App AI Automation Founders & SaaS
Nilkick
Our own product: a free scanner that checks if a site is ready to launch, which grew into a paid app, a Chrome extension and a public API in nine weeks.
- Client
- Croissant & Baguette, own product
- Industry
- Founders & SaaS
- Shipped
- 2026
- Live at
- nilkick.com
Before
Launching is the hard part.
Founders build something good, release it, and nothing happens. The advice they find is generic. The real work is finding the venues that matter (directories and communities), reading the live discussions where their users already talk, and writing every post. Nobody does that well at eleven at night.
We wanted a product of our own that did that work, with the founder still pressing the final button. It is also where we test how we build AI systems for clients: AI costs measured on every run, fixed rules where the answer must not change, and a person in the loop by design.
What we built · 01
A free score, set by fixed rules.
Paste in a product's web address. Our scanner reads the page, using a real browser if it needs JavaScript to show. It checks how search engines index the site and where it is mentioned, then scores it out of 100 against 27 rules in five groups: findable, clear, crawlable, indexed, referenced. A separate score for AI agents runs 19 checks, from llms.txt to MCP and OAuth discovery, the files agents look for.
Fixed, versioned rules turn what the scan found into the score. No AI model touches it, so the same site gets the same score on every refresh, and each report has a fixed, shareable link. The launch map loads in after it, using only real venues from a hand-picked list of 100, so the AI cannot invent one.
What we built · 02
Three products for the launch work.
A Launch reads the product's whole site and builds a brand and audience profile for the founder to approve. It then lists venues and live Reddit and Hacker News threads in order of value, with every post drafted, to go through one by one. The Reddit warm-up is a 28-day plan of purely helpful comments that build the account's reputation, and by design none mentions the product. The Scout finds new Reddit and Hacker News threads daily, with the reply already written, and emails only on days with something new.
The one rule: Nilkick never posts for you. A person presses the real button for every post. Around Nilkick sits a marketing site with more than 50 guides. Each has a plain-text copy for AI tools, and the llms.txt file and structured data (labels for search engines) come from the same content.
What we built · 03
A Chrome extension, and an API for developers.
A Chrome side panel shows the same list next to the venue's page. Copy each field, press submit yourself, and the item is marked done in the app. It only copies, by choice: it runs no scripts on other websites and fills in no forms. Developers can also drive it from their own code through a public API. It has limited-access keys, signed notifications (webhooks), claiming work in batches and rate limits, and only spends on AI when a draft is asked for.
What changed
The numbers.
9 weeks
to build a paid product
free report, three paid products, extension and API
46
checks in the score
27 for launch, 19 for AI agents, none decided by AI
5
programs running live
the app plus four for background and scheduled jobs
400+
automated tests
plus a 22-scenario simulation of the warm-up plan
Counted from the code and the live system, not estimated.
Runs on
Our own product, and our testing ground.
Nilkick is where we try things before a client pays for them: the right AI model per task, AI spend measured on every run, and fixed-rule scoring next to the AI's judgement. Client systems run on the same tools.
- SvelteKit
the app and the marketing site, with plain-text copies for AI tools
- Cloudflare
five Workers (small programs), long-running and scheduled jobs, and AI Gateway
- Supabase
the database, and sign-in by email code
- Stripe
credits, the subscription and the billing portal
- Claude
Haiku for quick checks, Sonnet for drafting, Opus for deep brainstorming
- Serper · Bright Data · Algolia
search results, Reddit and Hacker News data
- Resend
email digests and account emails
- Vite + CRXJS
building the Chrome extension
Same Oven
More work like this.
Same Oven
