Tool of the Week: One Claude (and the one setting that still belongs to you)
On Wednesday Anthropic folded Claude Cowork and chat into a single Claude. You open the app, ask a quick question or hand over a report due at noon, and Claude decides what the task needs. It is rolling out to Pro and Max plans over the next few weeks, then Team and Free. Enterprise admins get at least 30 days notice before anything changes.
Two things shipped with it. Claude Docs and Claude Slides are new: ask for a document and you write it together, ask for a deck and Claude drafts the slides, which you can edit in place, present from Claude, or download as PowerPoint or PDF. Claude Design now works inside a conversation instead of in its own place. All three are beta on paid plans. Everything you make lands at one shareable link that opens on a phone.
Why I care about this one. I run my side business out of Cowork. The agents that draft this newsletter, scan my inbox, and track invoices are Cowork scheduled tasks. So "which door do I use" is a decision I make a lot. As of Friday the merge has not reached my account, desktop or web, which matches the "over the next few weeks" in the announcement. Anthropic's stated reason for doing it is that people hated making that decision and that work started in one place did not carry to the other. When it lands, your Cowork chats, projects, artifacts, connectors, and skills are supposed to be where you left them, with nothing to turn on.
The part worth thinking about is not the feature list. Once Claude decides whether a task is a chat answer or a job that keeps running after you close the laptop, the routing decision is gone. The decision that remains is how it checks in with you. By default Claude asks before taking an action. You can switch that to "keep working, check in only when something needs a closer look." Anthropic's line is that you keep the final say. That is true, but the moment you flip the setting, "final say" means reviewing what already happened.
How I'd set it:
Leave the default on for anything that writes to a system of record. CRM updates, invoices, email sends, anything a client sees. Ask-before-acting is the same governance question from issue #16 (reading is a research problem, writing is a governance problem), now inside the app.
Turn it off for the boring middle. Weekly reports, research pulls, drafts you will read before they go anywhere. Interruptions on work you will review anyway are just latency.
Test Docs and Slides on something you would not miss. They are beta. Download the PowerPoint and open it in real PowerPoint before you promise anyone a deck from this path. The formatting that survives export is the formatting you can rely on.
Do not rebuild anything this week. The rollout is gradual and mine has not arrived. Check whether your account has it before you plan a workflow around it.
Scheduled work is the quiet win. The example in the announcement is a Monday report that starts on its own. If you have a recurring deliverable you build by hand every week, that is the first thing to hand over, with ask-before-acting left on for the first few runs.
A companion post the next day is titled "Projects redesigned: from folder to conversation." I have not read it yet; it is on the list.
Who this is for: anyone on a paid Claude plan who has been keeping two mental buckets, chat for questions and Cowork for jobs. Those buckets are about to be Claude's problem. Your problem is the approval setting, and it deserves five minutes of thought rather than a shrug.
Quick Hits
Anthropic is now training the small-business market I sell to. On September 15 Claude for Small Business added new workflows and integrations and kicked off the fall leg of its SMB Tour. Per Unite.AI's writeup of the post, the package is now 43 workflows and picked up 27 integrations, including Shopify, Salesforce, TikTok, Atlassian, Zoom, Xero, Gusto, Square, Stripe, and Zapier, with new sales-side workflows (lead follow-up, proposals, marketing campaigns) because roughly a third of owners asked for sales help. Forbes puts installs since the May launch at 900,000. The tour is free half-day workshops with partner Tenex in ten cities, including Raleigh, Savannah, and Tampa, plus 150-some approved trainers running workshops in their own towns and a partner webinar series through November. Why it matters: if you run a small business, the packaged version of "AI for the back office" is now free to learn in person, and Raleigh is a drive from Charlotte. Go see it before you pay anyone, including me. What the package does not do is the glue: getting your actual data into those workflows and deciding what runs unattended. That is where the work still is, and it is the same approval question as the Tool of the Week.
DeepSeek made the clock a line item. DeepSeek-V4.1-Flash shipped September 10 with native image understanding, a 1M-token context window, and a new model name, deepseek-flash. The number that matters is the price schedule: $0.15 per million input tokens and $0.60 per million output off-peak, double that at peak, and peak is defined as 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. In Eastern time that is 9 PM to midnight and 2 AM to 6 AM on weeknights. The whole US business day is off-peak; a nightly batch job at 3 AM ET is not. Two operational notes from the changelog: the old V4 Flash model names are now silently routed to V4.1 Flash, and V4 Pro, which was scheduled to go away September 14, is staying "in response to user demand." Coverage reports the weights are on Hugging Face under MIT; that part I have not verified against DeepSeek directly. Why it matters: two lessons from #24 in one release. A model alias that quietly moves under you is a reason to pin versions and re-run your eval. And if you have a Claude-SDK pipeline, DeepSeek exposes an Anthropic-format endpoint, so an A/B costs a base URL change, not a rewrite.
OpenAI is turning the website chatbot into ad inventory. On September 16 OpenAI began testing Sponsored Agents with select US advertisers. After a relevant ad in ChatGPT, a user can start a clearly labeled conversation with a business-sponsored agent, ask follow-up questions, and follow a link to the business's site when ready. Alongside it: an Ads Manager plugin that lets advertisers create, update, and analyze campaigns by prompt, suggested ad copy and imagery generated from the landing page, and new ChatGPT ads connections for HubSpot and Shopify. (Issue #22 covered the ChatGPT Ads launch; this is the next step.) Why it matters: the "chat with us" widget on your site is a product OpenAI now sells inside its own app, on its own terms, with the hand-off to your site at the end. Whether that is worth money depends on one question you can answer for free: are your customers asking ChatGPT about your category at all? If the answer is no, this is not your channel yet.
Prompt of the Week: The Unattended Task Spec
The one-Claude change means more work runs while you are not watching. Most of the failures I have seen in unattended AI work are not the model being wrong. They are the task being underspecified: no definition of done, no list of what needs a human, no check on return. This prompt turns a loose request into a spec you can hand over with the approval setting off and still sleep. I use a version of it before any scheduled task goes live.
I am about to hand this task to an AI assistant that will run it without me watching:
[paste the task in plain words, e.g. "Every Monday, summarize last week's pipeline changes from HubSpot, flag anything that slipped, and draft the five-slide leadership update."]
Turn it into an unattended task spec with these sections, and nothing else:
1. GOAL: one sentence. What exists when this is done that did not exist before.
2. INPUTS: every source it needs (systems, files, connectors) and what to do if one is unavailable. Default: stop and report, do not guess.
3. DONE CRITERIA: 3 to 5 checkable statements. Each must be verifiable by looking at the output, not by trusting the process.
4. FREE TO DO: actions that can run without asking (reads, drafts, summaries, files written to a scratch location).
5. MUST ASK FIRST: actions that need my approval before they happen (anything sent, posted, updated in a system of record, deleted, or spent). If unsure which list an action belongs in, put it here.
6. ON RETURN, I CHECK: the 2 or 3 things I look at in under five minutes to decide whether to trust this run.
7. FAILURE MODE: what a bad run looks like and what the assistant should do instead of pushing through.
Be specific to my task. Do not add generic advice.Run it on your next recurring deliverable. If section 5 comes back empty, you have not thought hard enough about what the task touches.
One last thing
Like what you're reading? Forward it to someone who is trying to figure out where AI actually fits in their work. And if you want a second set of eyes on your own workflow, book a free 15-minute audit and I will find your top 3 time-wasters.
Scott