Skills, sans Shamanism

I'm a slow adapter of LLMs and distrust the fuzzy logic inherent therein. The use of new tools, in my craft, has to be earned. As an example: I've long been skeptical of skills, on principle.

Skills are fuzzy logic by design. Someone writes prompts, descriptions, examples — and the LLM interprets them to produce behavior. There's no execution, no type-checking, no contract. Just text, parsed at inference time.

I'd been keeping a folder of text files. Prompts I'd refined over months — step-by-step debugging approaches, document formats I liked, how I wanted code reviewed. Copy-paste into the chat window, run. Copy-paste, run. Efficient enough. Just manual.

You might see where this is going...

How LLM skills work

The LLM adds a list of "skill" entities in every prompt. Let us say you're asking your LLM to debug a misbehaving API call. You might prompt something like:

When I call /ping on the dev server, it raises a 419... debug, why would this happen?

The LLM will receive, instead:

[System Prompt (your env/harness)]

Available skills:
- local-debugging: Step through local errors, read logs, run diagnostic commands.
- api-testing: Call endpoints, inspect responses, check auth headers.
- code-review: Review diffs for logic bugs, style, and test coverage.

User: When I call /ping on the dev server, it raises a 419... debug, why would this happen?

Then, if the skill matches — "debug" in the query, "local-debugging" in the list — the LLM pulls the full text of that skill file into the session, and your prompting continues.

This is exactly what I've been doing.

The culture

Why haven't I realized this before? Because the culture around it is aggressively misleading.

Pinecone titled their introductory tutorial: "Superpower LLMs with Conversational Agents." Microsoft announced they were "Unlocking Autonomous Agent Capabilities" and filed a case study under the headline "Pets at Home Unleashes an AI Revolution." Projections have the agent market hitting $199 billion by 2034. A 920% surge in framework adoption was announced, apparently with a straight face.

A skill is a markdown file.

OpenAI's own developer blog — the technical one, not the press release — has a tip: "Write skill descriptions like routing logic, not marketing copy." They know. They just have two voices.

Skills are just files, loaded on demand, whenever their keywords are present. Some tools let you invoke them via /skill-name explicitly - but the same applies.

Lazy loaded text, nothing mysterious. Nothing special. Nothing bespoke. Once you strip away the advertising, the basic concept is brilliantly simple. And it works rather well.

My take

I'm using skills these days, and I'm using them a lot. It's a lot easier to wrap a prompt into a reusable package and reuse it via /skill-name than to repeat myself all the time when prompting things.

I do, however, stick to a few me-isms:

• I don't use other people's skills out of the box, at most I'll rewrite them for my workflows. Their context isn't my context.
• I often create skills "post facto". I find myself doing repeated work and then set up a skill to nail down the pattern. The prompt already works; I'm just making it permanent.
• Shorter is better, I've found. Rather write 3-4 skills than 1 super skill. Skills that try to cover more than one workflow break routing.

The same principles repeat as when writing code. YAGNI, KISS, DRY. If a plain prompt works, a skill isn't needed. If a skill needs five hundred words to explain itself, it's a spec doc — it'll drift. If I'm pasting the same prompt into three sessions, that's my cue.

What a bad skill looks like: Two paragraphs of team philosophy. A numbered list of twelve steps. A "Note:" section with exceptions. A reminder at the bottom to "be concise." The LLM reads all of it and produces approximately what it would have produced with a two-sentence prompt. You've paid in tokens and confusion and gained very little if anything.

I'll write more about my processes with skills, and I'll share a fun story later this week on how I've had to act as a therapist to my AI.