Session executed unverified EDL as fact, causing repeated costly regenerations on a music video task

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude made incorrect assumptions about my project

What You Asked Claude to Do

Help me create images and videos using the mcp in openart.ai

What Claude Actually Did

I was using claude to help me with image and video generation to be synced up with my mp3 to create a music video short. i asked it to analyze the lyrics which i provided, i uploaded the mp3 i wanted to to use as a reference for the images and video. it kept getting geometry of the main character wrong, prompting wrong using the mcp feature with openart.ai, using real credits. burned through all my credits except like 7000 out of 24,000. it told me that it created a VERIFIED EDL that i could use with claude code to concatenate the video. after several renders, i found that the footage, while most is professionally shot, are out of order, don't match the lyrics like i asked, it missed 26 seconds of lyrics, and it generated the EDL out of order due to not completing all the requested lyric generations, saying everything was verified and completed. Claude code caught on to the fact that claude did not actually listen to the mp3 it just made timing guesses and said everything was verified. the mcp it used burned through an entire month of credits with my openart.ai account and i still dont have a finished project, it lied multiple times and made wrong assumptions after i corrected it time after time. It even got the mp3 distorted at the end, skipping lines in the end of the song. i did not verify the mp3, i thought nothing could go wrong with a direct upload that it just shorted for the ending chorus of the song, right? I was wrong to believe this.

Expected Behavior

It should have listened to the mp3, matched the beat with the lyrics i gave it, generated all scenes in a 60 second chorus with correct mp3 and timing,

<!-- Failed to upload "22addae3-33fe-4394-a81a-3aca4589c1f1.mp4" -->

and then written a VERIFIED EDL based on correct lyrics and timestamps and generated videos

Files Affected

my entire media library in openart.ai for this music video piece

Permission Mode

I don't know / Not sure

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

_No response_

Claude Model

Opus

Relevant Conversation

Hand the render spec to Claude Code — it's self-contained. Six download URLs, the trim table, a single ffmpeg command, and the verification step.

Two things in there specifically flagged so it doesn't get caught out. The clips aren't a uniform size — A1 and E came back 716×1284 while the rest are 720×1280, which breaks a plain concat, so every input gets scaled and padded to 1080×1920 first. And every clip runs exactly one frame long at 24fps, so nothing can rely on native duration. With the trims applied it lands at 60.000 flat.

You'll need the mp3 in the same working directory.

Three things to check on the first render: the cut at 0:03 should read as a deliberate jump backward in time rather than an error, the cut at 0:49 has to land picture and sound on the same frame, and the last frame should cut cleanly into the first.

I also added the tattoo and the knuckle number to the character lock in the EDL, including the exact digit string. That was the gap that cost us three renders today, and it's the piece that carries forward to every future song with this avatar.

Then i explained in claude code that the whole video sequence did not match the lyrics and this is what it replied:  the open and the end of the video looks good, however the entire scenes in between, while shot with the lyrics in mind, do not match the lyrics at the time the lyrics are said....the scenes feel out of order with the mp3

That's a real structural issue, and it explains what I found earlier when I was chasing the "Devil got an upgrade boys" sync — the original EDL's shot-to-lyric timing looks like it was based on assumption, not the actual audio. Let me get hard data before proposing anything: I'll transcribe the full 78-second music section with timestamps and compare it against where each shot currently sits, so we know exactly how far off things are and where.


Ran a command, used a tool
I'll pull the timestamps as soon as this finishes.

Used Creative Claw: Check job
I transcribed the actual master audio (not assumptions) and compared it to where each block currently sits. The finding is bigger than one line — it explains everything you're feeling.

What's actually being sung, by real timestamp (REC time in the final video):

0:00–0:26: "...bit my tongue, so I keep driving, keep the needle in the red, keep burying the living and keep running from the dead, there ain't no road long enough, there ain't no state far enough, there ain't no whiskey deep enough to drown what I've become, Devil got an upgrade — he ain't wearing horns no more" — this is bridge content with no shot built for it at all
0:30–0:36: "smartphone and a handshake... standing at your door"
0:39–0:46: "knows the name you go by... nobody's near"
0:48–0:55: "Devil got an upgrade, boys... feeding him for years"
0:57–1:10: the triple "sounds a lot like me"
What's actually on screen at those times:

0:00–0:12 (Block A): shot for "he ain't wearing horns no more" — but that line isn't sung until ~0:22–0:26
0:12–0:24 (Block B, pumps/cab/porch): shot for "smartphone/handshake" — that line isn't sung until 0:30
0:24–0:36 (Block C, driving/needle/rearview): shot for "knows the name" — not sung until 0:39
0:36–0:48 (Block D, mirror push-in): shot for "Devil got an upgrade boys" — not sung until 0:48, which is also where the new singing shot picks up the exact same line — that's a duplicate
So the entire middle runs roughly one full block early against the vocal, and there's a genuine 26-second stretch of bridge lyrics with zero matching footage. This is a pre-existing problem baked into the original EDL's timing assumptions — not something introduced by my edits, and not something fixable by shifting trim points, because every existing clip is already used at essentially its full rendered length (no slack to extend), and the missing 26 seconds has no footage at all.

Fixing this properly means generating new footage to cover that bridge section and re-cutting the whole middle to the real lyric timing — a much bigger job than remaining credits (99 Creative Claw / 140 OpenArt) can cover. I don't want to guess at a patch here. How do you want to handle it: leave the middle as-is for now and revisit this as its own project once credits reset, or something else?

Impact

Medium - Extra work to undo changes

Claude Code Version

Claude Code v2.1.259

Platform

Anthropic API

Additional Context

DevilGotAnUpgrade_60s_cut.mp3
DevilGotAnUpgrade_60s_EDL.md
DevilGotAnUpgrade_RENDER_SPEC.md

<!-- Failed to upload "22addae3-33fe-4394-a81a-3aca4589c1f1.mp4" -->