Always Ask... (A skills architect for YOU!)

I am including my /skill-architect here. Take it, customize it, go wild. (jump to the end if you don't want to know how I got here...) I've seen quite a bit of questions on here lately around skills, so I hope this helps.

TL;DR - If you aren't asking Claude questions within chat AND planning with the chat feature via word vomit, yapping, brain dumps, and simply asking good questions... you should really start. (I plan to discuss my entire workflow for pretty much anything in another post)

---

## Context & Process

I'm working on building an organizational structure for my personal brand. I am following the ICM framework, and heavily focusing on the 60% architecture before I even consider orchestration or AI integration.

Today, I'm working through my content creation pipeline, where I'm focusing down on individual tasks that I want each 'employee' within my 'organization' to do. Enter skills.

Now, I've worked for about 2 days on-and-off yapping on walks to develop a ROBUST product requirement description/document (PRD) and standard operating procedure document (SOP). After uploading this, I asked a simple question: If anything is unclear, if you couldn't exactly replicate this, [or] if I'm missing edge cases, I want you to ask me questions back and gather as much detail as possible. Interview me, one question at a time." This went on for about an hour or so before Claude decided that it had what it needed. It spit out a comprehensive document.

I opened a new chat, uploaded the file, and gave it one simple prompt:

[Can you break this down into a list of skills that would be relevant to have?

For reference, I define a skill (or a skills file) as an individual, granular capability that an AI agent needs to execute its job. Rather than a general instruction, a skill is a comprehensive document that uniquely describes exactly how to perform a specific task.]

What this did was take everything from that document and meticulously peeled out what the jobs-to-be-done are to achieve the overarching objective. I ended up with 51 skills and one-line descriptions of what each skill did. I saved this as a markdown.

I then gave Claude my 'MVP' workflow of the first objective I want my 'Marketing & Content Creation' department to do (spoiler, I took a flow that I admired and thought would fit my lifestyle and use case from a few content creators... mashed up. Don't recreate the wheel, people). I asked it "what are the MVP skills that must be created first to complete this flow?". It took the list, and identified 21 out of the 51 skills that need to be designed first -- in order of priority.

---

### The Secret Sauce

Now, how do I create these skills, you might ask?

I asked Claude. "If I'm working to create or develop a skill for you to use, what are all the questions you would want to ask me in order to create a robust skill that is both highly effective and also specific to me and my situation/use-case? Be comprehensive."

The output was 33 questions categorized by the following:

Purpose & Trigger Logic
Input Landscape
Output Specification
Process & Decision Logic
Quality Criteria & Evaluation
Your Context & Personalization
Maintenance & Evolution

This was great. I wanted to go one step further. So I asked another question, followed by providing it 2 'best-in-class' examples of skills-builder tips I've collected: "I actually want you to create this as a skill that I can load up and use. Or is this in your skill-builder already?

I also have this as a reference: [insert referecnce]"

I've attached the output note. TL;DR - Claude's skills-builder is basic; 4 question intake. Will require major iteration. Not great for immediate and repeatable value.

So... I now have a comprehensive end-to-end skills architect.

---

### My Gift to You

Sharing is caring. Here's what you're getting:

skill-architect — a unified, end-to-end skill creation engine that handles the full lifecycle in a single conversation: structured intake interview → specification brief → SKILL.md authoring → folder architecture → testing → iteration → packaged .skill file ready to upload.

The existing built-in skill-creator asks 4 surface-level questions and jumps straight to drafting. This replaces that entire front end with a depth-calibrated interview (8 questions for a simple formatting skill, up to 35 for a complex multi-tool routing system), then carries the output all the way through to a tested, iterated, packaged deliverable — no phase handoffs, no "now go use a different skill to finish."

The architecture follows the three-layer model:

Layer 1 (always loaded): Pushy trigger description — fires on "create a skill," "build a skill," "turn this into a skill," "automate this workflow," "systematize this process," and all the natural variants. Also catches skill improvement requests.
Layer 2 (loaded on trigger): The full six-phase workflow: interview → brief → author → test → iterate → package. 301 lines — well inside the 500-line ceiling. Includes writing principles (explain the why, one skill = one job, examples beat essays), the spec brief template, folder architecture decision logic, and the iteration loop.
Layer 3 (on-demand): references/question-bank.md — 47 questions across six categories, each tagged by depth level (essential/complex/personal), plus 5 "power questions" for when you're struggling to articulate requirements, and an adaptive interview guide that scales question count to skill complexity.

Key design decisions: the skill mines your conversation history and memories before asking a single question, so it never re-asks what it already knows. It produces a spec brief as an intermediate checkpoint (you approve before any SKILL.md gets written), and it handles skill improvement — not just creation — by reading the existing skill, running a targeted delta interview, and rewriting in place without renaming.

9 comments