The camera recorded an hour with no audio
I would normally lip sync that by hand. I asked the agent to fix it instead, and it went frame by frame and did it. That is the pattern for most of my editing work now.
I was editing and the camera had no audio. An hour of footage. The other angle recorded fine, so typically what I do is lip sync it and match it up. That takes me about an hour, mostly just making sure it is right.
This time I let the agent try. It went frame by frame, did some calculations with the audio, and fixed it.
I did not want to do that sync by hand. So I did not.
0:20
The old school way still has a use. You go and check for that little bump in the audio. If I go to the camera and I see it, I know that is where I started. That was the whole point of it.
Small time savings, stacked
Same thing with thumbnails. The agent pulls the frames, looks for the ones that are good, and builds the thumbnail. I am not saving hours on that one. I would do it manually. But it is done and I did not touch it.
Stop reserving agents for the big jobs. The work worth handing over first is the repetitive hour you already know how to do by hand.
The interface rewrites your prompt
The ChatGPT UI for images is incredible right now. I have an AI agent that generates images for me, but when I use the API it does not work the same. I asked why. Apparently when you type a prompt into that interface, they rewrite it. That is where the better images come from. It reminds me of what Mid-Journey used to do.
2:27
I can hand it a photo and say recreate this, and it does a fantastic job. Make me smile, change the other person's face. I had to tweak it over a couple of generations to say do not change my face until I got it. The output is usually lower resolution, but I scale it with AI.
A client sent me photos taken on an iPhone that were blurry and terrible. I said I need the same photo, do not change anything, just fix this line this way. It worked. I also had a logo I did not have the files for. It was sitting in a photo on Google Maps. It recreated it and I used it in the posts.
Same photo. Do not change anything. Just fix this line.
Once you find your style, write it down
Go back and forth until you find the look, then stop repeating yourself. I have a skill file for Claude that I share on my YouTube channel as a freebie. First you design your color palette. These are my colors. Then the font. Then the layouts.
I design the little pieces so the agent knows what goes where. Lower third on top, headline, the two cameras here, the screen recording here. One version for YouTube and anything horizontal, one for anything vertical. For the cards I use Remotion, which runs on React.
Write your palette, fonts and layouts into a file the agent reads every time. Otherwise you are paying for the same back and forth on every project.
Slides go stagnant, so steal the layout
Gamma is pretty good if you want something fast. The templates are decent. You can pick a slide and try a new layout, but you are consuming credits. I use the free one and I am fine. If you are doing a lot per month you will need a membership.
The real problem is that slides come out stagnant. Give it a brand deck and it repeats the same structure. It loves creating boxes. So I go to Canva and collect layouts I actually like, delete the slides I do not want, and download the deck as a PDF.
The PDF matters. It embeds the position and the font size, like HTML. If you save it as a JPEG, the agent has to take a screenshot and figure out where you wanted everything. With the PDF you just say recreate this presentation, and it does a fantastic job. The colors you can worry about later.
8:47
Search Canva presentation templates for your vertical and collect the layouts you like.
Delete every slide you do not want so only the layouts remain.
Download that deck as a PDF, not a JPEG, so position and font size come with it.
Hand the PDF to the agent and tell it to recreate the presentation.
One API instead of ten
There are a lot of models and a lot of services reselling them at different prices. The one a bunch of people are on right now is Kie AI. One API and you can reach nano banana, GPT, the video models, Grok. Cheaper than going direct. Higgsfield has no API, which is the problem. People love spending time in that interface, and I like it too, but not when I want something fast.
13:01
When I want images, the agent writes the prompts and gives me a plan. Three images per prompt to find the best one. Then a rundown of the cost, comparing Kie against the direct API, before anything runs. You have to be patient. The first pass is not going to work one hundred percent.
It tells me what it will cost before it spends anything.
Agents talking to agents
Often I will build something in ChatGPT first and use that as the reference image for the generator. I like this style, except X, Y and Z. Once I have the direction I want, I hand that photo over.
Then the agent posting on Instagram talks to the agent that generates the images. It needs a carousel post, it asks for the images, it gets them back, it builds the carousel. All polished.
17:21
Instead of me doing it manually. That is the point.
01
Give the agent the repetitive hour you already know how to do by hand, not just the big jobs.
02
Download reference decks as PDF so position and font size come with them. A JPEG makes the agent guess.
03
Write your colors, fonts and layouts down once so you are not re-prompting your own style every project.
Want the skill file?
The Claude skill file with my palette and layouts is on my YouTube channel as a freebie. If you want me to share the image generation agent or walk through the Canva to PDF workflow on your own deck, tell me what you are building.
Send me a text