Last year, I joked that LLMs are the new runtime for executing instructions written in human language.
I jinxed it.
Last week, I discovered the real power of skills in AI agents. We’d known about them for a while, but used them mostly as rules — descriptions of how something should be. Turns out, they’re full automation. Zero code. Just markdown.
What’s a skill?
An open standard that Anthropic released in December 2025. Microsoft, OpenAI, and GitHub adopted it within 48 hours.
The format: a folder with a markdown file. Inside — step-by-step instructions. No code. The agent reads the description, and when a task matches — it loads the file and executes it like an algorithm.
Essentially, it’s an SOP for AI agents.
The thing that blew my mind
I work on three features in parallel in the same repo. When I get into flow, my commits turn into “changed 1000 files, 15000 lines of code.” Complete chaos in the git history.
So I wrote a skill that:
- Analyzes all changed files
- Groups them by feature (mapping to specs)
- Stages each group separately
- Generates a meaningful commit message for each
- Asks for confirmation
- Commits each group individually
Zero code. Just markdown instructions: what to analyze, how to group, in what order to call git add and git commit.
It works. Every single time. The skill now lives in every repo I work with.
The key difference from prompting
- A prompt is a one-shot command. It lives in a chat session and dies with it.
- A skill is a multi-step process with logic, validation, and file operations. It lives in the repository. It’s versioned. It improves over time.
A prompt says: “Commit my changes.” A skill says: “Analyze what changed. Group by feature. For each group: stage only those files, generate a message following conventional commit format referencing the spec, show me the diff, wait for approval, then commit. Move to the next group.”
That’s not prompting. That’s programming — in a runtime that understands English.
Why this matters
If you have a repeating process that you explain to an agent (or a colleague) every time — that’s not a prompt. That’s a skill. Write it once.
What used to require pages of documentation, a wiki nobody reads, and a new hire onboarding checklist — is now a single file that the runtime actually understands and executes.
I used to joke about it. Now I build with it daily. Be careful what you wish for.
At Finsi, we have over a dozen skills that automate everything from smart commits to content publishing workflows. Each one is a markdown file. Zero code. Full automation. The LLM runtime is real.