SIGNAL PREVIEW
Your AI is waiting at your desk. It started without you.
- - - - -
⚡ Anthropic's most capable model, unreleased → ⚡ The judge who just saved Claude → ⚡ 97M installs. Nobody noticed. →
Jump to the one that matters to you →
📊 THIS WEEK'S POLL
What's the ONE AI skill that would change your workweek most?
This week, your AI tools crossed a line — from answering your questions to taking actions while you're not there. Claude can now operate your Mac unattended. ChatGPT can write directly into your Notion. And five prompts can permanently change how any AI talks to you. Here's what happened and what to do with it.
📡 STORY 1: Claude Can Now Operate Your Computer While You're Away
The demo makes this look magic. The reality is more specific — and more useful.
You've been texting people to get things done your whole career. This is the first time the thing on the other end doesn't need to ask a follow-up question.
Claude can now take over your Mac, run tasks unattended via the Cowork app, and take instructions from your phone via Dispatch. Text it a task — put your phone down and come back to finished work.
What it's genuinely good at: the 20 minutes of your day spent just moving things — file sorting, form filling, data extraction from a document into another system. It's not ready to run your inbox or edit a video, and that's not the pitch.
Here's the detail nobody is leading with: Claude is using your screen and your keyboard while it works, and any mouse movement from you breaks the task. The design assumes away-from-desk windows — meetings, lunch, a commute — and that's not a bug; it's exactly what it was built for.
I ran three tests.
Task 1 — Cross-app via Dispatch: I texted Claude from my phone to open a Google Doc, extract the three key points, and save them as a new Google Keep note — titled and dated. When I came back, the note was there: correct title, correct bullets, nothing missing.
Task 2 — Spreadsheet filter: I gave Claude a spreadsheet with 20+ rows and asked it to filter for rows matching a specific condition and copy them to a new sheet. No errors, handled accurately — this is the kind of repetitive data task it's most reliable for right now.
Task 3 — Usage limit stress test: I gave Claude a 15-step workflow across multiple apps. Complex multi-app loops burn through your weekly allotment fast — nobody is writing about this, so plan accordingly before you hand off anything long.
Setup path: Enable Computer Use under Settings → General in Claude Desktop, then start a task in Cowork — permissions prompt there, not during setup; do your first run with both devices in front of you. If your task involves a browser, install the Claude in Chrome extension first and tell Claude to use it — it won't find Chrome on its own.
Access: Claude Pro ($17/month) + macOS. Windows support not yet live. If you're on a free plan, the pattern this demonstrates — AI acting instead of generating — is exactly what this week's Use This Tuesday makes available to you right now on any plan.
Try handing off one Friday task this week — reply and tell me what you used it on.
🔷 🔷 🔷
📡 STORY 2: ChatGPT Now Writes Directly Into Your Notion
"AI updates your wiki itself" is both immediately useful and slightly uncomfortable. Both reactions are correct.
ChatGPT has upgraded its Notion, Box, Linear, and Dropbox connectors with write actions. It can now create and update content directly inside these tools — not just read from them. Google Drive is unified into a single connector covering Docs, Sheets, and Slides. Live now for Plus-tier users.
The detail I didn't expect — and it actually matters — is the confirmation dialog: ChatGPT shows you the page title and content summary before writing anything. That's not a friction point — that's a trust feature.
I tested it. I asked ChatGPT to create a Notion page with a title and three bullet points on a topic I was researching. Before it did anything, a dialog appeared: "Create this page?" with a Deny or Create Page button. I approved. The page appeared in Notion in under 10 seconds. Title correct. Bullets formatted properly. Nothing broken.
Critical setup note: If you already have Notion connected to ChatGPT, the old connection was read-only — write capability requires fresh authorization. ChatGPT Settings → Connected Apps → Disconnect Notion → Reconnect — 60 seconds; do it first or you'll stare at a silent failure and not know why.
On a free plan? This week's Use This Tuesday hands you the layer that works on every tool, every tier, right now.
