I built an AI browser extension in a month with an AI coding assistant. The code was the easy part.

I watch a lot of long content: hour-long lectures, 40-page PDFs, long articles in languages I read slowly. For months my workflow was: copy, switch tab, paste into a chat, ask, switch back. Dozens of times a day. Every switch broke my focus.

So I built SideDog, an AI assistant that lives in the browser's side panel and helps you understand the page you're on. It summarizes web pages, PDFs and videos, lets you ask follow-up questions, translates whole pages, and lets you screenshot a chart and ask about it. Video summaries come with timestamps you can click to jump to that moment.

I built it alone, pairing with an AI coding assistant, in about a month. This post isn't a tutorial. It's about everything around the code, which turned out to take most of the time.

The build: a month, ~80 small PRs

Roughly:

• Week 1: wrote a plan, then cut it down to the smallest loop that works: side panel, read the page, chat, sign-in, a daily free quota, an admin dashboard. Everything else went to a "later" list.
• Week 2: first working version, shared as a zip with a few friends.
• Week 3: images and screenshots, PDFs, full-page translation, video summaries from captions, one-click actions that change with the page type, 10 UI languages.
• Week 4: legal pages, store assets, a promo video, SEO, payment research, analytics.

Every change was one task → one branch → one PR → merge. Around 80 of them.

What worked when pairing with an AI coding assistant

1. You own the scope. The assistant will happily add features. Deciding what not to build is still your job.
2. Every change needs proof. I asked for tests plus a run in a real browser, with screenshots. "It should work" didn't count.
3. Let it push back. Several times it flagged a wrong assumption before writing any code: a reference project didn't actually have the feature I thought it had, or a payment provider didn't support the payment method I wanted. That saved days.
4. Write the rules down. "API keys never reach the client" lives in the project rules and is checked on every build. I don't rely on remembering it.

The hard parts weren't code

1. Store submission is its own project

• The privacy form is detailed. You explain every permission, every category of data you collect, and whether you run remote code. Vague answers get rejected. Optional host permissions, requested only when needed, made review smoother.
• Developer accounts have publishing limits. I hit the cap and had to unpublish an old item first.
• EU trader disclosure. If you plan to charge, you'll likely declare as a trader, and your contact details become public on the listing. Decide early what details you'll use.
• Your title is your pitch. "SideDog" alone told people nothing. "SideDog: AI Summary for Web, PDF & YouTube" did much better.
• Reviewers really sign up. Accounts named "Test Account" and "ms test" appeared in my dashboard during review. Make sure a reviewer can finish sign-in, and write clear test notes. Edge asks for them on every submission.

2. Half my early users couldn't open the store

My first real users came from China, where the Chrome Web Store isn't reachable without a VPN. People got interested, clicked "Add to Chrome", and hit a dead page.

What fixed it:

1. A zip download with step-by-step "load unpacked" instructions. It works, but a lot of people give up partway.
2. Publishing on Microsoft Edge Add-ons too. It's reachable there, installs in one click and updates automatically.
3. Both buttons on the homepage.

Lesson: the store your users can actually reach matters more than the biggest store.

3. Payments for a solo developer

I compared several merchant-of-record providers. Low setup effort, higher fees, and in some cases payouts only to a personal account, which raises tax and accounting questions you should ask an accountant early. Some providers also have extra review requirements for AI products: model disclosure, published pricing, refund policy, support contact.

My decision: launch as a free beta first. With single-digit users, conversion rates mean nothing. Prove that people use it and come back, then turn on billing.

4. You need your own numbers from day one

Store dashboards lag, and the numbers mix in reviewers and your own test installs. I added, on my own backend:

• Source tracking: every channel gets its own ?ref= link, recorded at sign-up.
• A sign-up funnel: saw the sign-in page → clicked → succeeded, failed or cancelled.
• An uninstall survey: one question, six reasons.
• In-extension feedback.

One surprise: my store stats said 71% of installs were from the US, but 86% of installs used Chinese as the language. Those were VPN users. Language was more trustworthy than region.

Honest numbers

• About 20 store installs in the first month. Several were reviewers or me.
• Real external users: single digits.
• The first real user came from a Chinese video site where I posted a demo. That first video got 37 views; the thumbnail didn't show the pain point.

Not impressive. But it told me who the users are (students and people learning from long videos), which channel works, and that installation friction was the biggest leak.

How SideDog is different

There are excellent all-in-one AI extensions. SideDog deliberately does less:

• One job: understand the page in front of you.
• Video summaries with clickable timestamps, on YouTube and other video sites.
• One-click actions that change with the page: video outline on videos, "paper at a glance" on papers, pros & cons on product pages, a job breakdown on job posts, a reply draft on emails. No prompt writing needed.
• Reads the page only when you ask, and keeps your chat history in your browser.
• Free during the beta, no card required.

Takeaways for anyone building a small AI tool

1. Ship the smallest loop that works, then add features.
2. Write down what you're not building.
3. Distribution beats features: stores, compliance, reachability, channels.
4. Instrument from day one. You can't improve what you can't see.
5. Small, honest numbers are still signal.

If you read a lot of long content, I'd love for you to try it: sidedog.7m7s.tech (Chrome and Edge).

I'm also launching on Product Hunt on November 7. If you want a reminder, you can follow it here: SideDog on Product Hunt.

What's the "too long to read or watch" moment in your day? I'll use the answers to decide what to build next.