I left a client meeting last week still thinking about a question that used to have an easy answer.
Build or buy?
For twenty years, "buy" won by default — and for good reason. Building meant hiring a dev team, waiting six to twelve months, and owning the thing forever. So you rented a seat on a big platform, paid per user per month, and quietly accepted that eighty percent of the features weren't built for you.
That trade just changed, and I don't think most owners have noticed yet. I wrote up what happened — and what it did to one client's books in a way I honestly didn't expect.
Read: Build vs. Buy: The Ground Just Shifted Under You →
Also in this issue: the AI that keeps working after you close the laptop, a ten-minute fix for the transcription tool you gave up on, and the two numbers in your business that never meet.
Almost everyone's first experience of AI is the same: you sit down, open a chat, and type. That version of AI costs you time. There's another version that gives it back, and most people who pay for AI every month have never turned it on.
Claude's scheduled tasks let you write the instructions once, pick a time, and find the work finished and waiting. It just got meaningfully better: sessions now run on Anthropic's servers instead of your machine, so a scheduled task executes with no device online at all. Laptop closed, asleep, or in a bag at the airport — the work still happens.
Here's the honest disclosure: this newsletter is the proof. The research behind every issue — the news scan, the tool checks, the analytics — runs as a scheduled task early Monday morning, before I'm at my desk. I don't do that work anymore. I review it.
Three worth setting up this week:
One guardrail, and it matters: have it draft, not send. Anthropic's own example leaves the follow-up email written but unsent, waiting for a human. Keep it that way for the first month. Then loosen it once you trust what comes back.
Scheduled tasks are on every paid plan. One note if you've been feeling generous limits lately — the doubled Cowork usage promotion ended August 5, so you're back to standard allowances.
You tried AI meeting notes once. It turned a part number into gibberish, spelled a crew member's name three different ways in the same paragraph, and you quietly stopped using it.
The fix takes ten minutes and nobody mentions it.
Open your transcription tool's settings and find the field for keyterms or custom vocabulary. (ElevenLabs just shipped this for realtime transcription; most tools have some version of it, usually buried.) Paste in the twenty words your business says constantly and the rest of the world doesn't: product names, SKUs, the three competitors you always mention, your crew's last names, the acronyms on your job sheets. While you're in there, turn on filler-word removal so "um, so, you know" stops eating the transcript.
That's it. Every meeting summary, action list, and follow-up email you generate from that day forward gets measurably better — because the thing writing them can finally spell your business.
The broader principle: most AI tools that "didn't work" were never configured. They were opened, tested cold on the hardest possible input, and abandoned. Ten minutes of setup is usually the whole difference.
Now take it one step further. A clean transcript is only worth something if it lands somewhere useful — and for most of us, meeting summaries go into an email, then into the void.
Check whether Claude has a connector for the places your client conversations already live: Slack, Microsoft 365 (which covers Teams), Google Drive, SharePoint. There are fifty-plus in the directory now, all built on the same open standard, so switching one on is a couple of clicks rather than a project.
Then build a Project for each client. A Project in Claude is a container with its own knowledge files, its own memory, and its own connected sources. Point that client's Project at where their conversations live, and let every meeting summary land there as you go. From then on, one client equals one place — call transcripts, emails, scope documents, the running history of what's been said — all of it somewhere AI can actually read.
Here's what that buys you. You can ask: "What have we promised this client, when did we promise it, and what's still open?" Today that answer is scattered across four systems and one person's memory. After this, it's one question.
Two honest caveats. Teams comes bundled inside the Microsoft 365 connector and stays read-only — Claude can search your Teams messages but can't post into them. And be deliberate about what goes into a client Project. Treat it like a shared folder, because functionally that's exactly what it is.
Connect the Dots: what you quoted + what it actually took. Two numbers live in every business and never meet. Your estimates, proposals, and quotes are in one place — a folder, a CRM, a stack of PDFs. Your actual hours and costs are somewhere else entirely — time tracking, payroll, the job file. You compare them on the jobs that go badly. You have never compared them across all of them at once.
Connect both to Claude, or just paste an export from each, and ask: "Across every job in the last twelve months, where did actual hours exceed the estimate — and is there a pattern in the type of work, the client, or the person who wrote the quote?"
The answer is almost never "we're bad at estimating." It's specific, and it's fixable: one category of work is consistently underbid by a predictable margin. Or one client's jobs always run long because of change requests nobody re-quotes. Or estimates written under deadline pressure are systematically optimistic in the same direction. You cannot see that job by job. You can see it instantly across a year.
What's a skill? A downloadable instruction pack that teaches Claude to do a specific job your way — like handing a new hire a perfect training manual. Install it once, and that expertise is on your team permanently.
This week's pick: Internal Comms — from Anthropic's own public skills library.
Internal communication is the work owners postpone until it becomes a problem. The weekly update that stopped going out in March. The policy change nobody wrote down. The incident where three people have three different accounts of what happened. Each one takes 30 to 45 minutes of staring at a blank page, so it slides — and then a team of twelve is running on hallway rumor.
This skill teaches Claude the formats: company newsletters, weekly team updates in the Progress/Plans/Problems structure, FAQ answers, status reports, leadership updates, project updates, incident reports. Ask for any of them and it follows the actual format instead of producing generic corporate mush.
Here's the part worth your attention, though. The formats live in plain, editable files. Replace the examples with two of your own updates, and from then on everything Claude writes comes out in your house format — your structure, your tone, your level of detail. That makes this the clearest example I've found of what a skill actually is. If you want to understand the concept in five minutes by reading one, read this one.
Worth noting: Microsoft just shipped the same idea into Excel and PowerPoint — reusable skills you can upload as a plain file. The pattern is spreading fast, and it's the same pattern either way.
How to get it: it's in Anthropic's free skills library (github.com/anthropics/skills) — download the internal-comms folder and upload it in claude.ai under Settings → Capabilities → Skills. If that sentence made your eyes glaze, reach out and we'll do it with you in ten minutes.
Try this prompt: "Write this week's team update in the Progress / Plans / Problems format. Pull from my calendar and my sent email for the last seven days. Keep it under 300 words, name who owns each open item, and flag anything that slipped without an explanation."
One email every two weeks — practical AI for small business owners. Free, and the forwarding kind of good.
Subscribe free →