Redo one scene of a Shorts episode without redoing the episode

Continue the same Format thread with previous_run_id and a new instruction to redo a single scene. Sume treats it as one more turn of the same conversation.

4 min readSume
All posts

Create a new Format run with previous_run_id set to the episode's run and an instruction that names the scene. Sume treats it as one more turn of the same conversation instead of a fresh start, so the agent still has the earlier turns.

The docs say a single-scene retry on the same thread costs a fraction of the original run's spend, which is the whole point.

The request

previous_run_id is one of the fields a run item can name, alongside instruction, input and attachments. An unknown or foreign id returns 404 previous_run_not_found. The Continue a run section of the runs page covers the thread rules.

Write the instruction as a delta: the scene number, what was wrong and what to keep. A vague 'make it better' re-opens the whole episode.

Redo a scene with previous_run_id (read 2026-10-07)
FieldValueReason
previous_run_idThe episode run idSame thread, same context
instructionScene 4 only, with the fixLimits the work
generation_spend_cap_usdA small numberCaps a runaway retry
Idempotency-KeyNew per retryA spent key replays the old run

Fit with batches

In a bulk queue an item may carry previous_run_id, so a season's reshoots can be queued the same way as its first pass: one item per scene that failed review. See the bulk post for limits.

Deadlines apply to each run separately, so a scene run does not inherit the 90-minute clock of the original.

  • Review the episode, list failing scenes, queue only those.
  • Keep the original receipt for the artifacts of scenes that passed.
  • Do not reuse an idempotency key for a different instruction.

What a good redo instruction looks like

Name the scene by its number and the problem by what a viewer sees. 'Scene 4: the lamp is on in the first shot and off in the second; keep the dialogue and the camera, fix the lamp' is a good redo. 'Fix scene 4' invites a rewrite of everything the agent remembers.

List what must not change. The agent holds the earlier turns of the thread, so the more you say about the parts that passed, the less it touches them. If an artifact from the first run passed review, say so by name.

Cap the spend on the redo. A small generation_spend_cap_usd turns a runaway redo into a clean failure that you can inspect, and it keeps the cost of a single-scene fix visible in the receipt as usage.billable_amount_usd_micros. The spend-cap post lists the values the API accepts.

Checks after the redo

Compare the redone scene with the one before and after it. A fix that is right in isolation can still break the cut, so watch the three scenes in sequence at full speed.

If the redo fails twice, stop. Start a fresh run for that scene with a rewritten instruction, or move the scene to the video model's reference mode as in the continuity post.

Sources

Related posts

More in Formats

All Formats posts

Written by Sume