Download directly from GITHUB, or reference in your coding agent:

https://github.com/jveritas/aistartupseo-skills-md

I shipped a full SaaS product — auth, payments, email campaigns, AI-powered reports, admin panel with 35+ sections — without writing most of the code myself.

Not because I can't code. But because I've found a faster way.

For the past year, I've been building AIStartupSEO using Google's Project IDX (now Antigravity IDE) paired with Claude Code running in the terminal. The IDE gives me the environment. Claude does the heavy lifting. And a single markdown file keeps the whole thing from going off the rails.

That file is skills.md — and I'm sharing the entire thing below.

What is skills.md?

If you've vibe-coded anything beyond a weekend project, you know the problem: AI coding assistants are incredibly capable for about 20 minutes. Then they start forgetting your conventions, overwriting working code, introducing abstractions you didn't ask for, and slowly turning your codebase into spaghetti.

skills.md is my solution. It sits in the project root and acts as a persistent instruction set that every AI session reads before touching anything. Think of it as a constitution for your AI coding assistant.

It covers:

• Tech stack defaults — so the AI never wastes time asking "should I use yarn or npm?"

• File management rules — three markdown files (todo.md, dev.md, deploy.md) that keep the project self-documenting

• The Golden Rule — "Don't Break What Works." Read existing code before modifying it. Never refactor unless asked.

• Code patterns — API route structure, database workflow, error handling conventions

• UI/UX principles — landing page structure, dark mode, Tailwind conventions, copywriting tone

• Security essentials — input validation, auth checks, rate limiting (non-negotiable)

• Git habits — commit messages, what to gitignore, when to commit

• Agent skills ecosystem — pointers to community skills from skills.sh, obra/superpowers, and Claude Code templates

Why this works

The key insight is that AI coding tools are stateless by default. Every new session starts from zero context. skills.md gives them institutional memory.

Without it, you spend half your time re-explaining how your project works. With it, you drop into a session and say "add Stripe webhook handling" — and the AI already knows your file structure, your error handling pattern, your API response shape, and that you use Prisma with cuid IDs.

It also prevents the most common vibe-coding disaster: the AI "helping" by rewriting 200 lines of working code when you asked it to change one thing. Section 3.2 exists specifically because I lost an afternoon to that once.

My actual workflow

Here's what a typical feature build looks like:

1. Open the project in Antigravity IDE (cloud-based, runs anywhere)

2. Start Claude Code in the terminal

3. Claude reads skills.md, CLAUDE.md, and the project structure

5. I describe what I want in plain language

6. Claude plans the implementation, modifies the files, runs the dev server

7. I test in the browser, give feedback, iterate

8. Claude commits when the feature works

For AIStartupSEO, this workflow produced:

• NextAuth with 5 OAuth providers + credentials

• Stripe integration with 3 plan tiers

• An AI report generation pipeline processing 1M+ queued reports

• Email campaigns via AWS SES with warmup progression

• A full admin panel managing users, content, settings, and analytics

The entire product. One developer. One AI. One markdown file keeping it all sane.

The file

Here it is — copy it, fork it, adapt it to your stack. The specific technologies matter less than the patterns.

https://github.com/jveritas/aistartupseo-skills-md

Tips for adapting it

Change the stack section to match your tools. Using Supabase instead of raw Prisma? Swap it. Using Svelte? Update the file conventions. The structure matters more than the specific technologies.

The three-file system is the most valuable part. todo.md keeps you on track. dev.md prevents context loss between sessions. deploy.md means you can always get back to a working state. Even if you ignore everything else, use these three files.

Section 3.2 will save you. "Don't Break What Works" sounds obvious until your AI assistant decides to refactor your entire auth system while fixing a button color. Being explicit about this in your instructions is the difference between shipping and starting over.

The Lessons section in dev.md compounds. Every mistake you document is a mistake the AI won't make again. After a few months, your dev.md becomes genuinely valuable institutional knowledge — even for human developers joining the project.

What I built with this

AIStartupSEO helps AI startups get discovered. It generates SEO reports, submits to directories, and runs targeted email outreach — all automated.

If you're building something with AI tools and want to see what a solo-built SaaS looks like at scale, check it out. Or just steal my skills.md and build your own thing. That's the whole point.