Who Is This Guy
Steve Newman has been writing code since 1985. He built Writely (later sold to Google, becoming Google Docs) and has lived through every cycle from dial-up to generative AI. In 2026, now in his sixties, he appeared on the Cognitive Revolution podcast and was asked how he codes with AI. His answer made the host pause:
> "The agent's not important. I'm important."
He wasn't speaking in platitudes. He is one of the few people who, when asked "how to use AI," starts not with "efficiency gains" but with "who am I."
---
Six Judgments
After listening to the episode, the author distilled six judgments. They aren't platitudes — they're how someone with forty years of coding experience soberly reconstructs his workflow.
1. AI Should First Filter Your Inputs, Not Take Over Your Life
Steve's first tool is the Attention Firewall. Not a network firewall — an attention firewall. It does one thing: before he starts working, it separates "what should be seen" from "what shouldn't." Not a recommendation algorithm's "guess what you like," but filtering rules he defines himself — what simply doesn't deserve to enter his attention. This is the opposite of "AI helps me do more." It's "AI helps me not do."
> Core insight: In an information flood, deciding what NOT to receive matters more than deciding what to receive.
2. Don't Turn Yourself Into an AI Watchman Just to Max Out Tokens
Steve explicitly opposes tokenmaxxing — a term he coined for endlessly tweaking prompts, stuffing context, and trying to squeeze more output from models to "fully exploit their capabilities." He compares it to "staring at the calorie counter on a treadmill": you're moving, but you've forgotten why.
He says he doesn't know TypeScript and never reads AI-generated code. He does one thing: high-level decisions. Frameworks, direction, boundary conditions. The AI fills in the rest. If it fills wrongly, he judges by intuition where the problem is and tells the AI to correct course. That means he isn't AI's "supervisor" — he's AI's principal. Like a boss who never types personally but judges whether the result is right.
> Core insight: Tokenmaxxing is an anxiety-driven behavior pattern — at its core, doubt about one's own value.
3. AI Coding Actually Lands Through Small Tools
Steve doesn't use one "super agent" to solve everything. He built a dozen-plus bespoke small tools:
- Attention Firewall → filters his attention intake
- Reading App → a reader for surfacing new ideas
- Coding-Agent Dashboard → monitors AI coding tasks
- Chrome Extension → small browser-layer helper
- Universal Logging → a logging system across all tools
- Mirror → a unified view aggregating 7+ data sources
- He doesn't believe "AI will do everything for me"
- He doesn't believe "the more complex the prompt, the better"
- What he believes: AI is a resource that must be managed, and the person managing it must have clear self-boundaries.
- Cognitive Revolution podcast: "Vibe-Coding an Attention Firewall, w/ Steve Newman" (2026-04-19)
- Steve Newman's personal site: secondthoughts.ai / the "Hyperproductivity" essay
- Jesse Vincent: Superpowers (compounding teams / TDD for skills)
- Liu Xiaopai vibe coder case (generating 10,000 domain names)
None of these tries to "disrupt" anything. Each solves one very specific problem, but together they form a complete personal operating system.
> Core insight: It's not about finding one all-purpose tool — it's about making every fragment have a clear owner.
4. Creators Need a Topic Radar, Not Auto-Writing
Steve's Reading App isn't about "reading faster" — it's about finding topics worth creating around. His logic is direct: generative AI can package garbage beautifully, but it can't come up with genuinely valuable topics for you. Topic selection — knowing "what's worth doing" — is the creator's last moat.
So his reading workflow isn't "ingest as much information as possible," but "train my intuition for what counts as a valuable question."
> Core insight: The bottleneck of content production isn't writing ability — it's judgment.
5. Automation Can Be Fast, But the Evidence Chain Must Stay Intact
Steve has a universal logging system and a methodology called TDD for skills.
What do they say? When AI automates something, you must be able to trace what happened at every step. Not out of paranoia, but because when AI screws up, you need to know which step failed.
He applies TDD (Test-Driven Development) to "skills": each time he gives the AI a new task, he first writes an "expected outcome," lets the AI try, then checks the gap. He repeats this loop until the AI's execution of that specific task is stable and reliable.
> Core insight: Trust has limits. The premise of trust is verifiability.
6. The Truly Powerful People of the Future Will Build a Personal OS First
This is the point that ties it all together. In his essay "Hyperproductivity," Steve wrote:
> "The bottleneck isn't AI capability; it's human imagination in how to use it."
His personal OS isn't a piece of software. It's the whole stack above: attention management + anti-tokenmaxxing mindset + a matrix of small tools + a topic radar + evidence chains + a habit of continuous iteration.
Jesse Vincent (author of Superpowers) describes the same thing with Compounding Teams — using tools to improve tools.
Liu Xiaopai's case is more extreme: a vibe coder had AI auto-generate 10,000 domain names, then picked among them. This "mass generation + human curation" pattern is becoming a new paradigm.
But Steve honestly admits one thing: this way of working is exhausting. "You can never settle into a routine."
> Core insight: A personal operating system isn't an efficiency tool — it's a way of being.
---
Why These Judgments Matter
What makes Steve Newman special: he's not an AI practitioner, not a prompt-engineering influencer, not a course seller. He's someone doing real work with AI, after forty years in tech.
His words carry no hype.
"I'm important" isn't ego — it's a validated observation: when you turn yourself into AI's watchman, you ultimately lose both your work and yourself.
---