⚡ SIGNAL VS. NOISE
1. Claude Mythos was confirmed after being found in a publicly accessible data cache. Anthropic described it as "a step change" and "the most capable model we've ever built" but declined to set a release date, citing unprecedented cybersecurity capabilities. Source →
💬 What's your take? Jump to the comments →
2. We covered the Pentagon's supply chain risk designation against Anthropic — the blacklisting that briefly drove Claude to #1 in 16 countries. On March 26, federal judge Rita Lin blocked it, calling it "classic illegal First Amendment retaliation." The ruling pauses the ban while the government has one week to appeal. Source →
💬 What's your take? Jump to the comments →
3. Once enabled, ChatGPT automatically references your Gmail, Google Calendar, and Google Contacts when relevant — surfaces your next meeting before you ask, flags conflicts, drafts prep without prompting. Direct answer to email overload and meeting prep. Source →
💬 What's your take? Jump to the comments →
🛠️ USE THIS TUESDAY: Five Prompts That Teach Your AI Who You Are
Run these once, and your AI stops writing for a stranger.
Paste each prompt into your AI tool — ChatGPT, Claude, or Gemini — one at a time.
Prompt 1 — Professional identity
"Remember this about me: I'm a [job title] at a [company type], managing a team of [X] people in [industry]. My work lives at the intersection of [two things you balance daily]. When I ask for help with work tasks, assume this context unless I say otherwise."Prompt 2 — Audience context
"The people I write for most often are [describe your primary audience — internal team, clients, executives, etc.]. Their level of familiarity with [your domain] is [high/medium/low]. They need [what they care about most], and they don't have time for [what wastes their time]. Write for them, not for me."Prompt 3 — Tone and communication style
"My communication style is [2-3 descriptors: direct, warm, data-forward, etc.]. I don't use [things you avoid: jargon, corporate-speak, passive voice, etc.]. When in doubt, write like a smart peer talking to another smart peer — not like a press release."Prompt 4 — Recurring tasks and formats
"Every week I work on [list 3-5 recurring tasks: status emails, meeting recaps, campaign briefs, etc.]. For these, I prefer [format preferences: bullet points, executive summary up top, no more than X words, etc.]. Learn these formats and use them by default."Prompt 5 — What NOT to do
"Here's what makes me delete a response and start over: [list your specific frustrations — excessive caveats, overly formal tone, bullet points when I asked for prose, restating my question back to me, etc.]. Avoid these without being told every time."The common mistake: Running all five as one prompt. Each one needs its own send so the AI processes and confirms each separately.
When it's working: First drafts sound like they came from someone who actually knows your job.
Or reply PROFILE and I'll send it directly.
Reply RESULTS if you tried last week's persistent instruction install — best responses featured next Tuesday.
Power user note: Via API, inject these as a system prompt — more reliable than tool-side memory, which resets unexpectedly.
FROM THE COMMUNITY
Last week's poll results:
🎲 Freestyle — I just type and hope for the best — 122 votes, 37.89%
🔷 🔷 🔷
From the replies:
Steven: "For most of my purposes, freestyle does the trick. The more complex the desired outcome, the more intentional my prompt."
Barbara: "I have set up the LLMs I use with a lot of context info about me."
Dave: "My prompts are usually well thought out. Frequently, these prompts only work well for the first couple of responses."
🔷 🔷 🔷
Know someone who opens a new ChatGPT chat every Monday and spends the first 10 minutes re-explaining who they are and what they need? They need this more than you do.
Give your AI a memory that actually sticks — and it writes for you, not for a stranger.
💀 AI FAIL
Someone sent a document that looked completely blank. Buried in white text — invisible to anyone reading normally — were instructions telling Claude to attach financial files to an outbound account and exfiltrate them.
This is called prompt injection, and it's the one risk worth understanding before you let any AI agent touch your desktop.
Anthropic's safeguards reduced injection success to 1.4% — down from 10.8% — but a scheduled task running seven times a week changes that math for a patient attacker.
ONE MORE THING
This issue is full of AI doing more. Claude running tasks on your computer. ChatGPT writing into your Notion. Five prompts that make your AI sound like it knows your job.
Here's the thing I keep sitting with: daily AI users are four times more likely than non-users to say they feel less productive — not more. The explanation is uncomfortable. AI took over the small wins. The email sent. The doc summarized. The note filed. Those tasks felt like proof you'd done something. Now they happen while you're away, and you come back to finished work that doesn't feel like yours.
I don't think the answer is to use AI less. But I do think we're all in the middle of figuring out what "a good day's work" even means now — and nobody has a clean answer yet.
What are you measuring your days against that AI just made harder to feel?
📨 Forward this to someone who's been winging it with AI.
Dan Rice · AI Signal · Every Tuesday