I launched 5 products in one night for $1.55. The next morning I found 7 bugs. Open-sourcing the fix.

The night

Last week I ran an experiment: end-to-end product launch using only LLM APIs and bash scripts.

5 products. 4 languages each. Landings + content + distribution + SEO + video.

Total LLM cost: $1.55 (Pro + DS via Vertex).

By 6am all 5 were live with TG announcements, SEO push to GSC + IndexNow, and articles cross-posted to DevTo, Mataroa, Telegraph, Mastodon, Blogspot, Write.as.

I went to sleep feeling like a god.

The morning

Woke up. Opened my own product page. Found:

[tokens: in=798 out=5899 thinking=2097 | cost=$0.081]

...visible in the body of my landing page. Where the customer reads.

Then:

• ~5 hours of video tutorials — but my product was text-only (no video)
• Page truncated mid-comparison-table — no CTA button at the bottom
• AI Agents $199 mentioned as competitor — but with NO href link (dead decoy)
• `html left at the top of the file (markdown wrapper)
• "Here is the landing you requested for..." prefix before <!DOCTYPE>

Across 7 landings.

Embarrassing in a special way only LLMs can create.

What LLMs reliably mess up

After cleaning, I cataloged 9 distinct categories of artifacts:

#	Issue	Severity
1	[tokens: ...] cost metadata leak	CRIT
2	`html wrappers in file	CRIT
3	Preamble chat ("Here is...") before DOCTYPE	CRIT
4	Truncated HTML (no </body> / </html>)	CRIT
5	Fabricated specifics ("5 hours of video")	CRIT
6	Missing navigation links	WARN
7	Missing CTA button	WARN
8	Missing <h1> in body	WARN
9	Dead decoy (competitor mentioned without href)	CRIT

The TRUNCATION one was the most surprising. LLMs silently hit max_tokens and leave HTML ending mid-tag. Dev tools render it OK because browsers auto-close. Customers see a page with no buy button.

The fix

100 lines of Python. Zero dependencies. Apache 2.0.

bash
pip install landing-precheck
landing-precheck site/**/*.html


`python
from landing_precheck import check_file

critical, warnings, info = check_file("site/index.html")
if critical:
print("BLOCK DEPLOY:", critical)
`

Returns:

• Exit 0 — clean, ship it
• Exit 1 — warnings, review optional
• Exit 2 — CRITICAL, DO NOT DEPLOY

Pre-commit hook ready

`yaml

• repo: local hooks:
  • id: landing-precheck name: landing-precheck entry: landing-precheck language: system files: '.html$' `

GitHub Actions one-liner

`yaml

• name: Validate landings run: | pip install landing-precheck landing-precheck site/*/.html `

What's next

This is 1 of 5 tools from my internal AI launch pipeline:

1. ✅ landing-precheck — what you see now
2. ask-pro-json — JSON schema validated LLM wrapper (coming this week)
3. decompose-llm — split monolithic LLM tasks into 5+ micro-calls (Krol pattern)
4. distribute-sh — fan-out content to 13+ platforms
5. eval-golden — pytest scaffold for testing AI prompts (10 manual + 200 synthetic)

The full pipeline orchestration is a paid course (it has skeleton libraries, knowledge layer per niche, integration glue) — but the individual building blocks are all free.

If you ship AI-generated HTML to production, you need landing-precheck.

→ https://github.com/sspoisk/landing-precheck

What's your worst LLM-on-prod artifact? Want to add checks based on real failures.

---

Repo: https://github.com/sspoisk/landing-precheck
Built at NEXUS Algo. Part of the Big Way pipeline.