Happy Thursday,

Apologies for the image layout issue last week. Just a bit of technical difficulty while I transition to the new format. Hope you're having a great week so far!

OpenAI gave Codex full control of your Mac this week, and Canva declared itself an AI company with design tools bolted on. Both shipped April 16. The agent era stopped being theoretical. Let's get into it.

---

📰 THIS WEEK'S RABBIT HOLE

OpenAI's Codex Can Now Drive Your Mac

On April 16, OpenAI shipped a Codex update that lets the agent see your screen, click, and type across any macOS app. It works in parallel. A sandboxed cursor operates the virtual workspace while your real cursor keeps doing whatever you were doing.

Same release also added memory (Codex remembers your preferences and corrections across sessions), an in-app browser you can comment on directly, image generation via gpt-image-1.5, 90+ plugins wrapping MCP servers and app integrations, and scheduled automations that let an agent wake itself up days later to keep working on a task.

macOS-only at launch. Not available yet in the EU or UK. Personalization features are gated from Enterprise, Education, EU, and UK accounts until rollout completes.

The Anthropic computer-use preview from last year was the proof of concept. This is the production version, and it ships with a plugin store.

See What Shipped →

---

🛠 TOOL OF THE WEEK

CraftBot: A Self-Hosted AI Assistant That Never Leaves Your Machine

CraftBot hit #3 Product of the Day on Product Hunt on April 18, with 263 upvotes. It's a self-hosted proactive AI assistant that runs locally on your own machine, interprets tasks, plans actions, and executes them 24/7.

The pitch is the opposite of Codex. Codex sees your screen because it's running in OpenAI's cloud with a streamed view. CraftBot runs the model, the planner, and the executor entirely on your hardware. Nothing gets shipped out.

If you've been holding back on agentic AI because pointing OpenAI's cursor at your documents feels like a terrible idea, this is the alternative worth trying. Tradeoff is the obvious one: you're renting your own GPU instead of someone else's.

See The Launch →

---

🧠 DEEP DIVE

Canva AI 2.0 Turns a Zoom Recording Into a Deck

On April 16 at Canva Create in Los Angeles, Canva launched AI 2.0 and repositioned itself: "an AI platform with design tools" instead of the other way around. The claim is that Canva's in-house multimodal models are 7x faster and 30x cheaper than rival foundation models.

The practical change is the connectors. Canva AI 2.0 pulls from Gmail, Slack, Google Drive, and Zoom. You can feed it a recorded Zoom meeting and it returns a structured report. Point it at your Gmail inbox and it builds a morning briefing doc. The output is fully layered and editable, not a flattened image.

Rolling out to the first 1M users as a research preview, then wider. If you already pay for Canva, the workflow shift is real. Most teams don't need to generate slides from scratch. They need to turn a meeting they already had into slides.

Read The Announcement →

---

⚡ PROMPTS TO TRY TODAY

3 Prompts Built for the New Codex Computer-Use Mode

With Codex controlling a real cursor on macOS, the game changed. You're no longer asking for text. You're giving an agent a goal and the keys to your apps. Specific instructions beat clever ones here. Tell it what success looks like, where the inputs live, and what to return.

These three are worth testing if you installed the April 16 update. Replace the brackets.

Prompt 1: End-of-day triager

It's 5pm. Open Slack, Gmail, and my calendar. Read anything from the last 8 hours that I haven't responded to. Group items into three buckets: needs reply today, can wait until Monday, FYI only. Draft replies for the "today" bucket but don't send them. Output a short plan for tomorrow morning's first two hours based on what's still open.

Good for catching up without spending an hour catching up. The "don't send them" clause matters.

Prompt 2: Research-to-slides pipeline

Research [TOPIC] using the in-app browser. Find 5 credible sources published in the last 6 months. Pull one quote and one stat from each. Open Keynote and draft a 7-slide deck: intro, one slide per source with the quote and stat, and a closing slide with three takeaways. Leave placeholders where you'd want a chart.

Turns "I need a deck on X by tomorrow" into "review this draft." The leave-placeholders line keeps it from inventing data.

Prompt 3: Application diff

Open the job posting at [URL]. Extract the 10 most important requirements. Open my resume at [FILE_PATH]. For each requirement, tell me whether my resume currently addresses it (with the line), partially addresses it, or misses it. For the "partial" and "miss" rows, draft a rewritten bullet I could add or swap in. Return as a 3-column table.

Different take on the resume prompt from last week. That one asked for rewrites. This one shows you exactly where you're short before you commit.

---

Cheers,

Rj