How Skills Are Changing the Way I Work with Agents

Working with coding agents used to take more setup than I expected. Before a task could begin, I would explain the project, describe the output, list the checks that mattered, then catch the constraints I had assumed were obvious.

The agent could still help, but I was rebuilding the same context every time. Skills have changed that part of the work. They give an agent a better starting point, so I spend less time recreating a process and more time deciding what the work should become.

For daily engineering work, I use Matt Pocock’s skills. They cover a broader set of workflows than the one in this article, but they made the value of reusable process clear to me. This note is about a different skill, one I used to make a video.

One video, four revisions

I used Video Shotcraft, a video-production skill, to make a short introduction for a follow-cam project. The video needed to move from the method to the original wide footage, then to the processed vertical output.

The first version began with a precise brief: which source time range to use and what the video had to include. For the second version, I rewrote the copy and asked for full sentences to stay on one line. Abrupt line breaks are something I often have to correct in Codex’s video text.

The third version changed the visuals. I asked for the 01 Method and 02 Before sections to use the original footage, with the person we were following on the left and the copy on the right. That was a sensible layout choice, but it created an awkward accident. The reticle used for the narrative happened to frame the target neatly. It looked like a real verification result, even though it was not connected to the diagnostic data.

This made me reconsider the scene. Explanatory graphics can make a process easier to follow, but they should not look like live diagnostic output when they are not. A stronger next version would connect the explanation to accepted reference images and verified frames.

The fourth version dealt with continuity. I asked for one continuous music track across the whole video. We then used an audio-timeline sampling correlation check to compare points in the finished video with the source audio and confirm that the cuts stayed at the correct timestamps. It gave us a way to verify the timeline instead of only judging whether the edit felt smooth.

What the skill changed

The video had to do more than show a tighter crop. It had to explain, at a high level, how the system established a reference, verified reliable positions, and used those positions to guide the camera path. The original and processed footage then became a before-and-after comparison.

Video Shotcraft gave the agent a process for that work. It covered source material, layout, timing, sound, and final checks, rather than treating the task as a request to make an attractive final shot.

That is what I find useful about skills more generally. A skill can say what to inspect first, when the agent needs input from me, what it must not make up, and how we check the result. It is more useful than a clever prompt in a notes app because it can carry the details I would otherwise forget when I am in a hurry.

For repeat work, I no longer have to remember every check at the moment I need it. I can keep my attention on the story, the compromises, and the feedback that changes the direction. I can also review the work against an agreed process instead of only asking whether it looks plausible.

What remains mine

Skills can be a poor fit. A workflow may be too broad for a small task, too rigid for new work, or simply wrong for a project. I still need to decide whether it applies and supply the material the agent cannot safely infer.

In the follow-cam project, the important changes came from feedback about the narrative, the layout, and what the video should claim. A checklist could not make those calls for me. I use skills for work with a known standard, a repeatable review path, or a set of constraints I do not want to rediscover every time. They help an agent handle the repeatable parts carefully and leave the meaningful decisions visible.