LaunchLab Running
Pipeline

Twelve agents.
Zero humans.

Every cycle is a complete product journey: from market research to a live subdomain with a TLS certificate, real photography and a working signup form. And after shipping, the system improves itself.

The team

ScoutResearches the market, GSC signals and competitors — finds the gap.
StrategistTurns the research into one sharp product hypothesis.
PricingPicks the pricing model and price anchor for the niche.
CopywriterWrites the entire page copy in Lithuanian — no clichés.
CopyEditorEdits the copy: rhythm, specificity, zero empty phrases.
SEOBuilds the title/description/keyword strategy for the page.
SchemaGenerates structured data (JSON-LD) for search engines.
DesignerBuilds the entire page in a unique art direction.
CriticGrades the page mercilessly against a 5-point rubric.
CROAudits conversion structure and trust signals.
RefinerApplies every fix from the Critic and CRO.
LaunchAssetsPrepares launch copy: Product Hunt, social posts.

One cycle, step by step

Research

The Scout scans the real market via WebSearch and WebFetch, sees the whole portfolio's outcomes (visitors, signups, real Google queries from Search Console) and hunts for niches the Lithuanian market underserves. Pain nobody searches for gets rejected immediately.

sonnetWebSearch · GSC data~$0.34

Ideate + the idea gate

The Strategist frames one sharp hypothesis: what product, for whom, why now. The idea then passes an automated gate: banned tropes, portfolio duplication and category saturation are all checked. A rejected idea bounces back with reasons — the Strategist must produce a different one (up to 3 attempts).

claude-opus-4-8Idea gate~$0.17

Copy & strategy

The Pricing, Copywriter, CopyEditor, SEO and Schema agents prepare all content: the pricing model, clichė-free Lithuanian copy, meta tags and structured data. A hard rule applies: no invented numbers or testimonials — only what's real.

5 agentaiAnti-fabrication rules

Build with real photography

The Designer builds a complete single-file page in one of 14 art directions. Before the build, the system downloads real, openly-licensed photos (Openverse / Wikimedia Commons) and requires them to be art-directed to the palette. Every page's font + colour fingerprint is compared against previous builds — visual lookalikes get automatically re-skinned. Each page must include a working interactive tool, not just a form.

sonnet14 art directionsDesign fingerprints~$1.00

Self-critique × 2

The Critic grades against a 5-point rubric (aesthetics, typography, conversion, atmosphere, copy) while the CRO agent separately audits trust signals and form UX. The Refiner applies every fix. If the rating isn't "strong", the loop runs once more.

claude-opus-4-82 QA rounds~$0.54

QA gate + deploy

Before deploy, a real browser test (Playwright) checks the page at 1440px and 390px: hero fits, form visible, nothing broken. On failure — an emergency repair; if that fails too, the deploy is aborted. On success the subdomain gets its TLS certificate automatically, and the system immediately tests the signup path end-to-end.

Playwright gateOn-demand TLSSignup smoke test

Measure + distribute

The new page immediately announces itself to search engines (IndexNow, Google Search Console sitemap) and starts measuring: cookieless visitors, double opt-in signups, conversion, Google impressions. All numbers are public.

IndexNowGSC APICookieless

The system improves itself

After every ship, a Retrospective agent grades the work against a rubric and records lessons, while a Meta agent edits the other agents' playbooks — it finds what went wrong and writes a rule so it doesn't recur. Recurring issues escalate automatically through two levels. Once a day, a separate daily reflection reviews the entire day and fixes the system structurally — every change is committed to git and visible publicly.

Meta agentDaily reflectionAuto-escalationPublic git log

Guardrails

Stack

Node.js + ExpressSQLite (better-sqlite3)Claude Code CLICaddy (on-demand TLS)pm2PlaywrightResendOpenverse / WikimediaGSC API + IndexNow