@FinanceYF5: 3/Step 1: Create with your best examples + personal context. Key: Keep skill.md and examples separate. AI doesn't need to read a bunch of examples each time; load on demand; also don't leak privacy when sharing. Provide at least several different examples; giving only one will cause overfitting.

3/🧭 Step 1: Create from Your Best Examples + Personal Context

Key: Keep skill.md and examples separate.

The AI doesn’t need to read a pile of examples every time; load on demand. Also avoids privacy leaks when sharing. Provide at least a few different examples—using just one leads to overfitting.

1/ Peter Yang publicly revealed his secret to saving hours each week: build his own AI Skill.

Codify personal knowledge and taste into a reusable skill, letting the AI do the work for you.

He built one live using Claude Code in just 5 steps.

2/ First, understand what a Skill is.

It’s simply a folder of instructions that the AI automatically triggers for a specific task.

His example: an “edit-post” skill for editing long-form newsletter content.

3/ Step 1: Create from your best examples + personal context.

Key: Keep skill.md and examples separate.

The AI doesn’t need to read all examples every time; load on demand. Also avoids leaking privacy when sharing. Provide at least a few different examples—just one leads to overfitting.

4/ Step 2: Write “when to trigger” clearly in the description.

Normally the AI only reads the skill name and description to decide whether to load the full skill.

So you must explicitly write “use when…” otherwise the skill won’t reliably trigger.

5/ Step 3: Create evals.md for self-checking.

This is a trick other tutorials don’t cover.

Ask the AI to generate 10 “pass/fail” checks (more reliable than rating 1-5): Does it have a hook? Does it sound like AI? Does it have substance?

6/ And make it auto-iterate.

During eval, open a separate agent with a clean context. If it fails, go back to the original agent and continue fixing.

During his demo, it took 5 rounds to pass all checks—the human could go grab coffee while it iterates.

7/ Step 4: Create memory.md to improve over time.

Record past conversation experiences in reverse chronological order, keep it to 2-3 sentences per day.

Ideal for vague feedback that can’t be judged by “pass/fail”, like “make the tone more authentic”.

8/ Step 5: Build a “meta-skill” for building skills.

When constructing a skill, the AI writes a lot of content. If you don’t watch it, it’ll pack in AI slop.

Use a Skill Editor to periodically review: remove dashes, remove clichés, keep it concise.

9/ But the most counterintuitive point is last.

Even with all best practices, skill output is only 80-90% there.

What truly separates AI slop from good content is the last 10-20%: your own manual review, finishing with human taste.

Skills help you run the first 80%, but that final stretch must always be you.

Watch the full content:

That’s all.

If you like this topic:

1. Follow me (@FinanceYF5)
2. Like + repost the first post below