Daily AI Zine · No. 059

Today's Big Thing

The one thing that matters

Test it

Hot signal: Perplexity lets Astra run closer to operations

Astra is being used beyond drafting and research, which makes it a real workflow-routing story.

OpenAI says Perplexity uses GPT-6 Astra to write communications, change software, and monitor production systems, with much less frequent check-in than earlier models. That is the important shift today: the model is being positioned as a work loop participant, not just a better answer box. The useful read is not “remove humans.” It is “move reviews to the right boundary.” For Adrian, the test is a low-risk operations lane where the model drafts the message, proposes the code change, and summarizes the production signal, while a human still approves deployment and external communication. Verdict: test one bounded lane, not broad autonomy.

T1OpenAI News09/14

My AI Ecosystem

Your actual stack

Test it

Japan administrative knowledge becomes an AI product surface

Polimill is using OpenAI GPT models and Codex to help municipalities search and use administrative knowledge while accelerating development. The practical pattern is clear: domain documents plus search plus workflow support. For Japan-facing teams, this is a prompt to package narrow institutional knowledge before trying to sell a broader AI transformation.

T1OpenAI News08/31

Test it

HydraFusion makes model routing a coding-cost test

GitHub says Project HydraFusion is available as a Copilot research preview. In controlled offline evaluations, its selective coding workflows matched or exceeded the evaluated Opus 5 baseline while reducing estimated workflow cost. Treat this as a test candidate for repetitive coding tasks, not a replacement for review.

T1GitHub Blog09/04

Watch it

Security remediation is moving toward traffic-aware triage

Cloudflare says its Managed Defense work with OpenAI Daybreak models uses production traffic and security signals to prioritize findings, prepare edge mitigations when safe, and propose code patches. The useful part is prioritization: fix what live traffic proves matters first, then let code agents draft patches under review.

T1Cloudflare Blog09/03

CoWork Corner

Claude CoWork, day to day

CoWork: keep repeatable skills outside project memory

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows. No meaningful CoWork product change was found in today’s candidates. The workflow worth revisiting is separation: keep stable procedures as reusable skills, and keep project memory for current facts, decisions, and handoffs. Anthropic’s Agent Skills direction supports that split, because repeatable abilities should travel across work without polluting each project record.

Tier 4 · quiet day, honest fallback

T1Anthropic Engineering Blog

GPT Desk

OpenAI, ChatGPT, Codex

Test it

Give Astra one internal work loop with a hard review gate

What changed: OpenAI says Perplexity uses Astra to write communications, change software, and monitor production systems with fewer check-ins. Why Adrian should care: Grey Group OS can use that pattern for Daily Ops and proposal operations without pretending the model owns the business outcome. Action: pick one internal loop this week where ChatGPT drafts the update, Codex proposes the repo change, and Adrian approves every deploy or client-facing message.

T1OpenAI News09/14

Test it

Use Polimill as the Japan knowledge-search demo pattern

What changed: Polimill uses OpenAI GPT models and Codex to help municipalities search and use administrative knowledge while accelerating development. Why Adrian should care: Goodsense and SET can borrow the structure for Japan-English knowledge work, especially decks, local research, and public-sector-adjacent briefs. Action: create one small Codex-built demo that ingests ten approved documents and returns sourced answers plus a “next action” field.

T1OpenAI News08/31

Test it

Turn security findings into a reviewed patch queue

What changed: Cloudflare says its OpenAI Daybreak model workflow can prioritize findings using production traffic and security signals, prepare edge mitigations when safe, and propose code patches. Why Adrian should care: Grey Group OS, Netlify sites, and API-connected tools need a lighter way to separate real risk from noise. Action: make a one-page security backlog template with fields for traffic evidence, proposed mitigation, code patch, and human approval.

T1Cloudflare Blog09/03

Small Money Systems

Small, repeatable, real

Use it

Small system: Japan admin knowledge starter pack

System: build a tiny searchable knowledge demo from a client’s public documents and internal FAQs. Customer: Japan-facing agencies, municipal vendors, or local business teams that need faster answers from scattered rules. Offer: a working private demo, ten source documents, bilingual answer examples, and a short rollout memo. Price: ¥35,000 for the starter pack. Existing assets: Goodsense, SET, Grey Group OS, Japan-English proposal experience, and Codex shipping habits. AI workflow: Codex builds the document interface, ChatGPT drafts answer formats, and Claude Code can clean the repo. First action: choose one public document set and create the intake checklist. Repeatability: each pack reuses the same loader, answer template, and sales deck. Effort: half day. Expected value: small paid diagnostics that can become retainers or implementation work.

T1OpenAI News08/31

Test it

Small system: AI video fidelity test pack

System: sell a short AI video test package that checks whether generated motion, prompt adherence, and visual fidelity are good enough for a specific brand or event. Customer: small brands, agencies, venues, and Street Attack Japan partners considering AI-assisted recap or concept clips. Offer: three prompt tests, one comparison board, and a recommendation on whether to use AI video, shoot normally, or combine both. Price: ¥25,000. Existing assets: production taste, shot planning, Street Attack Japan visual language, and Grey Group proposal templates. AI workflow: ChatGPT writes prompt variants, Claude reviews shot logic, and the chosen video tool generates test outputs. First action: write one three-shot test brief. Repeatability: reuse the same scoring sheet for each client. Effort: one hour. Expected value: paid pre-production revenue and better-qualified video leads.

T1Runway Research

Build Next

Deployable now

Test it

Build an agent operations review ledger

What to build: a simple internal page that records each agent-run operation: communication drafted, software change proposed, production signal checked, reviewer, and final decision. Why now: OpenAI says Perplexity is using Astra across communications, software changes, and production monitoring, so review boundaries matter more than prompt cleverness. Effort: half day. Expected impact: faster delegation with less ambiguity and a clearer audit trail. Dependencies: an OpenAI account, one active repo or operations workflow, and a reviewer who can approve or reject changes.

T1OpenAI News09/14

Test it

Build a coding model-routing scorecard

What to build: a small GitHub-linked scorecard that logs task type, model or agent used, result quality, reviewer fixes, and estimated cost. Why now: GitHub’s HydraFusion preview claims selective coding workflows can match or exceed an Opus 5 baseline while reducing estimated workflow cost in offline evaluations. Effort: one hour. Expected impact: better routing for coding work and fewer expensive default choices. Dependencies: GitHub access, a few recent coding tasks, and one consistent review rubric.

T1GitHub Blog09/04

Try This Today

One action, right now

Pick one low-risk workflow. Let ChatGPT draft the communication, let Codex propose the repo change, and require a human approval line before anything ships or leaves the building. Log the result in a five-column review ledger.