Articles & Notes

Building the x70 LinkedIn desk made the remote API call look like the easy part; truthful research, human approval, scheduling, reconnection, and remote publication all need distinct states.

x70 Tools launched with 36 local browser utilities, but the part that keeps the collection maintainable is a single typed registry shared by navigation, metadata, routes, and discovery.

The x70 blog Admin became useful when writing, reviewing, scheduling, Git changes, and deployment evidence stopped being compressed into one optimistic action.

Dashboard polish was not the turning point for x70 Social. The product became coherent when guest access, realtime engagement, developer APIs, and AI moderation each received a precise boundary.

Applying one SEO checklist to every x70 property would have been wrong: public product pages should be discoverable, while authenticated operational surfaces should make their non-public status explicit.

Moving authentication and application state onto Convex reduced the number of persistence systems in the x70 App, but the real lesson was to migrate boundaries and call sites—not just schemas.

Apple's 2026 model APIs made local, private-cloud, and third-party inference choices part of one native architecture instead of separate product bets.

Centralizing x70’s identity assets removed a quiet source of drift across the ecosystem, while leaving each product free to keep its own layout, copy, and release cycle.

Cursor 3.6’s layered Auto-review design is a useful reminder that agent autonomy should depend on the action, the sandbox, and the available evidence—not one global yes-or-no setting.

Google's Gemini 3.5 launch paired a new model with managed execution, showing how much of an agent product now lives outside the model itself.

Bringing Codex to a phone matters less for writing code than for steering, reviewing, and approving asynchronous work at the right moment.

Cursor's SDK exposed its agent harness as a programmable service, making the infrastructure around the model the more interesting product boundary.

GPT-5.5 made long-running agent work the headline. The harder question is how to evaluate whether an agent can finish responsibly.

Rebuilding the portfolio’s Work pages taught me that more screenshots do not make a stronger case study; each image, video, and document has to prove a specific part of the story.

Moving the portfolio’s editorial content out of a hosted CMS made publishing less magical, but it also made every public claim reviewable, versioned, and easier to keep in sync with the product.