Ben Tossell's Non-Technical AI Coding

Table of content

Ben Tossell runs Ben’s Bites, one of the biggest AI newsletters out there. He started it six weeks before ChatGPT dropped — talk about timing.

But here’s the thing: he can’t code. Never could. He built 50+ products with no-code tools at Product Hunt, made $0 from them, then started Makerpad (a no-code tutorial site) which Zapier acquired in 2021. Now he codes daily using AI agents. The gap between “I can’t code” and “I shipped this” has basically disappeared.

3 Billion Tokens in Four Months

Every single one through a terminal, watching an agent write code he couldn’t write himself. He calls himself a “vibe-coder” but pushes back on that term — there’s real skill involved, just like no-code required skill.

He doesn’t read the code. But he reads the agent output religiously. That’s his version of learning to program.

What He’s Actually Shipped

Personal site — redesigned to look like a terminal CLI tool
Feed — social tracker for Twitter mentions, Reddit posts, GitHub issues (open-sourced, 100+ stars)
Factory Wrapped — built first version, team loved it, baked it into the product
Custom CLIs — Pylon CLI for support, Linear CLI, Gmail CLI
Crypto tracker — opens/closes short/long positions based on predictions
Droidmas — twelve experiments touching memory, context management, vibe coding
AI-directed video demo system — agents watch the screen during recording and respond to issues
Telegram bot — chat with repos from messaging app, synced with local files

The Workflow

Terminal over web interfaces. Always. He starts with an idea or a pain, spins up a new project in Droid (Factory’s CLI), then:

Talk to the model a few times to feed context
Switch to spec mode to plan
Question everything — “why this over that?”
Link docs and GitHub repos for the agent to explore
Let Opus 4.5 with high autonomy rip
Watch the stream, jump in when needed
Start server, test, give feedback, iterate

He builds ahead of himself first. The gaps and issues that come up? Those are learning opportunities.

agents.md Setup

He treats this as the instruction manual. His setup lives in a root repos folder and includes:

Rules for each new repo
GitHub setup instructions (work vs personal account)
Commit conventions
End-to-end testing requirements

He constantly looks at other people’s agents.md files to borrow ideas. The doc keeps improving.

CLIs Over MCPs

He stopped using MCPs. CLIs are simpler — you usually only need a few tools that an MCP would include. With Supabase, Vercel, and GitHub, he uses CLI versions exclusively.

He builds his own CLIs for things. Linear CLI to query issues. Gmail CLI for triage. These are portable — they live at the root directory and can be pulled into any project.

Coding on the Go

Every repo gets the Droid GitHub app installed. He submits pull requests so Droid can review them and make fixes with custom prompts. Triggered from issues or PRs.

For always-on stuff (like the crypto tracker), he uses a VPS. SyncThing keeps local repos synced so he can pick up where he left off from his phone.

Slack works too — one channel per repo, fire off things whenever. He calls it “a great 1-person product (+ agents).”

The New Layer of Abstraction

Back in the no-code days, the abstraction layer was drag-and-drop tools — Webflow, Zapier, Airtable stitched together to feel like real software.

Now the abstraction layer is the AI agent itself. You don’t need to learn code from scratch. You need to learn how to work with the agent.