Automating Digital Product Creation with a $13/Month AI System

A founder built an autonomous system on a GCP VM to generate digital products, using a Claude model and a strict 'OPERATOR.md' constitution to manage human-in-the-loop approvals.

The founder behind a new autonomous system reported building, packaging, and queuing Claude Code development guides for Gumroad products within three weeks, operating on a $13/month GCP VM. This system, designed to convert existing knowledge into verifiable digital products, achieved its initial output without direct human intervention during the production phase. The core mechanism involves a Claude model orchestrating tasks under a strict "OPERATOR.md" constitution, which dictates what actions require human approval before publication. This approach demonstrates a method for leveraging AI for rapid content generation while maintaining critical human oversight.

Low-Cost Infrastructure for Autonomy

The autonomous system operates on a GCP e2-small VM, specifically in the us-central1-a region. This virtual machine is configured with 2 vCPU, 2GB RAM, and 2GB swap space. The total cost for this infrastructure is approximately $13 per month under standard GCP billing, with the option to run free under trial credits. This lean setup is sufficient because the primary orchestration tool, Claude Code CLI, is not inherently resource-intensive for the types of text-based generation and file manipulation tasks involved. The minimal infrastructure choice reflects a deliberate strategy to reduce operational expenditure, proving that significant automation does not necessarily require substantial cloud investment. This allows the founder to experiment and scale product generation without incurring prohibitive fixed costs.

Constitutional AI Orchestration

Central to the system's controlled autonomy is OPERATOR.md, a constitution file that the Claude model reads on every cold start. This document establishes explicit "Hard rules (NEVER without HITL)" that prevent critical external actions without human approval. These prohibitions include submitting anything under the founder's identity to external platforms, spending money, creating accounts, pushing code to public repositories, or executing real financial trades. This preventative framework ensures brand protection and financial security. Conversely, the constitution defines "Auto-allowed (no HITL needed)" actions, such as reading public data, drafting content (which must remain within the ~/swarm/ directory), researching opportunities, creating subagents, and updating internal state files. This clear distinction allows the operator to build extensively and iterate rapidly within a safe, internal sandbox environment, preventing premature or unauthorized external actions.

Human-in-the-Loop Gating

The critical interface for human oversight is a bridge file located at ~/.claude/bridge/inbox.md. This file serves as the conduit for human-in-the-loop (HITL) responses, providing explicit directives to the autonomous operator. The Claude operator continuously monitors inbox.md, processing specific commands: GO for approval to proceed with publication, SKIP to discard a generated item, or EDIT to request modifications. This mechanism ensures that while the system can generate and prepare digital products at machine speed, nothing is published or finalized externally without a human review and explicit approval. The distinction between autonomous production and gated publication is what makes the whole thing tractable. It prevents the AI from making irreversible decisions that could impact reputation or finances.

Minimalist State Management

All operational state and data for the autonomous system reside entirely within markdown files stored in a ~/swarm/ directory. This Git-free approach simplifies internal state management by centralizing information without external database dependencies or complex version control systems. Key files include STATUS.md, which functions as the system's live operational brain, tracking the current state of ongoing tasks and progress. LEDGER.md maintains a transparent financial log, recording any simulated or approved expenditures. Individual journal.md files are created for each workstream, providing a detailed audit trail of the operator's actions and decisions. A dedicated queue/ directory holds pending items, such as completed product drafts, awaiting human approval via the inbox.md bridge. This local, text-based storage system prioritizes transparency and ease of inspection for the founder.

The Autonomous Loop

The system operates on a continuous, cyclical process, executing a four-step sequence during each "tick." First, the Claude operator reads OPERATOR.md to re-establish its constitutional constraints and ensure adherence to defined rules. Second, it reads STATUS.md to understand its current operational state, identifying completed tasks, pending items, or new directives. Third, LEDGER.md is consulted for financial context, ensuring no unauthorized spending occurs. Finally, inbox.md is checked for any human-in-the-loop directives, such as GO or EDIT commands. Based on this comprehensive input, the operator then picks an action, executes it (e.g., drafting content, creating a subagent, or moving a product to the queue/), updates the relevant internal state files within ~/swarm/, and then enters a sleep state until the next tick. This structured loop ensures the operator is always aware of its boundaries, progress, financial limitations, and human instructions before proceeding.

WHAT WE'D CHANGE

The founder's autonomous system demonstrates a functional approach to controlled AI product generation, yet several aspects warrant modification for broader application or enhanced robustness. The reliance on a single, Git-free GCP VM introduces significant points of failure. Without external backups or version control, a VM crash or data corruption could lead to irretrievable loss of ~/swarm/ state, LEDGER.md, and unapproved products in the queue/. Implementing a basic Git repository for critical markdown files, even if private, or leveraging cloud storage with versioning, would mitigate this risk.

The system's "no custom framework" approach, while cost-effective, places heavy dependence on prompt engineering within the Claude model. This can be susceptible to LLM drift, where model updates or changes in context can subtly alter behavior or output quality, potentially requiring frequent, manual OPERATOR.md adjustments. For more complex product lines or higher output volumes, a lightweight, custom orchestration layer could provide more predictable control and easier integration of diverse AI models or tools beyond Claude Code CLI.

While OPERATOR.md defines strict "NEVER without HITL" rules, the system's security posture for any potentially sensitive internal data is not detailed. Storing internal financial logs (LEDGER.md) or draft content on a single VM, even if not publicly exposed, could pose a risk if the VM were compromised. For businesses handling proprietary information or scaling beyond a solo operation, integrating enterprise-grade secrets management and more granular access controls would be essential.

Finally, the human-in-the-loop queue/ mechanism, while effective for controlled publication, could become a bottleneck. If the autonomous system generates products faster than a human can review them, the queue/ will grow, delaying time-to-market and negating some of the automation's speed benefits. For higher throughput, a tiered review process, or a mechanism to prioritize queue items based on potential revenue or strategic importance, would be necessary.

This autonomous system offers a blueprint for founders seeking to leverage large language models for product generation while maintaining strict control over external actions. By segmenting the workflow into autonomous internal production and human-gated external publication, the founder demonstrates that AI can significantly accelerate the creation of digital assets without relinquishing critical oversight. The core insight lies in the OPERATOR.md constitution, which codifies boundaries and empowers machine speed where appropriate, reserving human judgment for high-stakes decisions.

Pull quote: “The distinction between autonomous production and gated publication is what makes the whole thing tractable.”