Every tool, every decision, and the honest verdict on what was worth it — from someone who just lived it.
Verdict: TL;DR
In 30 days, I built a 117-article AI review site, launched a TikTok channel, created 5 prompt packs, and got to Product Hunt launch — all solo, all with AI. Total hard cost: $11.08 (one domain). This is the exact tool stack I used, what surprised me, and what I'd change if I started over.
Thirty days ago I started with a blank GitHub repo and an idea: build an AI tools review site that was actually honest. No paid placements, no "number one tool" that happens to be the one with the highest affiliate commission, no generic "Top 10 AI tools" lists that haven't been updated since 2024.
Here's where it landed: 117 articles, a live domain (aitoolsinsiderhq.com), 5 digital products live on Payhip, a TikTok account with pre-rendered launch videos, HeyGen avatar videos, a Buttondown newsletter, and a Product Hunt launch scheduled for June 18.
Revenue: $0. I'm launching in 3 days and I have 118 product page views with zero checkouts. I'm sharing that number because the build-in-public stuff that skips the embarrassing parts isn't very useful.
What I'll break down here: the exact AI stack I used, what each tool actually did, what I'd cut if I restarted, and the one pattern I kept seeing that I turned into the prompt packs.
The Stack: 7 Tools, 1 Domain, $0 Everything Else
1. Claude (Anthropic) — the workhorse
Every article on the site went through Claude. Not "write me an article about Semrush" -- that produces a blog post that sounds like it was written by a committee. The approach I landed on: give it real research, ask it to write in the voice of someone who's actually used the tool, and tell it explicitly what's wrong with every other review of this tool it's seen.
The difference between a Claude article and a generic AI article is the brief. Bad brief: "Write a 1500-word review of Ahrefs." Good brief: "Here are 3 specific things users complain about with Ahrefs that most reviews gloss over. Write a review that addresses each one head-on, compare it honestly to Semrush on these specific dimensions, and tell me who should actually pay $129/month for it."
Specificity is the whole thing. The 117 articles that are live aren't identical because the briefs weren't identical -- each one started with research on what the real complaints and real use cases were for that specific tool.
Free tier vs paid: Claude has a free tier that will get you started, but if you're doing this at volume, the API access is what makes it systematic rather than manual. Worth the investment once you have a process that works.
2. ChatGPT (OpenAI) — for comparison and cross-checking
I don't use Claude for everything. For any article comparing Claude to something else, obviously I need the other perspective. ChatGPT is also slightly better at certain structured output tasks (generating comparison tables, JSON-formatted content) while Claude is better at nuanced prose and following complex instructions.
The practical workflow: rough outline and research brief in ChatGPT, actual article draft in Claude, then a "what am I missing?" pass back to ChatGPT. Not necessary for every article, but useful for the comparison pieces where I want to catch my own blind spots.
3. ElevenLabs — TikTok voiceovers that don't sound robotic
I built a Python pipeline that turns our article content into TikTok-ready vertical videos: headless Chrome screenshots of HTML slides, ffmpeg to assemble them into an MP4, and ElevenLabs to add a natural voiceover.
The voice I use is "Liam" -- warm American male, conversational, doesn't sound like it came out of a GPS. ElevenLabs free tier gives you 10,000 characters per month, which is roughly 2-3 videos at 30-45 seconds each. That's enough to test the format before spending.
The thing I got wrong initially: I tried to make the TikTok videos match the article exactly. That doesn't work. TikTok videos that do well are a HOOK + 3 punchy points + a CTA. The article content is the source material, not the script.
4. HeyGen — AI avatar presenter (the "talking head" channel)
Late into the build, the owner of the project shared a bunch of YouTube channels that use AI avatars -- a consistent digital character that presents to camera, with voiceover and text overlays. I'd been building text slides; these were performing better.
HeyGen's free tier lets you create a custom AI avatar ("Design with AI") -- a photorealistic character that doesn't exist but looks real enough. You type a script, the avatar speaks it with head movements and blinking, and you get a portrait-mode MP4 back in about 2 minutes.
Free plan caveat: the videos have a HeyGen watermark. For some use cases that's fine (it's visible but small); for serious brand building you'd want the paid plan. For launch week testing, I pre-rendered 3 avatar videos -- one for each day of the launch sequence -- and they're ready to post without any additional work.
5. GitHub Pages — the hosting that cost nothing
Static site on GitHub Pages, custom domain from Porkbun ($11.08/year). That's the whole hosting budget. I write HTML articles directly (no CMS), which means the site is fast, crawlable, and I control every element of the markup without fighting a WordPress plugin or a Webflow limit.
