This is the public lab where Frank builds real companies on the new agentic dev stack — Antigravity 2.0, Claude Code, Codex, Gemini API, Firebase, Android. Eight builds. One stack thesis. Receipts, not hype.
Why this matters now
The question is no longer which AI coding tool is best. The question is which builder knows how to command them — when to plan, when to dispatch, when to verify, when to ship. This lab is the receipts.
The 30-Day Build Log
Each build is a real company surface or working demo — not a sandbox. Status is honest: live, shipping, work in progress.
Shipped the flagship build-in-public lab and commercial hub on frankx.ai in a single Claude Code session, with cross-page interlinking and a reusable ACOS content-pipeline skill.
Building the SIS agent registry — the substrate that gives every Frank-operated agent persistent memory, capability scoping, and audit trails across Claude Code, Codex, and Antigravity sessions.
Building VibeClubs — the membership and offer engine for music-first creator clubs, with Stripe checkout, Resend onboarding, and a member dashboard powered by the Starlight substrate.
Prototyping the Arcanea creator tool — a world-building engine that takes a setting prompt and produces lore, characters, scenes, and visual references using Gemini for grounding and Antigravity for the UI scaffold.
A four-step chain that takes an idea from AI Studio, exports to Antigravity with context preserved, ships a Firebase-backed app, and produces a working Android target — all on the Google agentic stack.
A focused demo — single prompt into Antigravity, working Android APK out the other side. Tests the limits of agentic mobile development on the Google stack.
A content pipeline built on Gemini API Managed Agents that operates inside Gmail, Docs, and Sheets — researching, drafting, scheduling, and reporting without leaving Workspace.
A controlled benchmark — same dashboard feature, three agentic dev tools, full notes on time, output quality, failure modes, and which tool should own which layer of the stack.
The Stack
Picking one tool is the wrong frame. Each one earns a specific layer of the work. The architect routes the task to the right one.
Where new product surfaces are born. Generates UI from a brief, validates in a real browser, and produces walkthrough artifacts that double as demo assets.
Owns the existing codebase: deep refactors, test repair, cross-file edits, PR discipline. Plan-then-act mode keeps the diff tight.
Terminal-heavy work where the loop is shell-script-fast. Data validation, CSV pipelines, eval harnesses.
Grounded research with citations, Managed Agents for background tasks, AI Studio for prototyping, native Android target via Antigravity export.
The production backplane. Firebase for app data and auth, Vercel for the website, GitHub for the source of truth and CI.
Tool Matrix
A working architect's view — what each tool is best at, what it is acceptable at, and what still needs a human.
Methodology
Same loop every time. Specify, architect, build with agents, verify artifacts, ship with content. The ACOS skill handles step five.
Write the brief the way you would for a senior engineer joining tomorrow. Include the failure modes and the acceptance criteria.
Use a Plan agent to map files, components, data flow, and integration points before any code is generated.
Route the work to the tool that owns the layer. Antigravity for greenfield UI, Claude Code for repo work, Codex for terminal.
Every agent output becomes a reviewable artifact: a screenshot, a walkthrough, a diff, a test report. No invisible work.
Each build session produces the build-log entry, the LinkedIn draft, the demo brief, the diagram, and one reusable prompt. The ACOS skill runs the flywheel.
Templates
Every build adds one prompt or template to the pack. Forkable, documented, voice-spec clean.
The file every agent reads when it joins a repo. Roles, scope, escalation, banned operations.
Folder conventions, branch strategy, commit message patterns, PR review checklist.
Tool-specific reusable prompts: planning, refactor, test repair, content extraction.
A minimal eval scaffold to verify agent output against acceptance criteria before merge.
Want this for your team?
Field Notes
Honest write-ups from inside the lab. What shipped, what broke, which tool earned which layer. No threads, no thought-leader content.
No spam, ever. Unsubscribe anytime.