I used to install skills like browser extensions. See one. Install it. Hope it helps. Repeat until my skill directory had 40 folders and I only knew what 6 of them actually did.
Then I noticed something. The best skills aren't actually about their code. They're about a pattern — a specific way of thinking about a problem. That pattern is usually 10-20% of the skill's content. The rest is tool names, runtime-specific syntax, and scaffolding.
Once I could see the pattern, I didn't need the other 80%. I could adapt it to my own system in a few lines.
The three-step method
READ. Open the SKILL.md. Don't copy it. Don't install it. Read it and ask one question: what is the core mechanism here? Not the tool names. Not the file paths. The pattern.
EXTRACT. Name the pattern in one sentence. Not "it spawns subagents" — that's a tool. "It prevents the agent from giving up too early." That's a pattern.
ADAPT. Strip everything runtime-specific. Keep only the pattern. Fit it into your own rules, config, or existing skills. If it takes more than 8 lines, you're probably keeping too much.
Three examples from my own system
1. PUA skill → 6 lines in my agent rules
PUA is a Claude Code skill that uses pressure escalation to force agents into exhaustive problem-solving. 7-item checklist, 4-level escalation, 5-step methodology.
The core pattern? Don't let the agent give up without a concrete checklist. The original uses rhetoric like "How do I rate your performance?" — I don't need that.
After extraction — 6 lines: same operation fails 2x → no third attempt until: ① check logs for the specific failing line, ② read source code logic, ③ switch approaches. All three done + still unsolved → report the concrete blocker, not "probably an env issue."
2. Hermes skill format → standardized all 4 of my skills
Hermes has the most rigorous SKILL.md format: HARDLINE rules, platform gating, conditional activation. I didn't want to install Hermes skills into Claude Code — wrong runtime.
The pattern was: version tracking, related_skills, Skip For, Verification. I added all four to my own skills. They went from loose markdown to a real system without importing anything.
3. T-CBB pipeline → delivery-gate documentation upgrade
SwarmAI's T-CBB has 1520 lines, 45 runtime patterns, 8 invariants. I needed one concept: mechanical gates and reasoning gates are different.
My delivery-gate checks file timestamps + regex. Mechanical. My self-audit asks four questions. Reasoning. Before T-CBB I called both "quality checks." After, I understood they're complementary layers. The hook code didn't change. The docs went from "thinking quality" to "mechanical gate: timestamps + regex + disk usage."
Why this matters
The AI agent ecosystem is fragmenting. Claude Code, Hermes, Codex, Cursor — each with different tool names and idioms. Skills built for one runtime break on another.
The patterns don't care about runtimes. "Give the agent a checklist before it can give up" works everywhere. "Separate mechanical from reasoning gates" works everywhere.
Installing fewer skills and extracting more patterns isn't minimalism. It's literacy. You understand what you're using. Your system stays lean. Your skills survive the next runtime switch.
Start here
Next time you find an interesting skill:
- Read it without hovering over the install button
- Find the one sentence that captures the core mechanism
- Ask: can I express that in my own config in under 8 lines?
You probably can. And those 8 lines will outlast any installed skill.
Top comments (0)