Daily AI Zine · No. 011

Today's Big Thing

The one thing that matters

Use it

Claude Opus 5 lands where coding work already happens

The important change is not only the model release, it is its arrival inside GitHub Copilot.

Claude Opus 5 is the day’s real development: Anthropic introduced it as a step change for the Opus tier, aimed at long-running agents, coding, and professional work. GitHub has now made it available in Copilot, where it is positioned for complex coding tasks that need careful reasoning, effective tool use, and reliability. For Adrian, the useful shift is operational. The best coding workflow is becoming a model-routing decision inside the repo, not a separate model shopping exercise. Use Opus 5 for gnarly refactors, cross-file reasoning, and agent tasks that may run for a while. Keep cheaper lanes for routine edits, formatting, and small bug fixes. Verdict: use, but only on work that justifies the heavier lane.

T1Anthropic Newsroom07/25 T1GitHub Changelog07/24

My AI Ecosystem

Your actual stack

Use it

Copilot is now a routing surface, not one assistant

With Opus 5 in Copilot and GitHub’s own framing that Copilot includes workflow, policy, and harness value around raw model access, the practical question is when to pay for the managed coding lane. Use Copilot for repo-aware work and direct APIs for controlled batch jobs or custom tools. Verdict: use.

T1GitHub Changelog07/24 T1GitHub Blog07/22

Test it

Hot signal: coding-agent tests are getting closer to real work

Tencent WorkBuddy Bench is a new coding-agent benchmark, meaning a repeatable test set for agents, built across Code, Web, Office, and Security tasks. The notable part is task construction from real commits, pull requests, or business scenarios instead of recycled public issue text. This is useful for designing private Claude Code and Codex tests. Verdict: test.

T1Hugging Face Daily Papers07/22

Watch it

Japan adds nine GENIAC data-ecosystem themes

METI selected nine R&D themes under the GENIAC project, aimed at strengthening Japan’s generative AI development capability and social implementation. The summary is thin, so the action is light: track which themes become usable industry programs, datasets, or partner opportunities. Verdict: watch.

T1METI Press (English)07/25

CoWork Corner

Claude CoWork, day to day

Quiet day in CoWork

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows. No meaningful CoWork, Claude Desktop, Projects, or MCP product change appears in today’s candidates. Revisit the documented Claude Desktop Extensions pattern instead: one-click MCP server installation remains the cleanest way to add repeatable tools to a persistent workspace. Pick one low-risk connector and write down what it can read, what it can write, and what it must never touch.

Tier 4 · quiet day, honest fallback

T1Anthropic Engineering Blog07/25

GPT Desk

OpenAI, ChatGPT, Codex

Test it

Cars24 is the OpenAI lead-recovery template to copy

OpenAI says Cars24 uses OpenAI-powered voice and chat agents for more than 1M monthly conversation minutes and 12% recovery of lost leads. Adrian should care because Goodsense, SET, Street Attack Japan, and Grey Group all have inquiry, sponsor, partner, and vendor conversations that can decay after first contact. Today’s action: export 20 old unresolved inquiries, classify the reason each went cold in ChatGPT, then draft one recovery script and one human handoff rule. Verdict: test.

T1OpenAI News07/16

Use it

Codex belongs in the creative tool drawer

OpenAI says its creative team uses Codex to build custom creative tools, speed up ideation, and prototype with context-aware AI. Adrian should care because Grey Group OS can treat Codex as a toolmaker for production prep, not only repo maintenance. Today’s action: ask Codex for one tiny internal tool that turns a treatment into a shot checklist, deck-outline, and follow-up email draft for Goodsense or SET. Verdict: use.

T1OpenAI News07/16

Small Money Systems

Small, repeatable, real

Test it

Sell a missed-inquiry rescue system

System: build a lightweight lost-lead rescue agent that sorts old inquiries, drafts polite reactivation messages, and flags replies that need a human. Customer: small production companies, local service businesses, and event operators. Offer: one import, one lead-status report, three recovery scripts, and a weekly follow-up queue. Price: ¥88,000 setup, then ¥33,000 per month. Existing assets: Grey Group OS, Goodsense sales material, SET operating patterns, and Street Attack Japan sponsor language. AI workflow: ChatGPT or the OpenAI API classifies leads, drafts replies, and creates handoff notes. First action: paste 20 stale inquiries into a private sheet and label the next best action. Repeatability: every client has old leads and new missed replies. Effort: one hour. Expected value: recovered conversations, new proposals, and less manual follow-up. Verdict: test.

T1OpenAI News07/16

Use it

Package a sponsor recap toolkit

System: create a Codex-built recap generator that turns event notes, image selects, and rough results into a sponsor-facing recap pack. Customer: brands, agencies, venues, and local sponsors around Street Attack Japan or similar activations. Offer: one recap deck outline, social captions, a sponsor thank-you email, and a next-activation proposal starter. Price: ¥77,000 per event. Existing assets: Street Attack Japan formats, Goodsense deck taste, Grey Group OS templates, and past production-planning language. AI workflow: Codex builds the internal generator, while ChatGPT refines the sponsor copy. First action: define the five input fields every recap must include. Repeatability: every activation creates the same reporting burden. Effort: half day. Expected value: faster sponsor follow-up and more repeat-activation conversations. Verdict: use.

T1OpenAI News07/16

Build Next

Deployable now

Use it

Add an Opus 5 lane to coding triage

What to build: a repo note or issue template that routes coding tasks into three lanes: routine, uncertain, and long-running, with Opus 5 reserved for the last two. Why now: Opus 5 is available in GitHub Copilot, and GitHub is explicit that Copilot includes workflow and policy value beyond raw API calls. Effort: one hour. Expected impact: fewer wasted premium-model runs and cleaner reviews. Dependencies: an active GitHub repo, Copilot access, and at least five real backlog issues. Verdict: use.

T1GitHub Changelog07/24 T1GitHub Blog07/22

Test it

Build a real-commit agent test pack

What to build: a small private benchmark, meaning a repeatable test set, from three old commits and two pull requests where the correct fix is already known. Why now: Tencent WorkBuddy Bench shows a stronger pattern for coding-agent evaluation by deriving tasks from real commits, pull requests, and business scenarios. Effort: half day. Expected impact: better Claude Code and Codex delegation, with fewer impressive but unsafe completions. Dependencies: a repo with history, known-good commits, and a simple scoring note. Verdict: test.

T1Hugging Face Daily Papers07/22

Try This Today

One action, right now

Pick one messy repo issue that touches more than one file. Run it through Copilot with Opus 5, then record three notes: what it changed, what you had to correct, and whether the task deserved the heavier model lane.