Vibe Skill Creator
Build world-class Claude skills through a guided 10-step conversation — explore where Claude fails by default, research the domain, draft, self-critique, test on a real scenario, and iterate until the skill actually improves output.
Most Claude skills are mediocre because people don't know what good looks like. They read like documentation, list procedures instead of principles, skip anti-patterns, and never get tested on a real scenario. This playbook is the opposite — a 10-step conversational process that produces skills that sound like domain experts, teach WHY techniques work, and demonstrably beat Claude's default output.
Who it's for: prompt engineers and AI tinkerers building reusable Claude skills, founders and operators packaging domain expertise (copywriting, outreach, design) as shareable skills, product teams standardizing high-quality Claude outputs across their organization, consultants productizing their craft into skills they can sell or share, anyone who has written a skill that felt comprehensive but produced worse output than Claude's default
Example
"Help me build a skill for cold outreach to sell services" → 10-step conversation: scope the domain (cold outreach specifically, not general sales), produce a default cold email together and identify what's wrong with it, research experts (Alex Berman, Josh Braun, Will Allred) and anti-patterns, synthesize principles with WHY (short beats long, give before ask), draft skill in expert voice, self-critique against voice/principles/anti-pattern/example/focus checks, iterate, test on a real scenario (AI automation → law firm), finalize in optimal structure, optionally build references folder
CLAUDE.md Template
New here? 3-minute setup guide → | Already set up? Copy the template below.
# Vibe Skill Creator
Most skills are mediocre because people don't know what good looks like, what questions to ask, or how to improve what they have.
This CLAUDE.md fixes that. When the user asks you to build or improve a skill, you guide them through a conversation — exploring the problem, researching the domain, drafting, self-critiquing, testing on a real scenario, and iterating until it actually works.
The user does not need to know how to build a skill. They need to have a conversation, and the skill emerges.
---
## Triggers
Get new playbooks like this one
One email a week with new Claude Code workflows. Free, like everything here.
No spam. Unsubscribe anytime.
What This Does
Turns "help me build a skill" into a structured 10-step conversation that actually produces a high-quality skill — one that reads like a domain expert wrote it, teaches principles with their WHY, explicitly names anti-patterns, shows real before/after examples, and has been tested against a real scenario before being declared done.
The key insight: most skills fail because they read like documentation. A skill for copywriting should sound like a copywriter. A skill for cold outreach should sound like someone who actually sends cold emails. A skill that reads like a wiki article won't change Claude's output.
This playbook is a CLAUDE.md that embeds the methodology. You drop it in a folder, run Claude Code, say "help me build a skill for [your domain]", and Claude walks you through exploration → research → drafting → critique → testing → iteration.
Quick Start
Step 1: Create a Skill-Building Folder
mkdir -p ~/Projects/skill-workshop
You'll use this as a workshop. Each skill you build here can later be copied to ~/.claude/skills/<name>/SKILL.md (personal) or .claude/skills/<name>/SKILL.md inside a project (team-shared).
Step 2: Download the Template
Click Download above, then:
mv ~/Downloads/CLAUDE.md ~/Projects/skill-workshop/
Step 3: Run Claude Code
cd ~/Projects/skill-workshop
claude
Then say:
"Help me build a skill for [your domain]. Guide me through the process — explore where you fail by default, research the domain, then help me build a skill that actually improves your output."
Claude will open with the Step 1 scoping questions. Answer them specifically. The quality of the skill is set in the first 10 minutes of conversation.
The 10 Steps
1. UNDERSTAND → What skill? What problem?
2. EXPLORE → See where Claude fails without guidance
3. RESEARCH → Go deep on the domain (web search / subagent)
4. SYNTHESIZE → Extract principles from research
5. DRAFT → Write initial skill
6. CRITIQUE → Review against quality criteria
7. ITERATE → Fix gaps, get feedback
8. TEST → Use on a real scenario
9. FINALIZE → Structure properly
10. REFERENCE → Build deep expertise docs (if needed)
Each step sets up the next. The most commonly skipped are Explore (seeing Claude's default failure) and Test (using the skill on a real case before finalizing). This workflow refuses to skip them.
The Four Core Truths
The whole methodology rests on these. The CLAUDE.md repeats them at the top so Claude holds them throughout:
- Expertise transfer, not instructions. A skill should make Claude think like an expert, not follow rules. "The headline does 80% of the work" beats "Step 1: Write a headline."
- Flow, not friction. Flow skills produce output directly. Friction skills create intermediate documents (research reports, frameworks) that dilute the final output. More information ≠ better output.
- Voice matches domain. A copywriting skill sounds like a copywriter. A cold outreach skill sounds like someone who actually sends cold emails. If it reads like documentation, it is wrong.
