AI video editing

I alway dreamt of having a lifting instagram page but always struggled with editing the videos. managed to do so but it took a lot of time and energy from me which makes it hard to be consistent. i imagined what if i had an intern to whom i could just share the videos and they’d do the editing. idea seems nice BUT economic of it doesn’t :/ so that thing just lost track from my life and i moved on. few months down the line ai agents became really powerful. i was already using them for coding but then i realised that video editing is coding too isn’t it? think of preimeire pro.. it is a well-known software for video editing, keyword being ‘software’ – meaning adobe had wrote code to make UI and code for the actual CORE engine of premiere pro. I had also realised that for many media related tasks much of the world is relying on a single opensource program called ‘ffmpeg’.

so, in theory all i needed was bunch of videos and ask agent to use ffmpeg, python & ffprobe and edit me a gym video. one thing i forgot was TASTE. agents have got good at things that have verifiable success. What is a success of edited reel?
more impressions?
more likes?
quality of edit?

all of it are the outcomes that humans decide once the reel is created. essentially i was getting into RL territory and i absolutely did not want to. but i did want to try and maybe just letting agent do its thing could be a good first options always. so i gave gpt 5.6 sol max my video clips and an instagram reel for reference and asked it to replicate the edit with my videos.

okay - not bad. but then what followed blew my mind. look at these.

with these im really bullish on using ai for video editing. the real bottle neck is identifying how to create an edit using a reference reel. so this comes into a category where i simply ask it to copy the edit but when doing so i have created a skill with all the effects noted down so that no more time is spent on exploration on how to do a particular effect which is already done. i started this project in Jul-25-2026 and it’s been just over a month. this is far from complete and i think it has just started.

another problem is that is what we call TOKENMAXXXING. I burned a lot of tokens, after analysis of some 9 reels, it turned out I had used 487M input (99.6%), 459M cached input (95%), 1.6M output of GPT-5.6 Sol. If I was paying API prices instead of subscription it’d have been a $327 experiment. what needs to be done now 1. to make similar reels with 10x cheaper models like gpt-5.6 luna & 2. to improve the taste of the reels (this is my bottleneck too) also a point to highlight is that i would really like it to be a SKILL rather than build my own agent harness. the reasoning for behind this is that i’d like to use the subsidised api calls in codex, claude or any other sub that offers best intelligence for money given that it is a token intensive task. to further enhance artistic taste i have some things in mind like this oil painting art

or this water color art

both made using code js and python respectively. im not sure how i can use them in reels but probably some overlay. additionally sound effects and music are the most important part of a video so i am going to get a epidemic sound subcription for sound effects and also try new reel formats like motivational reel edits for delivering a message. another kind of edit im really excited about is this kind:

2x2 Split-Screen, Analog CRT & Video Glitch Emulation

I also need to work on how my clips are selected by the AI using contact sheets.

Guide my agents on how to segment things:

i’ve been also trying to play around with Kdenlive python apis.

WHen i was working on this i did see claude use swift code to segment me.

i also think that all of it can’t be automated but assistance of ai for video editing is really useful.

the way i think this would work out is it will eventually become its own harness once the cost isn’t a constrain and human can take over whenever they want.