the better path is the existing local dashboard job API plus the existing agent pipeline handoff stages because the research says the run already uses those surfaces.

Use one local screenshot reference that I provide. Treat the screenshot as wardrobe, hairstyle, jewelry, microphone, framing, and shot-style reference only. Do not treat it as a likeness target and do not copy the reference person's identity. The screenshot may guide cues like a high updo, gold earrings, a leopard-print high-neck top, a microphone, and a dark ribbed podcast backdrop when those cues are present in the operator request.

Carry the override through `job.input.operator_overrides` and the run-local `operator_overrides.json` file written into the `creative_plan` stage. The override must include the exact spoken monologue, a flag that allows a spoken monologue instead of the normal guest-question requirement, and the requested close-up podcast angle. Do not change normal podcast behavior globally; make the exception only for this run.

3. What you watch, read, or trigger on

Trigger only when I queue a local dashboard request for a standard `full_content_run` with `media_mode` set to video, `delivery_mode` set to spoken, `reusable_images` containing the local screenshot path, and `operator_overrides` containing `allow_spoken_monologue`, `exact_script`, and `podcast_angle`.

Read the operator-supplied monologue exactly as the source of truth. Read the local screenshot path as a reusable visual reference. Read the requested close-up podcast angle as the framing source of truth. Human review remains available at the normal image review and video review points; do not skip those review gates if the pipeline presents them.

4. What you do step-by-step at each event

When the queue request appears, run this sequence:

1. Confirm the request is a local dashboard queue request for a video, spoken delivery, a reusable screenshot reference, and operator overrides. If any of those are missing, stop and ask me for the missing field instead of guessing.

2. Queue the run with the exact spoken script and the screenshot reference. Preserve the operator sentence exactly. Do not add speaker labels, guest questions, extra intro lines, or cleanup rewrites.

3. In Writer, create `delivery_mode.txt`, `title.txt`, and `script.txt` from the exact operator override. Confirm `creative_plan/script.txt` exactly matches the operator sentence with no speaker labels. If `script.txt` contains a rewritten sentence, an added question, a speaker name, or extra framing text, stop and reroll or correct the Writer output before continuing.

4. In the creative-plan handoff, write or verify the run-local `operator_overrides.json` so downstream stages can see that a one-line spoken monologue is allowed for this run. The purpose is to bypass the normal guest question requirement only for this operator override, not to weaken validation for future podcast runs.