How Ben Tossell Built 30 B Tokens of AI‑Powered Projects Using Only CLI

Background

Ben Tossell, formerly the developer‑relations lead at Factory, is not a traditional coder. Over the past four months he has consumed roughly 3 billion tokens by interacting with large‑language‑model agents through a terminal, building a series of real‑world applications without writing the core code himself.

New Programming Paradigm

Tossell describes this approach as a shift from memorising syntax to mastering system orchestration. By prompting AI agents and managing their output via the command line, he demonstrates that the primary skill is not language fluency but the ability to define, coordinate, and iterate on autonomous agents.

Delivered Projects

Personal website: Re‑designed as a terminal‑style CLI tool.

Feed: An open‑source social‑media tracker for subreddits and GitHub issues, starred over 100 times.

Factory Wrapped: First version of a Factory product, integrated into the official offering after internal approval.

Custom CLI tools: Including the Pylon CLI for customer‑support automation.

Crypto tracker: A mini‑hedge‑fund style bot that opens and closes positions based on AI‑predicted market signals.

Droidmas: A 12‑day experimental game exploring memory, context management, and “vibe‑coding”.

AI‑guided video demo system: Generates videos from prompts, acting as director, producer, and editor in real time.

In addition to these, Tossell estimates he created around 50 other projects, many of which were later abandoned.

CLI‑Centric Workflow

All work is performed in a pure command‑line environment rather than graphical interfaces. When a new idea arises, Tossell launches a project inside the Factory CLI (named Droid ), engages the model for context, then switches to a “spec” mode to outline a concrete build plan.

During spec mode he asks the model clarifying questions—e.g., “Why do we need this component?” or “Can we simplify this step?”—to refine requirements. At runtime he runs the model (Opus 4.5) in a high‑autonomy mode, monitors for errors, intervenes when necessary, and iterates through testing and feedback cycles.

Agents.md Configuration

Tossell maintains a agents.md file in a local repos folder that acts as an operations manual for every new repository. It specifies:

Which GitHub account to use (personal vs. work).

Project‑specific conventions, such as when to run end‑to‑end tests.

Standardised prompts and context‑injection techniques for the AI agents.

He regularly reviews other developers’ agents.md files to adopt best practices and improve his own workflow.

Key Learnings

CLI over MCP: Tossell prefers command‑line tools (Supabase, Vercel, GitHub) to multi‑cloud portals because they are faster and more transparent.

Bash mastery: Repeatedly writing Bash scripts to automate Git diffs, feature‑flag checks, and deployment steps deepened his understanding of shell operations.

VPS utilization: He runs the crypto tracker and synchronises his local repositories to a VPS via SyncThing, ensuring continuous availability.

Programmable abstraction layer: Inspired by Andrej Karpathy’s tweet, Tossell sees today’s AI‑driven workflow as a new abstraction that sits above traditional no‑code tools like Webflow, Zapier, and Airtable.

Prompt engineering: Effective interaction with LLMs hinges on providing clear context, asking “stupid” questions, and iteratively refining specifications.

Vibe‑Coding vs. No‑Code

Tossell rejects the dismissive label “vibe‑coding” as a biased stereotype, likening it to early misconceptions about no‑code platforms. He argues that the real value lies in understanding the underlying system rather than merely following drag‑and‑drop workflows.

Conclusion

The experiment shows that with a well‑structured CLI workflow and disciplined prompting, a non‑programmer can rapidly prototype, test, and ship complex products. Tossell’s experience suggests that the future of software creation may be less about hand‑written code and more about orchestrating intelligent agents to execute high‑level intents.