Vibe-Coding an Attention Firewall, w/ Steve Newman, creator of The Curve
Episode
129 min
Read time
3 min
Topics
Software Development
AI-Generated Summary
Key Takeaways
- ✓Attention Firewall Architecture: Newman built a system pulling email, Slack, WhatsApp, Signal, and Twitter into a single classifier that applies a one-page rubric to flag only urgent messages. These surface on a dedicated second monitor alongside a three-hour calendar view. The result: he eliminated approximately 30 daily check-ins across five apps, recovering substantial focus time without missing time-sensitive communications. The rubric grew exception-by-exception over several months of iteration.
- ✓Universal Logging as Debugging Infrastructure: Every component of Newman's stack — backend services, JavaScript frontends, Android apps — logs to a single SQLite database hosted on Cloudflare. When something breaks, he runs a one-sentence "systematic debugging" prompt and Claude traces the failure through actual log evidence rather than guessing. This single infrastructure decision, combined with the Superpowers Claude plugin's systematic debugging skill, resolves nearly 100% of issues without manual investigation.
- ✓Anti-Token-Maxing Philosophy: Newman initially felt pressure to keep agents constantly fed with prompts to avoid idle compute time. He reversed this after breaking work into roughly 15 isolated microservice projects. Now he runs zero to five parallel Claude Code agents and prompts them when convenient, not when the agent is waiting. His framing: optimize the human's time, not the agent's. This shift reduced stress significantly while maintaining output.
- ✓WhatsApp Integration via Local SQLite: To achieve real-time WhatsApp message ingestion without risking account bans from unofficial API hacks, Newman reads directly from the unencrypted SQLite database that WhatsApp Desktop writes locally on macOS. A cron job polling approximately once per minute reads this file with no network calls to WhatsApp servers. This approach is undetectable to Meta and avoids the account-suspension risks associated with third-party WhatsApp API services.
- ✓RSS Summarization with Two-Level Summaries: Newman receives roughly 50 newsletters, blog posts, and podcast transcripts daily. His custom RSS reader pre-computes two summaries per item: a short version for triage decisions and a longer one-page version surfacing novel ideas and notable evidence. The prompt instructs Claude to flag ideas novel relative to its training data. He skims the short summary in about ten seconds per item, then decides whether to read, deep-dive, or archive.
What It Covers
Steve Newman, creator of what became Google Docs and founder of the Golden Gate Institute for AI, walks through 15 bespoke personal productivity applications he built using Claude Code. The conversation covers his attention firewall system, RSS summarization tool, agent status dashboard, Chrome extensions, and unified logging infrastructure, plus broader reflections on AI's trajectory and software engineering's future.
Key Questions Answered
- •Attention Firewall Architecture: Newman built a system pulling email, Slack, WhatsApp, Signal, and Twitter into a single classifier that applies a one-page rubric to flag only urgent messages. These surface on a dedicated second monitor alongside a three-hour calendar view. The result: he eliminated approximately 30 daily check-ins across five apps, recovering substantial focus time without missing time-sensitive communications. The rubric grew exception-by-exception over several months of iteration.
- •Universal Logging as Debugging Infrastructure: Every component of Newman's stack — backend services, JavaScript frontends, Android apps — logs to a single SQLite database hosted on Cloudflare. When something breaks, he runs a one-sentence "systematic debugging" prompt and Claude traces the failure through actual log evidence rather than guessing. This single infrastructure decision, combined with the Superpowers Claude plugin's systematic debugging skill, resolves nearly 100% of issues without manual investigation.
- •Anti-Token-Maxing Philosophy: Newman initially felt pressure to keep agents constantly fed with prompts to avoid idle compute time. He reversed this after breaking work into roughly 15 isolated microservice projects. Now he runs zero to five parallel Claude Code agents and prompts them when convenient, not when the agent is waiting. His framing: optimize the human's time, not the agent's. This shift reduced stress significantly while maintaining output.
