Key Takeaways
- A developer published 20 rules for working with AI coding agents, covering scoping, review, release, and incident follow-up.
- Google's Q3 figures show AI-driven search and Cloud growth, while one developer questions the claimed model lead.
- Levelsio shipped MCP access, invite-only hotel profiles, and a Quake 1 server milestone.
- Bazzly's MRR rose from $15K to $20K in 10 days, and an affiliate program launched.
- A Dell XPS 14 preloaded with Omarchy drew praise for its fast setup and AI-agent configuration prompts.
- Stack Overflow's monthly question count reportedly fell below 1,000.
1. AI Development Practice
- Tw93 compiled 20 rules for humans and agents to follow together: describe desired outcomes rather than implementations, supply context and a definition of “good,” and use the strongest model. New features, settings, or switches require confirmation first, and product rules should be recorded as they are made. The rules also define communication semantics: “look at/analyze” means read and report only, while “fix/change” authorizes edits; commits, releases, and deletions need the author's consent. AGENTS.md or CLAUDE.md should contain only what the model cannot infer, and pasted secrets are treated as leaked and rotated. Coding and debugging rules include having the agent explain changes and impact before writing code, maintaining a validation script, giving parallel agents separate worktrees, locating causes before fixing, adding logs after two failed fixes, and clearing context after three. Review and release rules separate writing from review, cap review at three rounds with only new content after the first, require personal verification for deletions and payment changes, and require each commit to state what changed, how it was verified, and how to roll back. Every production bug and user report becomes a pre-release test, and model switches trigger a rerun 1.
- Levelsio submitted his product to most MCP directories, which required adding OAuth login because directories do not accept a plain API key. Users can now get city rankings and itinerary recommendations inside Claude or ChatGPT, add, edit, and delete trips, and organize local meetups. He argued that all apps should expose functionally equivalent MCP/API access 1. Separately, Hotelist passed 700,000 monthly visits, becoming his most popular site, and launched invite-only user profiles with hotel ratings, voting on others' reviews, and karma; the idea, credited to Marc, is that reviews from friends and friends of friends are more trustworthy than random Google Maps reviews and naturally filter fakes 2. The site also works as an MCP connector for finding, booking, and reviewing hotels through an AI agent 3.
2. Platforms, Models, and Market Signals
- Jack Ellis questioned Google's claim that the Argon and Carbon models are “leading the frontier,” saying nearly all discussion he hears centers on Opus. The quoted post said Alphabet's Q3 annualized revenue reached $500 billion with 25% year-over-year growth, AI investment drove search query growth including commercial queries at the fastest pace in nearly five years, Google Cloud revenue grew 103% year over year, first-party model API volume approached 32 billion tokens per minute, up 40% quarter over quarter, and Genie 4 was called the most advanced new world model 1.
- Levelsio reported that Stack Overflow's question count fell below 1,000 last month, a new record 1. He also criticized X's zero-tolerance policy on AI replies, under which a single offense brings a permanent ban, calling the reply section worse than ever and AI replies an existential threat: if most replies are bots, users will leave 2.
- Levelsio unboxed a Dell XPS 14 preloaded with Omarchy sent by DHH, reporting first-boot setup in under a minute with no bugs. The onboarding asked about Git and AI agents including Codex, Claude Code, and OpenCode, with questions he found meaningful rather than the useless prompts he associates with Windows. He got stuck on shortcuts, was helped by jankeesvw, and planned to configure SSH to a VPS and customize Omarchy with Claude Code; he later found Action + K brings up all shortcuts and showed the screensaver 1 2 3.
3. Product and Revenue Experiments
- Tibo said Bazzly's MRR rose from $15,000 to $20,000 in 10 days, that he receives near-daily messages calling it the best Reddit platform, and that the partnership with Filip Panoski could produce another $100,000-level product. He opened an affiliate program offering reposts for posts or DMs and double support for articles 1; a repost framed the program around an $800 LTV and 50% trial-to-paid conversion 2. Promotional material said Bazzly finds Reddit “what tool does X” threads that rank on Google and get cited by ChatGPT, drafts replies mentioning the product, and posts them automatically 3. A longer repost described the channel's value, said Bazzly scans relevant threads, drafts replies and personalized DMs, and surfaces old ranked threads, with full automation or a Chrome extension for posting from one's own account at $199 per month including 170 credits refreshed monthly. It also cautioned that showcased testimonials are company-selected claims rather than guarantees, that every draft should be human-reviewed and authorship disclosed, that Reddit bans vote manipulation and undisclosed promotion, and that features such as “smart upvotes” and aged accounts require reading subreddit and Reddit rules first 4.
- Jon Yongfook contrasted one SaaS business at $300,000 monthly recurring revenue with 3,000 SaaS businesses at $100 MRR each, arguing that AI finally makes the latter model viable and calling it a moat 1.
- Nathan Barry relayed Every CEO Dan Shipper's judgment that newsletters are dead because AI agents will read email before subscribers do. Barry, whose 13-year career has centered on email, said he dislikes the conclusion but thinks Dan may be right, that there is still time to adapt, and that the show covers ways to keep email reaching audiences 1.
- Levelsio said his FTE-based browser Quake 1 became one of the most popular servers in the QuakeWorld community. dsvensson contacted him to add WebRTC and QTV spectator support, saying the server is full daily and attracts many newcomers whom the community hopes to bring into competitive play; Levelsio proposed organizing matches 1.
