Daily AI Zine · No. 037

Today's Big Thing

The one thing that matters

Test it

Hot signal: GitHub turns team chat into an agent room

Copilot is moving from solo assistant to shared work surface.

GitHub now lets a Microsoft Teams discussion become a collaborative Copilot agent session that people can see and help direct. GitHub also put Copilot agent capabilities into Slack in public preview. The practical change is not another model option, it is where agent work happens. A task can start inside the same channel where the brief, objections, and approvals already live. For Adrian, the useful pattern is shared supervision: one low-risk repo task, one visible thread, one human approval point before merge. Test this now if the team chat stack supports it. If not, copy the pattern into GitHub issues until the tooling matches the workflow.

T1GitHub Changelog08/21 T1GitHub Changelog08/21

My AI Ecosystem

Your actual stack

Test it

Copilot enters Slack in public preview

GitHub’s Slack integration now brings the agentic capabilities of GitHub Copilot CLI and the GitHub Copilot app into Slack in public preview. The useful bit is operational, not social: agent work can be requested where decisions already happen. Test it only on low-risk repo maintenance until permissions, logging, and approval behavior are clear.

T1GitHub Changelog08/21

Test it

Claude Opus 5 is the endurance lane

Anthropic positions Claude Opus 5 as an Opus-tier model for long-running agents, coding, and professional work. Treat that as a routing clue rather than a blanket upgrade. Use it where persistence and judgment matter, then keep cheaper or faster lanes for drafts, extraction, and routine code edits.

T1Anthropic Newsroom

Use it

GPT-5.6 guidance pushes smarter agent routing

OpenAI’s builder guide says startups are using GPT-5.6 to build faster, more cost-efficient agents with smarter model selection and new Responses API capabilities. The plain-English takeaway: stop treating one model as the whole system. Route planning, execution, review, and summarization as separate jobs.

T1OpenAI News08/13

CoWork Corner

Claude CoWork, day to day

CoWork gets a learning layer, not a product shift

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows. Today’s useful CoWork signal is training, not a new workflow surface: Anthropic is reported to have launched Claude Academy, a free learning site that explains Claude products including Code and CoWork. No meaningful CoWork product change showed up in the candidates. Use this as a maintenance day: pick one current project room and rewrite its standing instructions into three parts, purpose, allowed tools, and handoff rule. That keeps the workspace legible before adding more agent behavior.

T1Anthropic Newsroom T2ITmedia AI+08/21

GPT Desk

OpenAI, ChatGPT, Codex

Test it

ChatGPT on Mac is becoming a follow-up lane

What changed: ITmedia reports that ChatGPT for macOS now has an Apple Messages integration for searching past conversations, drafting replies, and sending with per-message approval. Why Adrian should care: Goodsense, SET, and Street Attack Japan all depend on fast, careful follow-up after calls, shoots, and informal Japan-business chats. What to do: test one approved thread today by asking ChatGPT to find the last concrete ask, draft a short reply, and stop before sending.

T1OpenAI News08/12 T2ITmedia AI+08/21

Use it

Make Codex write the acceptance check before the code

What changed: OpenAI’s enterprise research says companies are moving from AI assistance to execution with ChatGPT and Codex. Why Adrian should care: Grey Group OS needs repeatable delivery rules, not only faster commits. What to do: before the next Codex task, ask ChatGPT to produce a five-line acceptance checklist, then give Codex only the checklist and the repo context. If Codex cannot satisfy the checklist, the task was not defined enough.

T1OpenAI News08/12

Use it

GPT-5.6 needs a budget lane, not a prestige lane

What changed: OpenAI’s GPT-5.6 builder guide emphasizes cost-efficient agents, smarter model selection, and Responses API capabilities. Why Adrian should care: Grey Group, Goodsense, and the Daily AI Zine all have mixed work, from cheap extraction to higher-stakes reasoning. What to do: create one routing note with three lanes, cheap summary, normal build, premium review, then force every new OpenAI API task into one lane before it runs.

T1OpenAI News08/13

Small Money Systems

Small, repeatable, real

Test it

Small system: message-to-proposal follow-up pack

System: build a one-hour follow-up service that searches approved message history, identifies the last real buying signal, and drafts a concise next-step proposal. Customer: foreign producers, agencies, or visiting founders who had a Japan conversation but stalled. Offer: three polished follow-ups, one short proposal angle, and one suggested call agenda. Price: ¥30,000. Existing assets: Goodsense, SET, Street Attack Japan relationships, Japan-English business judgment, and Adrian’s production context. AI workflow: ChatGPT for message search and drafts, Codex for a reusable checklist or template if needed. First action: pick one dormant lead and generate the first approved draft today. Repeatability: the same template works for every stale thread. Effort: one hour. Expected value: lead revival and faster proposal conversion without starting a new sales cycle.

T1OpenAI News08/12 T2ITmedia AI+08/21

Use it

Small system: agent-room audit for small teams

System: sell a compact audit that turns one messy GitHub or chat-based workflow into a visible agent workroom with task owner, prompt, repo link, approval point, and merge rule. Customer: small companies already using GitHub but unsure how to control AI coding work. Offer: one reviewed repo flow, one agent-session template, and one approval checklist. Price: ¥50,000. Existing assets: Grey Group OS, Claude Code, Codex, GitHub habits, and Adrian’s delivery discipline. AI workflow: Claude Code or Codex drafts the template, then a human review locks the rules. First action: make the audit checklist from today’s GitHub Copilot chat pattern. Repeatability: each audit reuses the same checklist with client-specific edits. Effort: half day. Expected value: paid consulting revenue and fewer uncontrolled agent commits.

T1GitHub Changelog08/21 T1GitHub Changelog08/21

Build Next

Deployable now

Use it

Build an agent-session brief generator

What to build: a small GitHub issue template or script that turns a repo task into a shared agent-session brief with goal, files to inspect, forbidden actions, approval point, and done criteria. Why now: GitHub is moving Copilot agent work into Teams and Slack, so the brief needs to be readable in chat as well as in the repo. Effort: one hour. Expected impact: fewer vague agent runs and cleaner human review. Dependencies: an active GitHub repo and one low-risk maintenance task.

T1GitHub Changelog08/21 T1GitHub Changelog08/21

Try This Today

One action, right now

Pick one low-risk GitHub issue. Add five lines before any agent touches it: goal, context, files, forbidden actions, and approval point. Then run the task only if the brief is clear enough for another person to supervise.