Agent Skills I Actually Use

I do not use agent skills as magic commands. I use them as small operating procedures: ask harder questions, keep work simple, review with receipts, normalize design tokens, and start projects with less drift.

Adarsh Trivedi · August 20, 2025 · 6 min read

Takeaway: The useful skill is not the prompt itself. It is the repeatable boundary it creates: when to ask questions, what evidence is required, and what kind of output is acceptable.

How I think about skills

Skills are the lightweight playbooks I use with coding agents. I reach for them when a task has a repeatable shape: review this PR, challenge this plan, normalize these styles, or set up a project branch without losing the thread.

The best ones are boring. They do not try to be a second brain. They make one kind of work easier to repeat, and they make the acceptance criteria harder to skip.

General-purpose skills

Two skills are useful across almost any engineering task:

Grill me — a pressure-test before building. It is useful when I have an approach but want the weak assumptions challenged before code exists.
Ponytail — a simplicity pass. It pushes toward deleting, inlining, or deferring anything that does not need to exist yet.

I like these because they do not make the agent more ambitious. They make it more skeptical.

Custom skills I keep using

I also have a few custom skills that grew out of repeated work patterns. I keep their detailed notes separate so this article stays readable:

Peer review — reviews a teammate's diff against the repository's own rules, not memory.
Token normalization — cleans migrated UI styles into the design/token system after visual parity.
Project start — turns a task into a ticket, branch, and implementation outline so the work starts with less drift.

What makes a skill useful

The skills I keep are the ones that create a clear boundary. A review skill should not silently edit. A planning skill should not jump into implementation. A token cleanup skill should run after visual parity, not before.

That separation is the point. When the skill is small, I can tell whether it did its job.