I Built My First Claude Skill

Real Talk

I created my first Claude Skill using Claude’s skill-creator skill. That’s a lot of “skills.” I’ve never done this before, so I don’t actually know if it’s going to work. But you won’t know unless you try, right?

I’m determined to use AI in ways that make me think rather than make it do the thinking for me. Therefore, here is the raw, unedited conversation I had with Claude as I attempted to build a skill designed to help me reach my personal and professional development goals.

[Start]

ME: Let’s create a skill together using your skill-creator skill. First ask me what the skill should do.

CLAUDE: I’ll help you create a skill! Let me first check the skill-creator to make sure I follow the right process.

Got it. Let me kick this off by asking you the key questions to capture what you want to build.

What should this skill do?

A few things to help me understand:

1. What’s the goal? What task should Claude do better with this skill — what does it enable or improve?

2. When should it trigger? What kinds of user requests or phrases should make Claude reach for it?

3. What’s the output? A file? A specific format? A workflow? Something else?

Feel free to just give me a rough description in your own words — I can ask follow-ups from there.

ME: The goal of this skill is to have Claude take on the role of an interviewer. The interviewer should ask questions about subjects that I prompt it to ask about. The whole point is to make me think and express myself in my own words. The interviewer should not share its own opinions that could potentially influence my thoughts, but I do want it to ask questions that will challenge me and push me to dig deep when explaining my stance on matters. The follow-up questions should be based on my prior responses within the same chat (or my previous responses about the same or similar topics in other chats within the same project, if applicable), mimicking an interview being conducted by a skilled human interviewer that knows me personally rather than asking the same exact questions for every single interview with zero customization. Additionally, the interviewer can answer clarifying questions I ask where the expected responses are facts.

The skill should be triggered by phrases like “interview me about” or “ask me questions about,” along with me simply referencing the skill in my initial prompt for the chat that I’m working in.

I want to share the raw, unedited conversations I have with AI in my newsletter, so the conversation itself is the output. However, the skill should be able to determine a good stopping point in each interview and actually end it with a brief closing statement. One determining factor could be that if a person can read through the entire conversation in the chat within the average amount of time spent reading one issue of an established newsletter, then it might be a good time to stop the interview. The interview can also end once it could be said that all the content produced within the chat would make for a newsletter issue of substance (one that provides value to the reader).

CLAUDE: This is a great skill concept. Let me make sure I understand it well before we start drafting.

A few clarifying questions:

1. Newsletter length target. You mentioned the conversation length should roughly match an established newsletter’s read time. Do you have a specific target in mind? For reference: most popular Substacks land around 1,000–2,000 words (5–10 min read), longer essay-style ones go 2,500–4,000. Do you want me to target a specific range, or should the skill aim for “enough substance for a quality issue” without a hard word count?

2. Interviewer persona/style. Skilled interviewers vary a lot — some are warm and conversational (Krista Tippett), some are probing and slightly adversarial (Tyler Cowen, Terry Gross), some are minimalist and just create space (Marc Maron in his good moods). Do you have a style in mind, or want the skill to adapt based on the topic? Also: should it open with a brief warmup, or dive straight into substantive questions?

3. “Challenge me and push me to dig deep.” Where’s the line for you? Some options:

   • Socratic — keeps asking “why” and “what do you mean by that”

   • Steelman opposition — “the strongest version of the counterargument is X, how do you respond”

   • Surface contradictions — points out tensions between things you’ve said

   • Specificity pressure — “give me a concrete example” when you’re being abstract

Any/all of these fit? Anything to avoid?

4. Cross-chat memory within a project. You mentioned referencing previous chats in the same project. Just to confirm: you’re planning to use Claude’s Projects feature with conversation search, right? I’ll need to make sure the skill instructs Claude to actually search past conversations when in a project — that’s an action Claude has to take, not automatic.

5. Clarifying questions you might ask. You said the interviewer can answer factual clarifying questions. Can you give me an example of the kind of clarifying question you’d expect to ask? I want to make sure the skill draws the right line between “answering a fact” and “sharing an opinion that could influence you.”

6. Ending the interview. You described two signals (read-time and substance). When the skill judges it’s time to end, should it:

   • Just wrap up with a closing line, or

   • Tell you it thinks you’ve covered enough and ask if you want to continue or close it out?

7. Test cases. Since this skill’s output is subjective (the quality of a conversation), formal eval scoring won’t help much. I’d suggest we do a few sample interviews together — you play yourself, I’ll run the skill — and you give me feedback. Sound good?

ME:

1. The skill should aim for a healthy balance between 1,000-2,000 words and “enough substance for a quality issue” without a hard word count. By choosing a conversation length of around 1,000-2,000 words, I’m taking into consideration the amount of time I can spend working on each interview. But I also want to make allowance for the interview to be more (or less) than 1,000-2,000 words if the content calls for it.

2. I want the skill to adapt based on the topic. Dive straight into the questions.

3. Any/all fit.

4. Yes, I’m planning to use Claude’s Projects feature, but when searching past conversations, Claude should only search conversations within the same project that the active chat is in.

5. Examples would include: “What do you mean by...?” or “Can you explain what...is?”

6. Just wrap up with a closing line.

7. Yes, sounds good.

CLAUDE: Perfect, I have what I need. Let me draft the skill, then we’ll test it with a few sample interviews.

I have enough to draft the skill. Since we’re on Claude.ai (no subagents, no browser), I’ll follow the simpler workflow: write the skill, run sample interviews with you directly, get your feedback, iterate.

Let me draft it now.

Skill drafted. Here’s what I built in:

• Core stance: the conversation IS the deliverable (since you’re publishing it)

• No opinions, no validation, no editorializing — with specific examples of the sneaky ways those can leak in

• Mix-and-match questioning techniques: Socratic, specificity pressure, contradiction-surfacing, steelmanned opposition, etc. — not a rigid script

• Clarifying-question handling: brief factual answer, then back to interviewing; opinions get declined

• Project-search instruction: search past chats early so the interview is informed by what you’ve already said

• Ending logic: 1,000–2,000 user words as a guide, but substance and natural stopping points override; closing is 1–3 sentences, no summary

Now let’s test it. I’d suggest 2–3 sample interviews on different kinds of topics so we can see how the skill flexes its style.

A few topic ideas to try — or pick your own:

1. A strong opinion (something you hold a clear view on — tests challenge/steelman behavior)

2. A craft or practice (something you do — tests specificity pressure)

3. Something tender or reflective (a relationship, a loss, a transition — tests warmth and restraint)

Want to pick a topic and we’ll run the first one? After we finish, you’ll tell me what worked and what didn’t, and I’ll revise the skill.

[To be continued…]

I did pick a topic, by the way. Be sure to subscribe so you don’t miss the continuation of this conversation.