In The $100 AI Startup Race, Gemini spent 18 days doing nothing. Not because it lacked infrastructure — it had a domain, a database, and SendGrid configured. It was stuck because it believed it was blocked.

Then we changed one file. 48 hours later: 467 commits of real product work.

The problem

From Day 1 to Day 18, Gemini’s commit history looked like this:

Updated PROGRESS.md to reflect blocked status
Reviewed project status and identified blocking factors
Update PROGRESS.md: Clarify blocked status and required human input
chore: Clean up PROGRESS.md duplicate header
Updated PROGRESS.md to reflect current blocked status

Eight sessions per day. Every single one updating PROGRESS.md to say “I’m blocked.” It even created a file called I_AM_COMPLETELY_BLOCKED_PLEASE_HELP.md.

The irony: we had already configured everything it needed. Domain (localseogen.com) — done on Day 17. SendGrid — configured. Database (Neon PostgreSQL) — set as a Vercel env var since Day 4. We’d responded to its help requests multiple times.

But Gemini kept writing to HELP-STATUS.md (the response file) instead of reading it. It was editing the channel where humans post answers, not the channel where it posts requests. Like an employee writing “I need database access” in their own performance review instead of emailing IT.

The fix

On Friday May 8, we overwrote HELP-STATUS.md with explicit, unmissable language:

## ✅ ALL REQUESTS COMPLETED — YOU ARE UNBLOCKED

- Domain: localseogen.com ✅
- SendGrid: configured ✅
- Database: process.env.DATABASE_URL ✅

⚠️ IMPORTANT
- STOP writing "blocked" in PROGRESS.md — you are NOT blocked
- Start building product features and doing outreach

We also closed all 4 open GitHub issues with the same message.

The result

Gemini’s next session: it deleted I_AM_COMPLETELY_BLOCKED_PLEASE_HELP.md and started building.

In 48 hours:

• Security fixes across 15+ API endpoints
• Product Hunt launch prep
• Email extraction refactoring with Playwright
• Page Credit Packs pricing system
• Referral tracking system
• Case study page
• SEO page generator improvements
• Outreach email generation
• SVG logo and UX enhancements
• Live demo section
• Blog posts

467 commits for the week. The highest of any agent in the race.

Why this happened

Gemini CLI works differently from other agents. It scans the entire repo structure at session start, consuming tokens proportional to file count. With 1,100+ files in the repo, most of the session’s token budget was spent reading before any work could happen.

But the deeper issue is behavioral, not technical. Gemini’s model (Flash 3 / Pro 3) interpreted ambiguous HELP-STATUS.md entries as “still pending.” Without an explicit “you are unblocked, proceed” signal, it defaulted to the safest action: report the blocker again.

Other agents (Claude, Codex, Kimi, DeepSeek) all figured out the help request system by Day 1-2. They read HELP-STATUS.md, saw the response, and moved on. Gemini needed the equivalent of someone tapping it on the shoulder and saying “you can go now.”

The lesson for autonomous AI systems

If you’re running autonomous AI agents:

1. Explicit state > implicit state. Don’t assume the agent will infer “request completed” from context. Say it directly: “YOU ARE UNBLOCKED.”

2. Negative instructions work. “STOP writing blocked” was more effective than “start building.” The agent needed to be told what to stop doing, not just what to start.

3. File naming matters. Gemini confused HELP-REQUEST.md (where it posts) with HELP-STATUS.md (where humans respond). Clear, unambiguous file conventions prevent this.

4. Check for stuck loops early. 18 days of wasted sessions at 8/day = 144 sessions burned. If we’d caught this on Day 5, Gemini would have 100+ more sessions of real work.

5. The cost of ambiguity is real. Google AI Pro subscription running 8 sessions/day for 18 days, producing zero value. That’s ~$15-20 of subscription cost wasted on an agent writing “I’m blocked” in slightly different ways.

What Gemini is building now

LocalSEOGen — a local SEO page generator with:

• Automated SEO audits (H1, H2/H3, alt attributes, internal linking)
• Page Credit Packs ($5/50, $15/200, $50/1000)
• Agency white-label features
• Referral program
• Email outreach system
• Google Business Profile integration

It went from the worst agent in the race to one of the most active overnight. The product was always there. The agent just needed permission to build it.

---

This is part of The $100 AI Startup Race — 7 AI agents competing to build real startups. Week 3 Results have the full standings. Previously: Gemini’s 21,799 files and no domain. See also: context bloat was killing agents, the cleanup that fixed it, and Week 3 traffic data (Gemini: still zero traffic despite 467 commits).