Trimming Toil: Automating Repetitive Development Tasks

I maintain a small number of open-source tools that require regular patching. As I got more comfortable with coding agents, I finally addressed two minor annoyances I’d let linger for too long.

The first was having a coding agent produce a GitHub Workflow for moving forward a major version tag (v1) any time I created a new minor version. Major version tracking is a common approach for GitHub Actions where consumers want to reduce the number of updates to their workflow files.

Previously, this involved a couple extra commands to map the major version to the new commit of the latest release. While it was not “moving a mountain”, I’d put off automating it for far too long. I described what I wanted to Claude Code, and the workflow worked on its first use.

Another situation I found myself in regularly: dependabot would update a library, and the transpiled JavaScript in the dist folder would no longer match the source. The PR workflow catches this mismatch, which meant I had to check out the branch, rebuild, and push the updated dist. I tolerated running this sequence of commands 5-10 times over the past year.

Instead of a workflow, I addressed this with an agent skill. Skills are now an open standard that provide agents with a scalable way to automate common tasks. Instead of running a series of commands, I can simply tell my coding agent “rebuild dist on PR 123”.

Both examples share something: they’re deterministic by design. The workflow runs the same way every time. The skill gives the agent explicit steps rather than asking it to reason about dependabot PRs from scratch. Coding agents are powerful because they can improvise. But for tasks where I already know the right sequence, encoding that sequence is faster, cheaper, and more predictable.

Smaller iterative improvements like these add up. This is where I see coding agents actually providing value, not generating entire applications from a prompt, but chipping away at the friction that accumulates around the work. As these improvements pile up, I spend less time on maintenance and more on the tools themselves. I’m curious how these compound at scale when teams start chaining these patterns together.