Daily AI Zine
Tuesday, September 22, 2026
Issue No. 067 · Tokyo · full edition

✦ Today's Big Thing

Copilot adds another serious coding model

Grok 4.7 is rolling into GitHub Copilot, while OpenAI case studies point toward faster video-ad production and source-linked company memory.

5 min read · 7 sections

In Brief
  1. Grok 4.7 is now available in GitHub Copilot for agentic coding and complex, multistep work.
  2. OpenAI says Higgsfield used GPT-6 Astra to ship new video-ad features in a day.
  3. GitHub’s usage metrics API now covers skills, custom agents, MCP servers, slash commands, and plugins.
  4. Codex CLI worktrees are worth testing as a cleaner way to run parallel tasks without mixing branches.

Today's Big Thing

The one thing that matters

Test it

Grok 4.7 enters GitHub Copilot

The practical story is model choice inside the coding workspace, not another standalone chatbot.

GitHub says Grok 4.7, xAI’s latest reasoning model, is rolling out in GitHub Copilot. The model is described as designed for agentic coding and complex, multistep workflows. For daily shipping, the useful shift is that Copilot keeps turning into a live model bench inside the editor and agent workflow: one interface, more competing model behavior. Adrian should not move core work to it on trust. Use one non-sensitive repo task, compare it against the current default model, and judge on patch quality, instruction following, and review load. Verdict: test.

T1GitHub Changelog09/21

My AI Ecosystem

Your actual stack

Test it

Hot signal: video-ad tools are shipping faster

OpenAI says Higgsfield AI used GPT-6 Astra to ship new video features in a day, aimed at making video ad creation easier for small businesses. Treat this as a practical signal, not a benchmark claim: the useful point is that creative-product teams are compressing feature cycles around video ads. Test only if it produces sellable variants faster than the current edit path.

T1OpenAI News09/21
Use it

Copilot admins can see custom agent usage

GitHub expanded Copilot’s usage metrics API to include agentic CLI activity for skills, custom agents, MCP servers, slash commands, and plugins. That matters because the invisible layer of agent customization is becoming measurable. For a small technical operation, this is useful only after there are named agents or MCP servers worth comparing.

T1GitHub Changelog09/17
Use it

MCP still earns its place through control

A fresh MCP debate argues that full terminal agents can call APIs directly, but MCP still matters when you want defined service access and safer authentication. In plain terms: MCP is less about raw capability and more about deciding exactly which doors an agent can open. Keep it for bounded tools, not for every simple script.

T2Simon Willison Weblog09/20T1Anthropic Engineering Blog

CoWork Corner

Claude CoWork, day to day

CoWork has no confirmed product move today

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows. No meaningful confirmed CoWork product change surfaced today. One user report claims GitHub access became unreliable in CoWork after session handling changed, but treat that as a signal, not a verified platform update. The useful workflow is simple: keep repo writes outside the workspace when reliability matters. Use CoWork for planning, context, and review notes, then hand implementation to a local coding agent or normal GitHub workflow with explicit review.

Tier 4 · quiet day, honest fallback

T2r/ClaudeAI09/21T1Anthropic Engineering Blog

GPT Desk

OpenAI, ChatGPT, Codex

Test it

Codex worktrees make parallel tasks cleaner

What changed: Codex CLI 0.154.0 adds experimental support for starting or forking tasks in separate Git worktrees, with browsing, resume, and inline questions while Codex keeps working. Why Adrian cares: Grey Group OS and Daily AI Zine work both create small parallel fixes that should not collide in one branch. Do this: update Codex CLI and run two low-risk tasks in separate worktrees, then compare review friction before making it standard.

T2Perplexity Research09/21T1OpenAI News09/10
Use it

Turn scattered files into operational memory

What changed: OpenAI says V7 uses GPT-5.6 to turn scattered company files into context agents can use for complex, source-linked work. Why Adrian cares: Grey Group OS needs reliable project memory across proposals, production notes, decks, and follow-ups. Do this: pick one active Goodsense or SET project folder, load only final or current documents, and ask for a source-linked operating brief with open questions and next actions.

T1OpenAI News09/21

Small Money Systems

Small, repeatable, real

Test it

Small system: video-ad variant desk

System: build a rapid video-ad variant desk that turns one client idea, event recap, or product note into a short creative brief plus three ad concepts. Customer: local venues, niche retailers, event sponsors, and Street Attack Japan partners. Offer: a one-page concept pack, three short scripts, and edit direction for vertical ads. Price: ¥55,000 per pack. Existing assets: production judgment, Japan-English positioning, Street Attack Japan references, and proposal templates. AI workflow: ChatGPT or an OpenAI API flow drafts variants, Claude Code formats the pack, and human review selects the usable set. First action: make one sample from an existing non-confidential event idea. Repeatability: each pack reuses the same intake, script, and review structure. Effort: one hour. Expected value: small repeatable revenue and faster sponsor conversations.

T1OpenAI News09/21
Use it

Small system: source-linked memory pack

System: build a source-linked company memory pack for small teams with messy folders. Customer: founders, producers, and small agencies that lose time searching past decisions. Offer: one cleaned project brief, decision log, open-issue list, and reusable agent context file. Price: ¥88,000 for the first pack. Existing assets: Grey Group OS structure, proposal workflows, production planning habits, and Adrian’s document-cleanup patterns. AI workflow: ChatGPT extracts the first pass, Claude checks structure, and Codex or Claude Code turns the output into a reusable template. First action: create a before-and-after demo using one internal non-sensitive folder. Repeatability: each new client uses the same intake checklist and memory-pack format. Effort: half day. Expected value: paid setup work that can become a monthly update retainer.

T1OpenAI News09/21

Build Next

Deployable now

Test it

Build a Codex worktree intake lane

What to build: a tiny GitHub issue label and local script convention that sends each approved Codex task into its own Git worktree, then records branch, task, status, and review note in one markdown table. Why now: Codex CLI now has experimental separate worktree support for starting or forking tasks. Effort: one hour. Expected impact: fewer branch collisions and cleaner review when two small fixes run at once. Dependencies: updated Codex CLI, a Git repo, and one low-risk backlog item.

T2Perplexity Research09/21T1OpenAI News09/10

Try This Today

One action, right now

Pick one small GitHub Copilot task with clear acceptance criteria: a failing test, a copy tweak, or a narrow refactor. Run Grok 4.7 once, save the diff, then compare it against your normal model on review time and number of corrections.