Superpowers in Action: How AI Coding Turned Two Weeks of Work Into Half a Day

What Are Superpowers

Superpowers isn't a specific tool. It's a new way of working. When you combine AI coding tools (Claude Code, Cursor, Copilot) with a structured development workflow, the result isn't "a bit faster." It's a qualitative leap.

Traditional development: ideation → mockups → design doc → review → development → testing → deployment → documentation. Each step depends on different people and time windows. The whole pipeline stretches to weeks.

AI mode: describe the goal to AI → AI asks clarifying questions → produces design doc → AI self-reviews → generates code → AI writes tests → deploy. One person, half a day.

This isn't theory. What follows is my actual experience today.

From Zero to Live: Building a Blog End-to-End

The starting point: a vague idea

My original requirement was one sentence: "I need a personal tech blog to showcase my skills and attract consulting clients."

In a normal workflow, that sentence triggers: find a designer (2-3 days), frontend development (3-5 days), backend or CMS setup (2-3 days), deployment and domain (1 day), testing (1-2 days). Optimistically, two weeks.

What actually happened: I typed /office-hours into Claude Code.

Phase 1: Design diagnosis (10 minutes)

The AI — playing the role of a YC partner — didn't start writing code. It asked three critical questions first:

1. Demand reality: "Does anyone actually want this? Have clients asked for your portfolio?" I admitted I'm building ahead of demand. That honesty matters — it means I should control investment, not chase perfection.

2. Status quo: "How do you currently demonstrate your skills?" I don't. Zero online brand presence.

3. Target user: "What kind of client do you want?" I started with "anyone." The AI pushed back twice until I committed to "AI + automation tooling" as my niche.

Three premises were surfaced and challenged:

• Is a blog the right channel for client acquisition? Depends on content and SEO.
• Does it need to be fully-featured? Search engines rank complete sites better.
• Is bilingual necessary? The Chinese AI content space is underserved — differentiation opportunity.

Phase 2: Design review (15 minutes, 3 rounds)

The AI produced a design document, then launched an independent review agent. First round: 5/10 — missing data model, didn't account for Pagefind/SSR compatibility, Giscus Discussion workflow was factually wrong.

Fixed and re-reviewed: 6/10 — discovered a fundamental issue: Pagefind indexes static HTML, but Next.js App Router defaults to SSR. These are incompatible.

Third round: 7/10. Switched to static export (output: 'export'), corrected the Giscus configuration. Passed.

This "write → independent review → fix → re-review" cycle would normally require two senior engineers in a room for half a day. The AI completed it in 15 minutes, and the second round was stricter than the first — it remembered previously flagged issues and verified each one was actually addressed.

Phase 3: Implementation (2 hours)

After the plan was locked, five phases executed in sequence:

Phase 1 — MDX rendering. Replaced all hardcoded placeholder data with real data sources. Homepage fetches from getAllPosts(), article pages render MDX via getPost(slug, lang), code blocks get dual-theme syntax highlighting.

Phase 2 — SEO + distribution. next.config.ts set to output: 'export' (the entire blog is static). generateMetadata outputs OG tags and hreflang alternates per page. RSS feed and sitemap.xml generated at build time.

Phase 3 — Comments and search. Pagefind indexes all static HTML post-build via pagefind --site out. Search page embeds a client-side UI with 300ms debounce.

Phase 4 — Visual polish. Inter font, indigo gradient accents, hover animations, dark mode, 404 page, mobile hamburger menu.

Phase 5 — Testing. 15 component tests (vitest + testing-library). bun run build passed on the first attempt.

Phase 4: Deployment (30 minutes)

Registered the alex908.com domain, configured Vercel auto-deploy, DNS set up with automatic SSL. git push triggered the build. 18 static pages generated, Pagefind indexing complete.

The core of this workflow isn't how much code the AI wrote — it's that the AI handled everything you know you should do but often don't have time for: design review, edge case analysis, testing, SEO configuration, RSS generation.

Three Principles of Superpowers

1. Boil the Ocean — Do it completely, once

Traditional intuition says iterate: build the core, ship, add based on feedback. But AI makes completeness nearly free.

Write all tests. Cover all edge cases. Configure full SEO and distribution. It looks like "doing too much," but because of AI's speed, the complete version costs less than the incremental version with all its round-trips.

2. Adversarial Review — Let AI review AI

A single AI model's output always has blind spots. But when you use one model to generate content and another (or an independent instance of the same model) to review it, you create a mini peer-review system.

The most valuable finding in this build — the Pagefind/SSR incompatibility — wasn't discovered by me writing the design doc. It was caught by the review agent in round two. That bug, if discovered post-deployment, would have at least doubled the fix cost.

3. Decision Audit Trail — Traceable choices

Every auto-decision the AI made was logged: which principle applied, which alternative was rejected, why. This matters. Three months from now, when you look at the code, you know why that choice was made — you're not guessing from the aftermath.

This isn't replacement. It's amplification.

After using AI tools, I'm still writing code, making decisions, and controlling quality. What changed is that AI absorbs the cognitive overhead — I don't need to remember every API signature, don't write boilerplate, don't manually check every SEO best practice.

Superpowers isn't AI thinking and coding for you. It's AI covering all the repetitive, mechanical, easy-to-miss parts while you're thinking and coding. You only need to do the one thing AI can't: decide what's worth building.

---

Full source code for this blog is on GitHub.