- •WhatsApp Integration via Local SQLite: To achieve real-time WhatsApp message ingestion without risking account bans from unofficial API hacks, Newman reads directly from the unencrypted SQLite database that WhatsApp Desktop writes locally on macOS. A cron job polling approximately once per minute reads this file with no network calls to WhatsApp servers. This approach is undetectable to Meta and avoids the account-suspension risks associated with third-party WhatsApp API services.
- •RSS Summarization with Two-Level Summaries: Newman receives roughly 50 newsletters, blog posts, and podcast transcripts daily. His custom RSS reader pre-computes two summaries per item: a short version for triage decisions and a longer one-page version surfacing novel ideas and notable evidence. The prompt instructs Claude to flag ideas novel relative to its training data. He skims the short summary in about ten seconds per item, then decides whether to read, deep-dive, or archive.
- •Microservice Project Structure for Context Management: Newman deliberately splits his 15 projects into separate GitHub repositories, each with its own Cloudflare database and deployment. All repositories sit adjacent in one Docker container directory, allowing agents to read across project boundaries when needed. Claude Code runs in the container with dangerously-skip-permissions mode enabled, but all source data remains in its original location — Gmail, WhatsApp, Slack — so a catastrophic agent error loses only the custom interface, not the underlying data.
- •Voice-to-Prompt Brain Dump Workflow: For capturing ideas during walks, Newman dictates into the Gmail compose window using the iOS keyboard's built-in dictation button, addressed to himself. The raw brain dump then gets pasted into an LLM with a single instruction: organize this into a Claude Code prompt. He typically skips reviewing the cleaned output and lets the agent proceed, correcting misunderstandings after the fact rather than proofreading the prompt first — a deliberate shift away from his prior measure-twice engineering habits.
Notable Moment
Newman, a software engineer since 1985 who built Google Docs, admitted he went four decades without a second monitor — then purchased one specifically to display his attention firewall dashboard. The detail underscores how fundamentally these self-built tools are restructuring decades-old working habits, not merely adding convenience on top of existing workflows.
You just read a 3-minute summary of a 126-minute episode.
Get Cognitive Revolution summarized like this every Monday — plus up to 2 more podcasts, free.
Pick Your Podcasts — FreeKeep Reading
More from Cognitive Revolution
Nested Learning: Ali Behrouz on the Quest for Continual Learning & Illusion of AI Architectures
Jun 3 · 180 min
The Journal
How AI Is Being Trained to Do Your Job
Jun 4
More from Cognitive Revolution
Inside Nathan's Second Brain: Daniel Miessler, Security Expert & Creator of PAI, Audits My AI Setup
May 30 · 152 min
The Vergecast
Microsoft's plan to catch up in AI
Jun 4
More from Cognitive Revolution
We summarize every new episode. Want them in your inbox?
Nested Learning: Ali Behrouz on the Quest for Continual Learning & Illusion of AI Architectures
Inside Nathan's Second Brain: Daniel Miessler, Security Expert & Creator of PAI, Audits My AI Setup
Your Biggest Lever: Designing your AI Career for Maximum Impact, with 80,000 Hours founder Ben Todd
All Compute Is Food: Palisade's Jeffrey Ladish on AI Shutdown Resistance, Self-Replication & Ecology
The Model Eats the Scaffolding: DeepMind's Logan Kilpatrick & Tulsee Doshi on 3.5 Flash, Omni & More
Similar Episodes
Related episodes from other podcasts
The Journal
Jun 4
How AI Is Being Trained to Do Your Job
The Vergecast
Jun 4
Microsoft's plan to catch up in AI
The AI Breakdown
Jun 4
How Companies Are Becoming AI Token Efficient
The Bulwark Podcast
Jun 4
Jonathan V. Last: We Got a Billionaire Problem
The Startup Ideas Podcast
Jun 4
Codex Sites Clearly Explained (and how to use it)
Explore Related Topics
This podcast is featured in Best AI Podcasts (2026) — ranked and reviewed with AI summaries.
Read this week's Software Engineering Podcast Insights — cross-podcast analysis updated weekly.
You're clearly into Cognitive Revolution.
Every Monday, we deliver AI summaries of the latest episodes from Cognitive Revolution and 192+ other podcasts. Free for up to 3 shows.
Start My Monday DigestNo credit card · Unsubscribe anytime