← AI Portfolio · Case study

Remastering an event with Codex.

I worked with Codex on graphics and timeline automation for two installments of a recorded event. The workflow had to keep up with changes I was making by hand.

Part 1: keeping the timing traceable

The first installment began with information cards. I had the copy in a spreadsheet and an existing animation to work from. The job grew to include still images, removals, and audio repairs.

There was also a Spanish recording made against the earlier English version. Every change to the edit could affect where that translation belonged.

I had mostly deleted things and left gaps, but there were a few moments where I had to shift things, and some ended up extending a bit. Keeping the gaps was no longer enough to preserve the original timing.

Three timelines to keep track of

The same material, at different times

1. Original recording

The Spanish recording followed this version.

2. Revised edit with gaps

An extension moves later material. A removal leaves empty time.

3. Final edit

Close the selected removal gap. Keep the extension and the pause.

  • A B C Retained passages
  • + Editorial extension
  • × Selected removal
  • Intentional pause

Conceptual example, not the event's timeline or scale. In the saved production pass, 4 selected closures removed 2,464 frames: 98.56 seconds at 25 fps.

We needed to map where each retained passage came from and where it ended up. The version labeled "with gaps" already included my extensions, so its name couldn't tell us which clock we were looking at.

Some empty stretches were there because I wanted a pause. I wanted a little breather at the end, for example. Those needed to stay when we closed the gaps left by removed material.

A map between the versions

Codex turned the editing notes into a cue sheet, with timing recorded in integer frames at the sequence's verified 25 fps. I reviewed the planned actions before each editing pass, and the saved project was checked afterward.

The initial pass placed 48 visual inserts: 19 information cards and 29 still images. Removals first left gaps. Once I had finished my refinements, 4 selected gaps were closed from the end backward, with the other sequences protected.

For the Spanish audio, the map selected 7 retained source segments and placed them in the final edit, with 6 intentional silent intervals. Each segment kept its playback speed. The original recording stayed on a muted track, and the conformed audio went on a separate track for review.

We also made a bilingual timing-review document to help locate passages. Its speech-recognition text could guide listening, but still needed checking against the audio.

What needed another pass

Moving the clips hadn't settled the transitions. Their positions were stored separately and needed a correction pass. We also had to refresh the English transcript after the final timing changes.

The first graphics delivery contained rendered cards and structured copy data, but the promised editable After Effects builder was missing. That became one of the jobs for the second installment.

Part 2: rebuilding the editable system

The next installment needed 35 information cards and 39 supplied still images. I was still reframing photos and adjusting transitions while the automation work was underway, so each pass needed to start from the current saved edit.

We rebuilt the graphics system from the approved renders and previews. I approved 3 prototypes before the batch: a long card, a biography, and a short card. Their sampled transparency bounds matched the references, with text-line endpoints no more than 2 pixels apart.

This time, the delivery included the editable After Effects project, the builder, and structured text data, alongside 35 transparent renders and 35 previews. New types of layout would still need their own prototype review.

I set the treatment for the photos by hand. Codex continued that pattern for the remaining stills: replace the picture, keep the program audio, and use the demonstrated 30-frame transitions.

A video contribution inside a nested sequence also needed repair. Two broad attempts failed their checks and stopped without saving. A narrower repair fixed the picture join and restored the known transitions, then checked the saved result.

The two cards I had moved

The final audit reset all 35 cards to the default framing. I had deliberately moved 2 of them because faces were low in the frame. Making those cards match the others undid an editing decision.

After I pointed it out, their exact approved offsets were restored and checked in the saved project. The check needed to account for those exceptions, and the rule it had been using didn't.

35 cards, with 2 deliberate exceptions
Check the approved framing for each card

33 Default framing

Keep the shared scale and position.

2 Approved exceptions

Keep the offsets I chose to leave faces visible.

Schematic of the corrected framing. The 2 positions were restored after the audit had reset them. A reusable exception registry is a proposed next step.

What each installment produced

Part 1

20 rendered information cards, including a later short card. The initial automated pass placed 19 cards and 29 stills.

The Spanish audio file matched its planned sample count, with zero-valued samples in all 6 intended silence windows. The saved listening checklist and 20 translated-card reviews were still open.

Part 2

An editable graphics system with 35 renders and previews. All 39 stills were integrated through my manual work and the automated continuation. Final verification matched 10 protected sequences to their reference snapshots.

I exported the locked English master and reported that it was broadcast. The export's metadata and stream analysis matched the technical delivery format and reported no errors.

This account draws on the two exhibit packages and their saved reports and artifact checks. It does not add an independent full-playback review of either installment or independent observation of the broadcast.

What still needs work

Part 2's Spanish conform is still open. Part 1 produced a technically checked review file, but its saved record also stops before listening approval. The second run doesn't establish that the language workflow has been tested end to end again.

The next improvement is to record approved exceptions before running a check: which card changed, the exact setting, why I changed it, and the saved version where I approved it. Once I say an edit is ready to lock, any further change needs my go.

The reusable workflow documents carry those rules forward. Graphics, captions, and language review keep their own approval steps, and a changed final edit needs a new timing map.

← Back to AI Portfolio Book a call