Newman built a system that ingests all messages (email, Slack, WhatsApp), uses an LLM with a personal rubric to determine urgency, and displays only critical items on a dedicated monitor. This reclaims focus time by eliminating the need to constantly check multiple apps for important updates.
Newman's most critical infrastructure for AI-assisted development is a universal logging service for all his apps (front-end, back-end, mobile). When a bug appears, he can tell an AI agent to "debug this," and it can analyze the comprehensive logs to find the root cause without guesswork.
When building a tool to manage newsletter overload, Newman abandoned plans for a complex system that would track his reading history to surface novel information. He found a simple, stateless LLM summary of each article was efficient enough to help him decide what to read, making the complex version unnecessary.
Traditional software engineering valued meticulous upfront planning to avoid costly coding and debugging cycles. Newman argues that with AI agents, the cost of building and iterating is so low that the old "measure twice, cut once" philosophy is obsolete. The superior modern approach is to build quickly, even incorrectly, and rapidly iterate.
The predicted explosion of AI-driven phishing and deepfakes hasn't happened. Newman finds this confusing but notes it's not unprecedented. He compares it to historical events like the Tylenol poisonings—a simple, devastating attack that could be easily replicated but rarely is. The sociological factors preventing widespread misuse remain a puzzle.
Most AI power users focus on creating agentic "skills" or "verbs" (e.g., summarize this). Steve Newman's personal toolkit highlights the power of building custom UIs or "nouns"—like a dashboard for agent status. This visual layer makes interacting with AI-processed information far more efficient and is an underexplored frontier.
To make his personal AI development manageable, Steve Newman structures his work as a suite of microservices. Each of his 15+ apps is its own project with a separate GitHub repo and database. This modular approach keeps the context window for the AI coding agent small and focused, which he believes is crucial for its effectiveness.
Decades of software development created established patterns and best practices. Steve Newman argues AI invalidates many of them. The most valuable engineers now are not those who know the old rulebook, but those who are comfortable with ambiguity, can think outside the box, and can invent new methods on the fly in a world without a map.
We often focus solely on model improvements. Steve Newman argues this is too narrow. True impact is a multiplicative function of eight factors: pre-training, post-training, inference compute, agent scaffolding, app design, user aptitude, workflow refactoring, and adoption. All are advancing simultaneously, creating a blistering pace of change.
Instead of trying to code on mobile, Steve Newman uses his time away from the desk for high-level thinking. He dictates unstructured thoughts about a project into his phone, then simply pastes the entire "brain dump" into an LLM. The AI's task is to organize the ramble into a structured, actionable prompt for his coding agent.
Contrary to the "token maxing" culture of keeping AI agents busy, Newman optimizes for his own time and mental state. After finding the pressure to constantly feed the AI stressful, he now gives it tasks when he's ready, not the other way around, prioritizing his own workflow over the agent's uptime.
Beyond just coding, improving AI models requires subtle skills like designing effective reinforcement learning environments or managing human expert feedback. Newman questions how close we are to recursive self-improvement by asking if AIs can automate these tasks, which rely on nuanced "taste and judgment" rather than just raw computational ability.
Getting real-time data from services like WhatsApp is notoriously difficult and risky. Steve Newman found a clever workaround: the WhatsApp Desktop app stores all messages in a local, unencrypted SQLite database. His system simply reads from this file, piggybacking on WhatsApp's own sync mechanism without violating terms of service or using fragile APIs.