Daily AI Zine · No. 074

Today's Big Thing

The one thing that matters

Use it

GitHub puts Claude Sonnet 5.5 into the daily coding lane

The important part is not another model name. It is where the model now lives.

Claude Sonnet 5.5, Anthropic’s newest Sonnet model, is now generally available in GitHub Copilot. GitHub frames it for well-scoped everyday work such as building features and fixing bugs, which is exactly the layer where coding agents either become useful or create review debt. The practical shift: model choice is moving into the normal GitHub workbench rather than staying in separate chat tools. For teams already shipping through repositories, this makes it easier to test one bounded issue, compare output against the current default model, and decide by merge quality rather than benchmark claims. Verdict: use it for contained implementation tasks, not broad architecture handoff.

T1GitHub Changelog09/28

My AI Ecosystem

Your actual stack

Test it

Hot signal: Cloudflare’s agent browser adds WebMCP

Cloudflare updated Kitesurf, its Workers-based browser for AI agents, with WebMCP support, better DOM performance, and terminal-based rendering. Cloudflare says more than 730,000 Web Platform subtests now pass, which matters because agents need ordinary web pages to behave predictably before they can do useful browser work.

T1Cloudflare Blog09/28

Watch it

Faster JavaScript tooling now matters for agents too

Cloudflare says VoidZero has delivered more than 80 releases since joining, including faster JavaScript compilation, linting, and testing, a 10x faster React compiler, and Vite+ 1.0. The operator point is simple: agents iterate through build and test loops. Faster loops mean cheaper review cycles and less waiting.

T1Cloudflare Blog09/28

CoWork Corner

Claude CoWork, day to day

CoWork: no product move, review one execution boundary

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows. No meaningful CoWork product change surfaced today. The useful revisit is a connector drill: pick one MCP-linked tool, confirm what data it can read, what it can write, and where execution happens. MCP is the protocol that lets an AI workspace call external tools through controlled connectors; code execution through MCP is a reminder that the boundary is operational, not just descriptive.

Tier 4 · quiet day, honest fallback

T1Anthropic Engineering Blog

GPT Desk

OpenAI, ChatGPT, Codex

Test it

OpenAI’s ad tools deserve a private pitch rehearsal

What changed: OpenAI describes AI-powered advertising experiences, including Sponsored Agents, marketer tools, and HubSpot and Shopify integrations. Why Adrian should care: Grey Group can turn this into concrete proposal language for Goodsense, SET, or Street Attack Japan without committing to an ad buy. Do this: make one private deck page showing a Sponsored Agent concept, the user journey, and the data it would need before any client-facing pitch.

T1OpenAI News09/16

Test it

Use OpenAI call agents as an inquiry filter test

What changed: Ringg uses GPT-5.6 for multilingual agents across voice, chat, WhatsApp, and web, resolving up to 65% of customer calls at 90% lower cost versus GPT-4.1. Why Adrian should care: Goodsense and SET both benefit from faster qualification of inbound requests, especially bilingual ones. Do this: collect ten real inquiry types and draft the first-pass triage script before touching any voice deployment.

T1OpenAI News09/23

Small Money Systems

Small, repeatable, real

Test it

Small system: storefront AI ad checklist

System: build a paid readiness checklist for brands that want an AI ad or Sponsored Agent but have not organized product data, FAQs, lead capture, and approval rules. Customer: small Japan-facing brands, restaurants, studios, and pop-ups. Offer: a short audit, one sample agent conversation, and a one-page risk list. Price: ¥35,000. Existing assets: Grey Group proposal structure, Goodsense strategy language, and Street Attack Japan activation experience. AI workflow: ChatGPT drafts the agent journey, then Claude Code or Codex turns the checklist into a reusable form. First action: make the ten-question intake today. Repeatability: each audit reuses the same scoring sheet. Effort: one hour. Expected value: paid discovery and warmer leads for larger campaign work.

T1OpenAI News09/16

Test it

Small system: missed-inquiry recovery pack

System: create a lightweight service that turns missed calls, web inquiries, DMs, and WhatsApp messages into a prioritized follow-up list with suggested replies. Customer: small service businesses, event teams, and production vendors that lose leads during busy periods. Offer: one import, deduping, bilingual reply drafts, and a next-action sheet. Price: ¥50,000. Existing assets: Goodsense client-communication habits, SET production intake patterns, and Grey Group OS operating templates. AI workflow: OpenAI models classify intent and draft replies, then Claude Code packages the results into a simple report. First action: mock the report from five sample inquiries. Repeatability: the same template can run monthly. Effort: half day. Expected value: recovered leads and less manual sorting.

T1OpenAI News09/23

Build Next

Deployable now

Test it

Build an agent-browser check for quote forms

What to build: a small test that sends an agent browser through one public inquiry or quote form, records confusing fields, and stops before submission. Why now: Cloudflare’s Kitesurf update adds WebMCP support, better DOM performance, and terminal-based rendering for agents. Effort: half day. Expected impact: fewer broken intake flows and better lead capture. Dependencies: access to an existing public form, a safe test account, and a clear stop rule before any real submission.

T1Cloudflare Blog09/28

Try This Today

One action, right now

Pick one low-risk GitHub issue: a bug fix, copy update, or small component change. Run it with Claude Sonnet 5.5 in Copilot, then record three things only: time to first patch, review fixes needed, and whether you would assign that class of task again.