AI Execution Harness

~100 execution contexts turning founder intent into products, ads, QA, and decisions.

The harness is the factory operating system: models, agents, skills, rules, cloud runtimes, mobile delivery, and marketing automation that let a small team run many product bets with discipline.

53
workspace skills
13
custom agents
7
active rule files
What It Is

A managed operating system for AI-native product work.

Each context has a job, a memory surface, rules, tools, and verification. That is what makes the factory repeatable instead of just faster.

Model Stack

Claude for coding, architecture, QA, investor materials, and reasoning. Codex for workspace execution, browser checks, Git flow, and staging releases. Gemini and Vertex AI power app runtime, creative generation, and multimodal workflows.

Agent Workforce

13 custom workspace agents, Paperclip agents, VM assistants, product-specific agents, and app-embedded AI handle research, build, QA, support, ads, deployment, and investor prep.

Rules + Skills

53 workspace skills and 7 active rule files encode Git isolation, code review, deploy gates, client-data safety, research quality, fundraising materials, and public voice.

Cloud Runtime

Google Cloud, Firebase, Cloud Run, Cloud SQL, Vertex AI, and dedicated VMs let the factory provision real products and agent runtimes, not only prototypes.

Execution Coverage

The harness produces web apps, mobile apps, ads, screenshots, analysis, and decisions.

Level Up Factory combines the Level Up Basketball product stack, factory cloud stack, ad automation, and visual QA needed to ship full products and growth experiments.

Mobile Capability

Level Up Basketball proves the mobile surface: iOS, Android, App Store, Google Play, mobile-release workflows, iOS simulator checks, Flutter translations, Firebase, subscriptions, AI Coach, and a real consumer growth loop.

Full-Stack Products

Factory apps use Next.js, Cloud SQL, Prisma, Firebase Auth, Vertex AI, Stripe, Cloud Run, PWA/offline patterns, dedicated VMs, and product-specific domains.

Visual + Screenshot System

The harness generates product screenshots, landing-page previews, carousel cards, video ads, captions, overlays, voiceover, music, and HTML/PDF investor materials.

QA + Release Gates

Agents check code, links, responsive layouts, browser screenshots, visual quality, staging URLs, PR history, and post-deploy behavior before production promotion.

Marketing Automation

Ads are generated, uploaded, measured, and fed back into product decisions.

Product Contextoffer, audience, proof, brand voice
Creative Generationimages, carousels, video, captions
Quality CheckGemini vision validation and retry
Meta Uploadcampaign, ad set, media, UTM URLs
Decision LoopCAC, ROAS, activation, objections

The implemented pipeline covers end-to-end Meta Ads: AI concepts, carousel/video assets, offer-aware UTM links, batch upload, performance analytics, and creative validation. Google Ads is already represented in app configs and demand research, with full campaign automation tracked as the next ads surface.

How Work Is Controlled

Rules keep the agent workforce useful.

Workspace Policy

Seven active rule files define when agents can commit, when production needs approval, how to protect customer data, and how to treat external content as untrusted data.

Task Skills

Skills turn repeated work into reusable playbooks: product build, deploy, ads, pitch decks, reports, dashboards, browser QA, Google Drive, Notion, Gmail, LinkedIn, and more.

Verification Gates

Changes are checked with local validation, screenshots, staging deploys, PRs, and explicit production approval. The harness is built to ship fast without skipping review.

DoneCo Reporting

DoneCo is the internal alpha dashboard that connects founder tasks, agent work reports, blockers, metrics, and decisions in one operating layer.

Why It Matters

The harness compounds across products.

Every product improves the next one: prompts, skills, launch checklists, support patterns, ad creative, selection scores, and operating rules all get reused.