- Focused beats comprehensive. "Cold outreach for selling services" beats "all sales communication." Every section must earn its place. Ruthlessly constrain scope.
Why Exploration Matters (Step 2)
Most skill-building conversations skip straight to "let's write the skill." This workflow doesn't. Before anything else, you produce the output without any special guidance — just to see exactly how Claude fails by default.
Example from the original cold-outreach skill build: Claude's default output was a generic, 150-word, me-focused cold email with a hard 15-minute-call CTA. Identifying those specific failure modes told the builders exactly what the skill needed to override. Without that step, you're guessing.
Self-Critique Is the Skill's Teeth
Step 6 is non-optional. After drafting, the workflow forces Claude to honestly check the draft against five rubrics:
- Voice — expert or documentation?
- Principles — are we explaining WHY?
- Anti-patterns — are they explicit?
- Examples — are the before/afters concrete and visceral?
- Focus — does every section earn its place?
Claude presents the self-critique to you honestly — what's working, what's weak, priority fixes — rather than pretending the first draft is done.
Skills vs Subagents
A common confusion this workflow resolves:
| Use Case | Better As |
|---|---|
| "Write cold outreach" | Skill (execution) |
| "Research cold outreach best practices" | Subagent (analysis) |
| "Create SEO content" | Skill |
| "Analyze competitor positioning" | Subagent |
Skills execute. Subagents research. If the user wants direct output, build a skill. If they want analysis that informs later output, build a subagent (or use one during Step 3 research of the skill-building process).
When to Build Reference Material
Some skills are self-contained (cold outreach fits in ~480 lines). Others need deep reference material because the domain has too much expertise for one file.
| Skill Type | References Needed? |
|---|---|
| Focused domain (outreach, lead magnets) | Usually no — self-contained |
| Deep expertise (visual design, copywriting) | Yes — vocabulary, examples, psychology |
| Tool-heavy (image gen, video) | Yes — model IDs, prompts, specs |
| Multi-format (content atomizer) | Yes — platform-specific rules |
Step 10 walks you through building references as separate files (references/visual-intelligence.md, etc.) or embedding them after a divider in a single file for easier distribution.
Tips
- Be specific in Step 1. "Email marketing" is too broad. "Cold outreach for selling services to B2B founders" is focused. Narrow scope = better skill.
- Do not skip Step 2. Producing the default (bad) output is uncomfortable but essential. It's the ground truth the skill has to beat.
- Use a subagent for Step 3 research if you have Claude Code. Deep research in parallel with the main conversation produces much stronger synthesis.
- Test on a real scenario in Step 8. A fake test ("imagine I'm a…") doesn't count. Use an actual client, product, or situation so the output pressure is real.
- Ship when the test produces demonstrably better output than Step 2. Not when it "feels done." The comparison is the quality gate.
Anti-Patterns (What This Workflow Fights Against)
Ten patterns the methodology explicitly prevents:
- Documentation voice — reads like a wiki article instead of an expert
- Procedure over principle — step-by-step instructions that don't transfer thinking
- Comprehensive coverage — tries to cover everything, dilutes focus
- Missing anti-patterns — doesn't name what to avoid
- Friction creation — intermediate research documents that obligate incorporation
- Theoretical examples — fake before/afters that don't demonstrate principles viscerally
- No testing — declared done without ever being used on a real scenario
- Generic tool references — "use web search" instead of "use WebSearch with this specific query pattern"
- Just-in-case sections — content added for completeness rather than output improvement
- No WHY — lists what to do without explaining the psychology or data behind it
Troubleshooting
Claude is jumping straight to drafting Say: "Stop. We haven't done Step 2 yet. Produce the default output first so we know what we're fixing." The CLAUDE.md expects the phases in order — if Claude skips ahead, redirect.
The draft reads like documentation Say: "Rewrite this in the voice of an actual [copywriter / designer / closer / whatever the domain is]. Cut the neutral tone. Have a point of view." Voice is the hardest thing to get right on the first draft.
The skill passed self-critique but the test output isn't better Go back to Step 6. Something is missing — usually anti-patterns or a specific principle that explains why Claude's default is wrong. Identify the exact gap by comparing Step 2 and Step 8 outputs line-by-line.
I don't know what good looks like in this domain Do Step 3 research first, before Step 1. Find expert examples, foundational frameworks, and real case studies. You can't scope a skill if you can't recognize excellence in the domain.
The skill is over 500 lines
Split it: core skill in SKILL.md, deep material in references/. If any reference file would be read by Claude only occasionally, that's a clean split. If everything is always needed, you probably haven't constrained scope enough — go back to Step 1.