I Built a Startup Resource Hub With 157 Programs in 7 Days
Photo by Oliver Kiss
I spent seven days building a searchable resource hub with 157 startup credit programs across 17 categories. Total estimated value: over $600,000 in non-dilutive credits. It is live at driftingdesk.com/startup-hub.
This is the build story. What worked. What broke. What I would do differently.
Why I Built It
I have used startup credit programs for years. AWS credits. Google Cloud. Azure. Deepgram. Stripe Atlas. They saved months of runway at the MVP stage.
But finding them was always painful:
- Hours of Googling for each new program
- Reddit threads with half the information outdated
- Official pages that bury the real requirements in fine print
- No single place with credit values, difficulty ratings, and eligibility in one view
I wanted a tool. Searchable. Filterable. With real community feedback instead of marketing copy. So I built it.
The 7-Day Breakdown
| Day | Focus | Output |
|---|---|---|
| 1-2 | Research (Gemini Deep Research, thinking mode) | 75 programs, 18 data fields each |
| 3-4 | Scale database (4 more Gemini batches) | Expanded to 157 programs |
| 5 | Architecture design (Claude) | 3-tier page structure, full deployment spec |
| 6 | Build (Cursor + Next.js) | 220 static pages, deployed via Coolify |
| 7 | Polish + launch | Mobile fixes, OG images, analytics, sitemap |
Day 1-2: The Research
I used Gemini Deep Research in thinking mode. Regular Gemini gives surface-level answers. Thinking mode forces it to reason through each data point. The quality difference is significant.
18 data fields per program:
| Field | Why It Matters |
|---|---|
| Credit value | The actual dollar amount, not marketing fluff |
| Validity period | 6 months vs 2 years changes your planning |
| Company required | Huge blocker for early builders without an entity |
| Solo founder eligible | Many programs quietly require a team |
| Difficulty (Low/Medium/High) | Based on community feedback, not official claims |
| Approval rate | What percentage actually get approved |
| Geographic restrictions | Some US-only, others global |
| Stage requirements | Pre-seed, Seed, Series A |
| Community blurb | What Reddit and HN founders actually say |
Two days gave me 75 programs across 15 categories. But it was not clean. 19 programs had credit values rendered as images instead of text in the export. Had to pull those numbers manually.
Lesson: AI research tools are powerful but unreliable at scale. Verify everything.
Day 3-4: Scaling to 157
75 was not enough. I wanted the most comprehensive list available.
What failed:
- Asking Gemini for all 85 new programs in one prompt
- Result: 59/85 credit values missing, 75/85 blurbs blank
- It choked on the volume
What worked:
- 20-22 programs per batch
- Thinking mode on every batch
- Completion rate jumped from 30% to 95%
Final count after merging and deduplication: 157 programs, 17 categories.
Day 5: Architecture
I used Claude to design the system. Not the code. The architecture.
3-tier SEO structure:
- Hub page (
/startup-hub) - all 157 programs, search, filters - Category pages (
/startup-hub/cloud, etc.) - 17 landing pages - Program pages (
/startup-hub/cloud/aws-activate) - 157 individual pages
Total: 175 indexable URLs, each targeting different search keywords.
Claude also designed:
- Thumbs up/down rating system (coupon-site style verification)
- Community forms (report issues, submit new programs)
- Share buttons (LinkedIn, X, copy link)
- Sidebar quick view (see details without leaving search results)
- Complete Cursor deployment prompt (2,000-word build spec)
No code was written on day 5. Just system design.
Day 6: The Build
Fed the spec and JSON data to Cursor. It built the full Next.js structure in hours:
- All pages and routing
- Card grid, filter bar, sidebar modal components
- API endpoints for ratings, reports, and submissions
- Search logic with URL parameter sync
- 220 static pages generated at build time
Deployed through Coolify to my $8/month VPS. Three minutes from push to live.
Day 7: Polish and Launch
| Fix | Detail |
|---|---|
| Mobile responsive | Filter bar ate half the screen - replaced with compact icon bar |
| OG images | Dynamic generation failed on Docker - switched to static fallbacks |
| Analytics | GTM + GA4 installed, 30+ tracked events wired |
| Sitemap | 175+ URLs generated, submitted to Search Console |
| Launch post | LinkedIn, honest about the build process |
The Stack
| Tool | Role | Cost |
|---|---|---|
| Gemini (thinking mode) | Deep research, data collection | Subscription |
| Claude | Architecture, prompt writing | $20/month |
| Cursor | Code generation | $20/month |
| Antigravity | Scaffolding | ~$20/month |
| Next.js | Frontend | Free |
| Coolify | Deployment | Free (self-hosted) |
| Hostinger VPS | Server | $8/month |
What I Learned
Data collection took longer than the build. 60% research, 40% development. We can generate code in minutes. We cannot generate verified data that quickly.
AI research needs batching. One big prompt fails. Many small prompts succeed. 20 programs per batch was the sweet spot.
Community feedback beats marketing pages. Reddit and HN threads reveal actual approval timelines, hidden requirements, and whether credits are real or just asterisked numbers.
A directory beats a blog post. Nobody reads a list of 157 programs in an article. A filterable tool gets bookmarked.
SEO compounds. 157 programs across 175 pages creates a wide net for long-tail keywords.
What Is Next
- Community-driven updates (submit programs, flag outdated info)
- Reddit quotes, FAQs, pros/cons on every program page
- Star ratings based on user feedback
Free. No email gate. No paywall.