我用 Antigravity 建置 AI 履歷工具，卻被 LinkedIn 封鎖

Gradient Log 1: Build in 2 days, Get Banned in 5

I spent last week building an AI resume tool with Antigravity, and getting banned from LinkedIn.

Hey everyone,

This is my first newsletter, and I’m excited to share what I’m learning as I build in public.

I’ve been sharing my journey on X, but the platform is shifting toward TikTok-style viral content. So I decided to give Substack a try. It seems like the new place for deeper insights. I also think it’s going to grow in 2026, as I hear more and more people are jumping on it.

I’ll be sending updates on what I’m building, marketing experiments (successful or disastrous), and interesting findings I come across.

My latest updates: I launched sweprofile.com but I realized finding users is brutally hard. I also explored more ways to build fast with AI.

Share

📦 Building & Marketing sweprofile.com

The Idea

While browsing trustmrr, I noticed AI resume tools making serious money. Rezi.ai caught my eye in particular. I’ve spent weeks building polished MVPs before, only to launch and get zero revenue. This time, I wanted to try something different: lower the bar, build fast, get free users, maybe collect some emails. No massive SaaS commitments or revenue goals yet.

I decided to enter the resume optimization space, but keep it simple. Just a straightforward tool that optimizes your resume for a specific job description.

That tool is sweprofile.com.

The Build

A friend recently shared his excitement about Antigravity, Google’s AI coding IDE that’s currently in free beta. I’d been using Cursor, but this felt like the perfect opportunity to try something new.

Antigravity enforces a structured workflow that keeps the AI on track by organizing information better in the context window. I used to manually implement something similar with the Cursor agent, but Antigravity does it automatically:

Processes my build request

Generates a list of tasks

Returns an implementation plan

Executes and writes a summary of what was built

Each artifact requires my review and approval. The UX is smooth, as I can add comments to files just like Google Docs.

For the tech stack, I stuck with what I know:

Frontend: Next.js (Static Export) on GitHub Pages. For $4/month, I can host unlimited websites that run fast.

Backend: Pocketbase on a Hetzner instance (~$15/month). It’s a lightweight alternative to Supabase, managed via Coolify so I can spin up new instances instantly. This single instance can easily handle up to 50 websites.

Total monthly cost: ~$19. Working product ready in 2 days.

The Marketing

This is where the trouble started.

Building was smooth. I had the workflow figured out and it was just a matter of execution. But when it came to marketing, I didn’t have a plan.

I remember Alex West saying marketing is not a science. Indeed, it felt more like alchemy, where my only option was to try different things.

I used Grok to find people complaining about layoffs and the job market. Found some tweets, mostly from small accounts with low engagement. I DM’d them links to my tool.

Result: Zero replies.

I knew Reddit is hostile to self-promotion, as I’ve had posts removed before. So I first built up my karma by leaving helpful comments. To my surprise, one comment got 35+ upvotes. When my karma hit 40, I felt confident enough to post a demo in r/SideProject. It was a ~20-second faceless video of me using the tool, with text explaining how I built it.

Result: “Sorry, this post was removed by Reddit’s filters.”

That didn’t feel good, but at least I knew it was time to move on.

I created a company page and posted my demo, re-using the same video from Reddit. Then I spent hours commenting on posts about the bad job market, offering the tool for free.

Result: My analytics showed ~50 users from social media, likely from LinkedIn. But the next day, my company page was gone, likely flagged as spam.

After the bans, I went back to Twitter and re-posted the demo video as a regular tweet. Eventually, it got >1k impressions, which totally unexpected. My impression is that Twitter rewards video content more, and I should double down on it.

After all this, I managed to get some traffic and collect one email address. Not much, but it’s something.

Lessons

One thing is clear: AI has made building incredibly easy. Anyone with a laptop and access to AI tools can ship a working product in days. The real moat isn’t code anymore, but distribution. And I think we are in a golden age to make money with apps, before everyone finds out and the market gets flooded.

If I want to thrive in this AI-powered world, being a developer is no longer enough. I need to become a great marketer. Actually, I’d prefer the term Growth Hacker, as the it feels less professional and more like someone learning, breaking things, and having fun. Someone willing to try the impossible to get users, even when the odds (and spam filters) are against me.

I’m not sure what to do next. Part of me wants to extend sweprofile into an app that automatically applies to jobs. It’s more complex, but I can easily build it with Antigravity. The thing that worries me is I can’t find competitors, which might mean there’s no market. And finding users is clearly my real problem.

But here’s another idea: if AI can write my code, perhaps it can help with marketing too. It could suggest strategies, edit posts, or find leads. I’ll keep experimenting and report back.

Thanks for reading Gradient Logs! Subscribe for free to receive new posts and support my work.

🚀 Launch Update: ScribePilotAI on TinyLaunch

Last week I also launched a previous project on TinyLaunch. The app is scribepilotai.com, an SEO tool that analyzes your website, suggests article ideas, and writes them for you.

I never put much effort into marketing this. I tweeted about it a few times and got some free signups, but no paid users. I decided to submit it to TinyLaunch since it costs nothing and could lead to some revenue.

I didn’t have high expectations, but to my surprise it ranked #10 out of 137 launches with 14 upvotes. Still no sales, but I would definitely consider this a win!

🤖 Interesting Findings: Long-Running Agents

A common theme that’s emerging is the idea of letting agents run for extended periods. It sounds scary as they could break things, but others are pulling it off successfully.

I saw a developer build a Rust-based image compressor without writing any code by hand. The final product achieved performance close to mozjpeg, a 30-year-old library. The crazy part? His longest-running agent session was ~3 hours.

This made me realize I might be under-leveraging these tools. I could build much more complex MVPs if I just let them run longer.

We created AGI and called it Ralph Wiggum. It’s actually not quite AGI yet, but it enables Claude to work on coding tasks autonomously for extended periods.

The core of enabling long-running scripts lies in configuring the agent harness for persistence and using hooks, particularly the stop hook. Hooks are shell commands that execute at specific points in the Claude workflow, allowing for deterministic actions in an otherwise non-deterministic LLM pattern.

The stop hook is crucial because it prevents Claude from exiting a task. Instead, it re-feeds the task back into the prompt, creating a continuous loop until a completion promise is met or max iterations are reached.

This is where ”Ralph Wiggum” comes into play. It’s essentially a bash loop that keeps feeding the same prompt into Claude until completion or a given number of iterations is reached. It works very well when feeding Claude a list of tasks and asking it to implement them one at a time. It helps to use test-driven development, because asking Claude to come up with tests and running them keeps it on track.

Thanks for reading.

I’ll keep posting updates as I figure this out. Hopefully next time I’ll have more wins to share. If you want to see if I can actually figure out this marketing thing (or just get banned from more platforms), subscribe below to follow along.

If you’re tired of generic AI resume writers, give my free tool a spin: sweprofile.com

Ready for more?