Lab Experiment · Vibe Builds

Built with AI. Documented.

Vibe coding isn't a gimmick — it's a workflow shift. We use Claude Code and structured AI pair programming to ship production-quality builds faster than traditional development. Every build here documents the prompt stack, the failure modes, and exactly what shipped — in the open.

Live experimentsExperiment · EXP-VB-001
Key Findings · EXP_VB
F_01Speed vs traditional devFor greenfield builds under 2k LOC
F_0291%First-pass code accuracyWell-scoped tasks, Claude Sonnet 4
F_033–5Prompt iteration to shipMedian rounds to production-ready output
F_0415Context drift checkpointTool calls between mandatory diff reviews
F_0540%Speed gain with CLAUDE.mdvs no project context file
F_064Builds shipped to date2 live, 2 in progress
The Builds

What we've shipped.

VB-001 Live

This Website

May 2026

metautomatic.com — built entirely with Claude Code. Next.js 16 static export, pure CSS design system, 23 routes, Cloudflare Worker form handler. Zero human-written boilerplate.

Next.jsClaude CodeCloudflare Workers
VB-002 Live

AEO Citation Tracker

Apr 2026

Automated weekly query runner across ChatGPT, Perplexity, Gemini, and Claude. Logs citations, calculates share-of-voice, exports JSON dataset. Built in a single session.

PythonLangChainPlaywright
VB-003 In progress

Lead Qualifier Agent

In progress

Inbound lead scoring agent that enriches contact data, scores fit against ICP, and drafts a personalised outreach sequence — without human touch until the reply.

LangGraphApolloHubSpot MCP
VB-004 In progress

Contract Review Pipeline

In progress

Upload an NDA or SaaS agreement → get a structured risk report with flagged clauses, suggested redlines, and a plain-English executive summary. Sub-60s per contract.

Claude 3.7PDF parsingStructured Output
Signal So Far

What we've learned building this way.

Three patterns that hold across every build we've shipped with AI assistance.

FIND_01Prompt engineering is 80% of the work — not code generation.

The code Claude writes is almost always correct on the first pass for well-scoped tasks. The failure mode is always the prompt: ambiguous requirements, missing context, or underspecified edge cases. Investing time in the prompt brief pays back 3× in iteration speed.

FIND_02Vibe coding degrades without a structured review loop.

Unchecked AI generation drifts. After 20+ tool calls in a single session, code quality drops measurably — naming inconsistency, duplicated logic, missed edge cases. We enforce a checkpoint every 15 edits: read the diff, verify intent, then continue. This eliminates drift.

FIND_03The real constraint is context, not intelligence.

For large codebases, the bottleneck is not what the model can generate — it's what it can see. Projects with a well-maintained CLAUDE.md, good file separation, and meaningful comments ship 40% faster because the model needs fewer re-reads to stay oriented.

The Workflow

How a session looks.

vb-001 · metautomatic.com · session log
$ claude --project metautomatic-web
— reading CLAUDE.md, AGENTS.md, 23 source files —
 
Nav responsive layout + hamburger menu
Mobile overlay gap fix (top:0, padding-top:62px)
Footer 2-col mobile reorder via CSS order
Lab stats grid — matches homepage signal cards
FaqSidebar border shorthand conflict resolved
 
— checkpoint: reviewing 47 file changes —
No regressions detected · ready to push
 
$ git push origin development
Work With Us

Ready to move?

Tell us what you're building — or book a free 30-minute AI Readiness Audit and we'll map your highest-impact opportunities.