Your Skills Are the Moat: 141 Claude Code Sessions, 611 Commits, and What Actually Matters

Hey folks! A few months back I wrote about Vibe Coding with Steroids — the claim was that custom skills, parallel agents, and superpowers turn Claude Code into a productivity machine.

That was theory. Today I’ve got the receipts.

I ran Claude Code’s /insights command and the numbers from the last 30 days hit different:

• 141 sessions analyzed
• 611 commits
• 96% goal achievement rate
• 952 TaskUpdate calls, 506 Agent invocations

That’s shipping-focused orchestration at scale. And the one thing doing the heavy lifting in those numbers isn’t the model. It’s the skills.

The Agentic Coding Connection

I recently helped review a book called Agentic Coding: Claude Code for Developers — and one of the core ideas stuck with me: agents aren’t magic, they’re disciplined systems. The magic happens when you give them the right context, the right process, and the right boundaries.

That’s literally what custom skills do. They’re not just prompts — they’re distilled expertise that survives between sessions. You teach Claude once, it remembers forever.

Which brings me to the thesis of this post:

Your skills are the moat. The model is a commodity.

Everyone has access to the same Claude. What separates a 10x workflow from a frustrating one is the skills you’ve built around it.

3 Skills That Paid Off

Let me show you what that looks like in practice.

1. write-like-iago — Voice consistency

I write blog posts, LinkedIn posts, Reddit drafts, and UGC video scripts. Without a skill, every session starts from zero — “write in my voice” means nothing without examples.

---
name: write-like-iago
description: Use when creating content for Iago Cavalcante
---

## Tone
- Conversational and warm
- Didactic without being condescending
- Start with "Fala, pessoal!" or "Hey folks!"

## Ground concepts in reality
BAD: "Hot reload improves productivity"
GOOD: "You know when you'd create a skill and had to restart the session?"

Result: 20+ articles shipped this quarter, voice stayed consistent across PT-BR and EN.

2. elixir-skills — The router that saves me from myself

This one is sneaky good. It’s a skill that routes to other skills before Claude explores the code:

Elixir/Phoenix/OTP task → Invoke skill FIRST → Then explore → Then code

Why? Because skills tell you what to look for. Without them, Claude explores blind and often adds a GenServer when a plain function would do. The router catches that early.

3. lovable-web-development — Design tokens or bust

Red flag: text-white  → Do: text-foreground
Red flag: bg-black    → Do: bg-background
Red flag: hex inline  → Do: CSS variables with HSL

Built the entire Respira fintech design — landing page plus 10 mobile screens — with zero hardcoded colors. Future me can reskin the whole thing in 20 minutes.

The Friction (Keeping It Honest)

The report also called out where I still hit walls:

• Sed-based renames silently corrupted a class (SystemPromptBuilder → SystemAgentBuilder). Lesson: always pair bulk renames with tsc --noEmit.
• Token limits cut off 3+ debugging sessions. Fix: dump findings to .claude/scratch/ instead of inline.
• Pre-push hooks amended formatting into the wrong commit. Force-push recovery isn’t fun.

None of this is the AI’s fault. It’s missing skills. Every friction point is a skill waiting to be written.

What I’m Codifying Next

Based on the report, here’s what goes into my ~/.claude/skills/ this week:

1. /pr-review — Fetch Gemini review comments, classify valid/invalid, apply fixes, resolve threads, commit, push. I do this 5+ times a week manually.
2. /ota-release — Bump version, commit, eas update --platform ios then --platform android separately (never --platform all because web isn’t configured).
3. Pre-commit hook — Auto-run prettier and eslint --fix on edited TS files. Kills the force-push recovery cycle.

Small skills, big leverage.

How to Start

You don’t need 141 sessions worth of data to start building skills. You need one painful, repeated workflow.

1. Notice you’re doing the same thing twice → candidate for a skill
2. Write a 20-line markdown file in ~/.claude/skills/your-skill/SKILL.md
3. Add the frontmatter with name and description
4. Use it three times. Refine. Ship.

The description field is the most important part — Claude uses it to decide when to invoke the skill. Be specific about the trigger.

The Real Insight

After a month of data, here’s what I actually believe now:

Agentic development isn’t about the model getting smarter. It’s about you getting better at teaching it.

The 96% goal achievement rate isn’t because Claude is magic. It’s because every week I’m feeding it a sharper version of my patterns, preferences, and constraints. The skill folder is the compound interest.

Vibe coding with steroids gets you started. Skills are what make it stick.

---

Got a painful workflow that repeats? That’s your next skill.

Find me on Twitter or LinkedIn — curious to see what you’re codifying.

Happy coding!