Claude Code edits video from frame-level feedback
Aziz reports a Claude Code video workflow with 25 corrections on one edit, turning reusable feedback into Markdown rules that later editing sessions read.
marvinaziz.de
Original prompt
Shared by the source author. See source notes for context.
hell no i hate the underline sound
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Marvin Aziz
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Aziz reports a Claude Code video workflow with 25 corrections on one edit, turning reusable feedback into Markdown rules that later editing sessions read.
- Human involvement
- Aziz reviews motion, timing and taste decisions and resolves conflicting rules. One video still required 25 comments before the published version.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Keep scripts, footage, transcripts and prior style context in a session folder.
- Have Claude orchestrate transcription, cleanup, cutting and motion graphics.
- Review the edit and attach comments to the affected frames.
Marvin Aziz describes making videos with Claude Code inside a repository containing his notes, scripts, prior posts and style guidance. Whisper produces transcripts, audio tools clean recordings, FFmpeg performs cuts and HyperFrames renders motion graphics written as HTML. His Takes app keeps scripts, takes, voice tracks, edits and social posts together in each video session.
Claude prepares an edit, then Aziz reviews it by placing comments directly on the affected frame. He describes one published video that required 25 comments. The agent fixes the marked problem and replies; comments relevant beyond that single video become Markdown rules with their originating feedback. Across his videos, he reports 163 comments becoming 65 rules that new sessions read before editing.
The rule file includes constraints about false starts, repeated shots, dimmed screens and sound effects. It reduces repeated corrections but does not remove editorial judgment. Aziz says Claude samples frames rather than understanding continuous video, so motion and timing still need his eyes. Some rules also contradict one another until he cleans them up. The concrete result is an iterative editing process with remembered taste decisions, not a measured guarantee of better audience performance. Its copyable detail is to save durable feedback beside the work so the next agent session starts with the lessons from the previous edit.
Prompt
Original prompt
Source: Blog · Marvin Aziz
Frequently asked questions
What turns a correction into reusable context?
Feedback that matters beyond one video becomes a Markdown rule, linked to its original comment and read before the next edit.
Does the agent understand continuous motion?
Aziz says Claude samples individual frames, so he reviews motion and timing himself; Gemini helps describe old footage for B-roll selection.
Are repeated corrections eliminated?
No. Rules reduce repeats, but taste decisions and contradictory instructions still need human review.
More details
Remaining steps
- Ask Claude to fix and rerender those issues.
- Promote reusable feedback into a Markdown file read before later edits.
About this account
Claude Code selects and orchestrates media tools, prepares edits, responds to frame feedback and applies stored rules in later sessions.
Firsthand account by Marvin Aziz; original publication date verified against Primary article byline October 7, 2026. Outcomes are the author’s reports, not independently reproduced. Original revision instruction transcribed verbatim from the author’s feedback examples; this is a correction, not the full initial brief.
Credit: Marvin Aziz
Source published
Added to Agents In Use
Source checked
Tools and topics
Claude Code · Takes · Whisper · FFmpeg · HyperFrames · Markdown rules
Frame feedback · Video editing · Creative memory
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.