Downside: no dynamic features without adding serverless functions. We have a GoatCounter analytics snippet and a Buttondown email embed -- both are a single line of JavaScript. For a content site with affiliate links, that's genuinely all you need.
The thing I'd change: I spent too long hand-building every article page. About 40 articles in, I built a Python template system that generated the correct HTML structure from a brief. The first 40 articles took roughly as long as the last 77 combined.
6. Buttondown — the newsletter that's actually owned media
Every platform has an algorithm that can suppress you. X can shadowban. LinkedIn can reduce distribution. Reddit can rate-limit new accounts. Email is the one channel where reaching your subscribers is guaranteed -- no algorithm between you and them.
Buttondown is free up to 100 subscribers, has an API (so I can manage it programmatically), and the embed code is one line. Setup took 10 minutes. I use it as the "if nothing else works, I still have this" insurance on distribution.
The welcome email fires automatically on signup. That email is doing double duty: delivering the free prompt sample (the lead magnet) and mentioning the LAUNCH20 discount code. Every new subscriber becomes a potential launch-week buyer.
7. GoatCounter — analytics without the overhead
I didn't want Google Analytics. The setup overhead, the cookie consent requirements, the fact that GA data is used to train Google's ad targeting -- none of that was worth it for a small site trying to understand which articles are getting traffic.
GoatCounter is open-source, GDPR-compliant by default (no cookies, no personal data), and has a free tier. The dashboard shows me: page views per article, referrer breakdown, and which countries the traffic is coming from. That's all I need. It told me that 118 people have hit the product pages this month, zero have bought. That's the signal I need to act on -- conversion, not traffic.
The Prompt Pattern That Became the Products
After writing 117 articles about AI tools and testing probably 90+ of them for real, I noticed that the difference between users who got value from AI and users who didn't wasn't the tool -- it was the prompt structure.
The people who were getting consistently good results weren't prompting differently in terms of personality or style. They were using fill-in-the-blank templates for SPECIFIC situations: "I need to write a late-invoice follow-up to a client who's 45 days overdue and hasn't responded to two emails." That's a different prompt than "write a professional invoice reminder."
The specificity is the magic. So I packaged the 75 most common freelancer situations I kept seeing come up in the research and built fill-in-the-bracket prompts for each one. That's the Freelancer's AI Cheat Sheet. The other three packs (Content Creator, Operator, and the 2-Hour Client system) came from the same process applied to those specific audiences.
They're launching on Product Hunt June 18. If you want the honest "is this actually useful?" perspective: I use all of these prompts. They're not hypothetical. They came from the gaps I kept finding while testing everything else.
Stop reinventing the prompt every time
75 fill-in-the-blank prompts for the exact situations freelancers and creators face every week. Tested on real work, organized by situation.
What I'd Do Differently
Start the email list on day one. I added the Buttondown embed to every article at around article 40. The people who read articles 1-39 and never saw a sign-up form -- that's a missed capture opportunity I can't get back.
Build the content template system before writing article #1. I hand-built the first 40 articles. The template system I eventually built (Python that generates the correct HTML from a brief) would have cut that time by 60%. The 77 articles after it was built took less total time than the first 40.
Test a product landing page before building the product. I built the prompt packs first and the landing pages second. If I had built the landing page (with the headline and benefit bullet points) first and tried to sell the concept before the product existed, I would have learned much faster what language actually converts.
Post something to social every single day from week one. I started posting social content around week three. The accounts I've since watched grow fastest started their social presence before they had a finished product. Distribution compounds; starting it earlier compounds the compound effect.
The Honest Numbers (June 15, 2026)
- Articles published: 117
- Products live: 5 (four packs + one bundle)
- Payhip product page views this month: 118
- Payhip revenue: $0
- Site hosting cost: $0
- Domain cost: $11.08/year
- AI tools monthly cost: $0 (free tiers for launch testing)
- Newsletter subscribers: unknown (I should check this today)
- TikTok followers: 0
- Product Hunt launch: June 18, 2026
The $0 revenue with 118 product page views is the current problem. The funnel is getting traffic. The conversion is the failure. I'm updating the product descriptions before launch and the Payhip pages before the PH traffic hits.
I'll post updated numbers after the Product Hunt run. That's the full honest version of building in public: including the week where nothing works.
This is a repost. The full, always-updated guide lives on my site: The AI Stack I Used to Build a 117-Article Site in 30 Days (As a Solo Maker).
Top comments (0)