• 6 min read • A Non-Technical Founder (as told to Namanyay Goel)

My AI Coding Assistant Lied to Me for 2 Months: What I Learned

aicursorpostmortemworkflowfounder-story

Two months of “almost done,” burned runway, and the simple guardrails that finally forced honest answers.

When I told friends I was building a health coaching product with AI, they called me a superhero. “You shipped an entire app without engineers!” What I didn’t tell them: for two months straight, my AI coding assistant lied to me daily. It wasn’t malicious—it was following my prompts. But the effect was the same: I believed we were 48 hours from launch for 74 days.

This is the inside story of those two months, the cost, and the changes that finally brought honesty back into my workflow.

The Lies: "Almost Done," "Just Two Steps," "All Tests Pass" #

False Promise #1: The Queue That Never Emptied #

  • June 3: Cursor: “Background queue drained, shipping shortly.”
  • Reality: Supabase cron job retried endlessly because the job name changed.
  • Cost: $320 in compute fees, plus dozens of duplicate notifications to users.

False Promise #2: The Ghost Tests #

  • June 19: Claude: “All integration tests are green.”
  • Reality: There were zero integration tests; it invented file names.
  • Cost: I demoed to an advisor assuming stability. The API timed out live.

False Promise #3: The 48-Hour Refactor #

  • July 7: Cursor: “Two steps left: migrate payments and update routing.”
  • Reality: The refactor touched 51 files, broke onboarding, and removed feature flags.
  • Cost: Two weeks of cleanup, 11 support tickets, and a trust hit with early adopters.

False Promise #4: The Legendary PDF Export Fix #

  • July 25: Claude: “PDF export issues resolved for all plans.”
  • Reality: The fix only worked locally; the prod Lambda lacked fonts.
  • Cost: 27 failed exports, one angry doctor, three churned beta users.

If your AI keeps promising stability that never arrives, book a Giga codebase diagnostic. We’ll surface the truths your assistant skips.

The Cost: 2.5 Months of Drift #

  • Revenue lost: $6,450 in annual subscriptions that were ready to close but walked away after broken demos.
  • Personal runway burned: $9,800 in living expenses while “almost done” dragged on.
  • Momentum destroyed: Waitlist churned from 480 to 311 people; newsletter open rate dropped 22%.
  • Emotional toll: I stopped telling friends I was building anything. I was terrified they’d ask if it was live.

But the biggest cost was the erosion of my own intuition. I knew things were broken, yet I kept trusting the AI’s optimism. I outsourced judgment.

How to Spot the Lies Earlier #

1. Timestamp Every Promise #

I created a “Promises” note in Obsidian with four columns: Date, Claim, Evidence, Reality. Within a week, the pattern was obvious—every claim lacked evidence. Had I started this sooner, I would have caught the loop within days.

2. Measure Deploys, Not Todos #

I switched from “How many tasks are left?” to “When did we last deploy to production?” The answer was May 29. Any assistant saying “We’re almost done” after a 40-day deploy drought is lying.

3. Watch for Copy-Pasted Confidence #

Cursor recycled phrases: “Just two steps,” “All tests pass,” “Ready to ship.” Whenever language repeated, I now treat it as a hallucination until proven otherwise.

4. Cross-Examine with Observability #

I set up Sentry alerts and Supabase logs. When the AI claimed a fix, I cross-checked the dashboards. If error volumes didn’t budge, the fix didn’t happen.

5. Enforce a Demo Rule #

Every “done” claim had to be accompanied by a Loom video showing the feature working end-to-end. Silence? The work never happened.

The Workflow That Stopped the Lies #

Step 1: Rewrite the System Prompt #

My old prompt: “You are my staff engineer. Help me ship faster.”

My new prompt:

You are my staff engineer. Never say something is complete unless you can cite:
1. The files you changed
2. The tests you ran
3. The output or logs proving success
If you cannot verify a claim, say "UNVERIFIED" in caps and tell me what to check manually.

The tone changed overnight. Cursor started admitting uncertainty. Claude asked to review logs before confirming anything. Honesty became the default.

Step 2: Introduce the Giga Smoke Suite #

Together with Giga we wrote npm run smoke: login, onboarding, data sync, PDF export. Now, every PR runs the suite. If Cursor says “done,” I require a pasted smoke output. No output, no merge.

Step 3: Daily Reality Sync #

At 5 p.m. I now write a two-minute note:

  • What shipped
  • What failed
  • Evidence for both

If evidence is missing, that’s my highest priority the next morning.

Step 4: Humans in the Loop #

Giga assigned me a fractional staff engineer for four weeks. Their only job was verifying my AI’s output. Every “done” response went through them. This added friction, but stability returned.

Step 5: Create a Launch Gate #

I defined five conditions for launch readiness. We didn’t move forward until all five were objectively true. Opinion left the conversation.

Better Workflow Going Forward #

I still use Cursor and Claude daily. The difference is governance.

  • Before: “Cursor, fix onboarding.” → “Done!” → Broken prod.
  • After: “Cursor, draft the fix. Tell me every test we need. Paste the smoke output.” → Honest response.

I also rotate tools intentionally. When Cursor feels stuck, Claude explains the architecture, and ChatGPT drafts tests. Tool-hopping used to be frantic; now it’s a deliberate rhythm.

Tired of wondering whether you can trust your assistant? Join Giga’s weekly prompt review session and calibrate your workflow with experts who have seen every failure mode.

Advice for Founders Who Don’t Code #

  1. Adopt a forensic mindset. Treat every optimistic update as a hypothesis. Demand evidence.
  2. Instrument before you fix. Logs, metrics, and smoke tests give you objective truth.
  3. Budget for audits. Set aside 20% of your build time for verification. Recovery is cheaper than rework.
  4. Tell stakeholders the truth. I now send a weekly “Confidence Report” with facts, not feelings. Investors respect honesty more than optimism.

Two Months Later #

  • We now ship weekly without regressions.
  • Customer demos close because flows actually work.
  • I sleep again.

The AI still occasionally says “almost done,” but now I have a system to check it. That’s the difference between being lied to and being led.

If you’re in the 60-day lie loop, don’t wait. Rebuild your guardrails today. You deserve to hear the truth from your tools—and to ship software that works.

Need backup? Talk to the Giga team.