Know your audience
It starts with knowing the audience.
- What viewers care about
- What they already know
- How much to explain
- Examples you choose
- What they should take away
You can’t edit a video without knowing what you’re making. Codex also needs to know what files it has, what was said, and what to inspect in the picture before it starts making cuts.
Answer these questions in order so Codex does not have to guess.
You can’t edit a video without knowing what you’re making. Codex needs this even more than a human editor.
It starts with knowing the audience.
Once we know who we're making it for, we need to give the story some structure. Codex can definitely help here but make sure it's locked in before asking it to edit.
There is no perfect video length. A video should be as long as the story needs it to be. That said, having a general range will help Codex make better decisions.
Is this video a long form 16:9 video or are we creating a 90-second reel for social media? Being explicit with these decisions helps Codex when choosing where to place elements on screen and composing clips in the edit.
This one is a little harder to specify to Codex but even something as simple as “upbeat launch animation” or “documentary style explanation” will help Codex understand the type of video it is making.
Picking up where we left off yesterday, now that we know what we’re making we need to create an inventory of the assets we’ll use for the video.
What footage was captured?
This usually starts with camera footage. Have Codex analyze all of your footage so that it knows what it is working with.
Check every clip, flag picture or sound problems, and group different versions of the same line or moment.
Are the demos complete?
If this is a demo that will include screencasts, have Codex take a peek at these too. Is everything readable? What are the focal points of the demo? (e.g. cursor movement, typing, etc.) Is sensitive information shown that might need masked? Codex can help with this and more.
Watch the recording from start to finish. Mark key clicks, hard-to-read text, missing steps, and anything private that should be hidden.
Which sound should we use?
While Codex can't “listen” to audio, it does have ways to analyze it. For dialogue, have Codex transcribe it both in a traditional and word-by-word format. Have it analyze for director cues like “that was perfect” which indicates a good take.
Transcribe dialogue in normal and word-level formats, then search for cues like “that was perfect.”
What visuals are needed?
Help Codex with any other graphics that are needed: titles and L3 styles, including any pre-installed templates; brand guidelines; motion graphics templates and guidelines; and any other logos, screenshots, etc. that you want to use. Specify these upfront to avoid surprises!
List every file, template, and brand rule. Locate or request anything that is still missing.
Up next we have one of the core fundamentals of Codex video editing: transcription. Codex can’t “watch” a video the way we do. Transcription gives it the words; the next phase covers what it can inspect in the picture.
Models are fantastic with words though and many videos have a lot of them.
Finding any line quickly and seeing exactly what was said in each take
Audio transcription has a lot of options. The most important one for editing with Codex is creating both a normal and a word-level transcript. Word-level transcripts give Codex extra precision in knowing when words are spoken and where pauses occur which helps with making cuts.
Making cleaner cuts and jumping back to the original clip
Another transcription must is speaker diarization. This tells Codex who is speaking. Codex can distinguish between a presenter and director or between multiple people in an interview. This can help with auto-switching cams in interviews or ignoring non-presenter lines in cuts.
Switching cameras when the speaker changes and leaving directions out of the cut
The transcript is usually what decides the best clips to use. Codex can determine: Did they stutter? Did the producer react positively? Which take matched the script best? Which take had the best pacing? This is very powerful for a model that can't “see” or “hear.”
Making a short list of the clips worth using first
Codex can compare the words, timing, and reactions in each take, then explain why one should be used.
We’re editing video so let’s talk about what Codex can see 👀
If you have a demo, let Codex take a look at it. Some things this can help with:
Are you revealing unreleased info? (we would never...) Do the demos show what the script says they should? Is it easy to read what's on screen?
I've seen Codex catch things I 100% missed.
Consistency is very very important when making videos. Viewers are sensitive to a lot of things they might not even be aware of (e.g. keeping camera moves consistent).
Codex can be an extra set of “eyes” to look at an edit and determine whether anything is inconsistent.
Speaking of consistency, one of the biggest sources of inconsistency is simple errors that might be present in one shot but not others.
If Codex finds a mistake in one take you can have it look to see if it can find a take without the mistake automatically!
Codex is great at analyzing camera footage. It'll use ffmpeg and ffprobe to analyze it. You can get as fine grained with this as you want. I often start with scene detection and go from there.
After Codex knows where the scenes are you can get even more detailed.
For instance, if I want to replicate an animation I like I will provide a video of it to Codex and have it analyze it frame by frame and figure out the animation curve. It's pretty powerful!