charli decouze

I am a product designer, not a coder, and so "lowcode" code editors leave me lost. It simply takes too long to ask Claude to step-by-step me through anything I want to do.
I have projects in the works in Lovable.dev and Medo.dev, Anything and Dazl - but yet to get things to work after migrating them to Vercel breakl out of the co-branding that these app-builders like to sneak in (notice the rainbow you see here and there on Lovable-built sites, even whn they use their own domain...?)
Gemini's suggestions:
AI code generation is prone to the "90% Problem," where the final 10% of polish requires significant manual intervention.  

• Bolt.new: Users report that while blazing fast (app skeleton in 20s), it often "shits the bed" when projects get complex, leading to massive token consumption to fix errors .

• Lovable: Praised for UI aesthetics, but its 7-message memory limit leads to "recency bias," causing the AI to get "drunk" on different approaches during long debugging sessions .

• Tempo Labs: While it provides excellent planning features (PRDs and diagrams), some users have reported that its export to GitHub is unreliable, providing only "fragments of code" rather than a full app .

• Blink.new: Emerging as a smoother alternative for full-stack builds, handling authentication and backend integrations with fewer breaks than Bolt .

Strategic Implementation Path

1. Selection (Phase 1): Use Dyad.sh or Anything if you want the most stable backend/auth foundation .

2. Generation (Phase 2): Build your MVP on the Pro tier for 1 month (~$19-$25). Use Discussion Mode to plan changes without burning credits .

3. Sync & Export (Phase 3): Sync your project to a private GitHub repository immediately .

4. Deployment (Phase 4): Connect the repo to Cloudflare Pages for free, unlimited bandwidth or Coolify on a $5 VPS for total control .

5. Refinement (Phase 5): Use Cursor (locally) for further edits. It provides a full IDE environment with better project context for $20/month, allowing you to maintain your "clean codebase" without builder overhead .
   I then asked, about a cheap solution starting with Baidu's Medo.dev - because Medo gives us many credits to start, 100 more/day free, 20 more for 20 clicks on other's projects... So you can build a lot, download, but then have to sort out login and other backend problems. Possible, but takes some doing.
   Gemini's response:

   1. In Supabase Dashboard: Go to Authentication > URL Configuration. Change the "Site URL" from http://localhost:3000 to your actual production domain (e.g., https://your-app.com).

   2. In Vercel Dashboard: Go to Settings > Environment Variables. Add a new variable called NEXT_PUBLIC_APP_URL and set its value to your domain. This tells the backend where to send users when they click the email link.

   3. In Cursor: If the error persists, open your project in Cursor and ask the AI: "Search for all instances of 'localhost:3000' in the authentication logic and replace them with the environment variable for the site URL.".
      But what about you? What is your process if you are no-code building and shipping?

Hello,

I ve been working on a passion project I called Sapling (use.sapling.mobi) on and off since 2016.

I have spent the last 12 months improving it because I kept seeing small charities lose huge chunks of their donations to complex platforms that take high fees or guilt-trip donors into "tipping" the software. It felt wrong, so I decided to build a cleaner, simpler alternative.

While testing Alpie Core beyond benchmarks, we noticed something unexpected.

On tasks like step-by-step reasoning, reflective questions, and simple planning ( help me unwind after work , break this problem down calmly ), the model tends to stay unusually structured and neutral. Less fluff, less bias, more explicit reasoning.

It made us wonder if training and serving entirely at low precision changes how a model reasons, not just how fast it runs. Sometimes the chain of thought itself is something you d actually want to read to understand the reasoning behind the final answer.