ISSUE 01 · AUGUST 4, 2026
AI is all we do. We keep up so you don't have to.
WHAT WE'RE SEEING IN THE MARKET
Most companies haven't told their AI tools what the company actually does.
We hear this from more companies every week. People across the org use AI constantly, but nobody ever sat down and told it what the business actually is. Every new conversation starts from zero, so everyone ends up re-explaining it themselves, a little differently each time. The fix is writing down what you already know, once, and putting it somewhere your team can actually find it. More on how below.
THE WHY
Five people, five answers, and no way to tell which one is right
Ask five people on your team to draft the same customer email and you'll get five different answers about what's true, what's allowed, and what tone to use. The model is doing exactly what it's told. It just gets told something different by every person who opens it, because nobody agreed on what to tell it in the first place.
Anthropic ran into a version of this themselves this month, from the inside. Their team had spent a year and a half cramming Claude Code's instructions full of rules written for older, more literal models, and by July those instructions were fighting each other, with the model spending its effort resolving contradictions instead of doing the work. Their fix was blunt: cut over 80 percent of it and trust the model's judgment on the rest.
Take from it this: the fix is one page, written in plain language, covering the handful of things your team already knows about the business. That's it. No exhaustive rulebook required.
The fix takes about an afternoon. Take customer support replies as an example. Once you've written down your tone rules and what counts as routine versus what needs judgment, you can see that most replies are simple lookups a fast, cheap model handles fine, and only the tricky ones need your best model. That's the budget conversation you can't have today, when everything runs through the same expensive model because nobody's sorted routine from complex.
YOUR MOVE
Write the page, put it in a Claude Project, add your team
Pick the AI task your team runs most often. Write its one page using the prompt below, then save it to a Claude Project everyone on that task can open. Every conversation started inside that Project already has it, so nobody pastes anything and nobody re-explains the business.
The page. A Team AI Context Starter:
WHO WE ARE
[1-2 sentences. What you do, for whom]
WHO WE SERVE
[Your customer in plain terms. Who they are, what they need from you]
HOW WE TALK
[3-5 notes on tone. Words you use, words you don't]
RULES WE ALWAYS FOLLOW
[The non-negotiables. Compliance, brand, format]
WHERE A HUMAN CHECKS FIRST
[What AI should never send or act on without a person looking at it]The prompt that writes it for you. Paste this into Claude:
"Help me write a knowledge base document for a Claude Project my team will use for [the task or department]. Ask me questions one at a time about what my company does, who we serve, how we talk, our non-negotiables, and where a person has to check the work before it goes out. Write the result as a single document with clear headers, ready to upload as Project knowledge, and easy to add to later."
Then set it up:
Open Claude and start a new Project. Name it for the task or team it serves, like "Customer Support Replies," not "AI Project."
Open the Project's settings and add your page under Project knowledge. If you want every reply nudged toward it automatically, paste the same content into the Project's custom instructions too.
Add your team to the Project. Everyone who works inside it inherits that context in every new conversation.
Where it goes from here. The page grows by accretion. When a chat inside the Project goes sideways because something wasn't covered, add a paragraph. When a policy changes, update it so everyone's working from the same version. A few months in, you'll have a real knowledge base, built exactly where your team needed it, one real gap at a time.
WHAT TO SKIP
You don't need the wiki yet
Somebody is going to suggest building a full internal AI knowledge base off the back of this. That's a real project and it can wait. Start with one page in one Project, built around the task you picked above. Expand it once people are actually using that one.
THE ROUNDUP
Signals from the noise this week.
A scorecard for the AI age (OpenAI): Four questions worth asking about any AI spend, and they hold up no matter whose model you're buying. Bring the cost-per-successful-task one to your next budget meeting and watch the room go quiet. Read it here
Introducing Claude Opus 5 (Anthropic): Their new model lands near flagship Fable 5 quality on real work, at half the price, and it's already the default on Claude Max and the strongest option on Claude Pro. If your team has been paying premium prices out of habit, this is the week to actually test whether you still need to. Read it here
An opinionated guide to which AI to use to do stuff (Ethan Mollick, One Useful Thing): His Summer 2026 guide to which AI to actually use. He is most useful this year on agent mode versus chat mode, Cowork on the Claude side, Work mode on the ChatGPT side, and what changes once you hand Claude a task instead of a question. If the naming has confused you, he untangles it better than anyone. Read it here
Claude Cowork is coming to mobile and web (Claude by Anthropic): Scheduled tasks now run overnight, work keeps going after you close the laptop, and Claude still asks before anything ships. It is the same review-first habit the page above tries to build into your team, now built into the tool itself. Read it here
AI is more likely than humans to form biases when hiring (MIT Technology Review): New research on models inventing their own stereotypes out of almost nothing. The part worth your attention is that relevant context made them fairer and irrelevant context made them worse, which is this whole issue's lesson arriving from a much less comfortable direction. Read it before AI touches any more of your hiring. Read it here
If you're not sure which task to point this at first, tell us where you are with AI. We'll give you honest input on where to start, either way.
That's issue one, and that's the shape every week: the one thing worth knowing, why it matters for you, and something you can actually do about it by Friday. Hit reply and tell me what you're stuck on. I read all of them.
Justin
Why of AI