Every Monday I sit down for 20 minutes and go through what happened the week before. Two weeks ago that review told me to stop doing the thing that was working.
I'd started making carousel posts on LinkedIn, the kind you swipe through. The first one got 53 views. A plain version of the same idea, which took me 10 minutes instead of an hour, got 130. Easy call. I wrote down "stop making these" and moved on.
Last Monday I checked those two again. The first one wasn't at 53 anymore. It was at 109. It had doubled while I wasn't looking. And the next one I made, the one I almost skipped, turned out to be the best thing I'd published in 3 months.
The review wasn't wrong. I read it 3 days too early. This issue is about how to run a weekly review that catches that mistake instead of making it.
───
🗞️ THIS WEEK IN AI
1. Anthropic released Claude Opus 5, and you can now dial down what it costs you
Anthropic released Claude Opus 5 on July 24. The part worth your attention is a new control. You can set how much effort it spends on a task, low, medium, or high, and pay for what you asked for. Anthropic is pitching this one at everyday business work rather than research projects.
Why this matters: Most people run every question at full power, including "rewrite this email." If you use Claude regularly, the effort setting is the difference between a tool that scales with your work and one that quietly runs up a bill. Use low for drafting and cleanup, high when the answer actually has to be right.
2. Claude's promotional API pricing ends August 31
Anthropic confirmed that the introductory rate on Claude Sonnet 5 ($2 per million input tokens, $10 per million output) runs through August 31. On September 1 it goes to $3 and $15, roughly a 50% increase.
Why this matters: This only hits you if you or one of your tools pays Anthropic by usage rather than a flat monthly subscription. Plenty of small business automation tools do exactly that and pass the cost through. If you run anything on Zapier, Make, or a custom workflow, check what it costs you in September before it surprises you.
3. OpenAI's ChatGPT Work keeps expanding what it will finish for you
ChatGPT Work, which launched in early July and is still rolling out, turns plain-language instructions into finished documents, spreadsheets, presentations, and simple websites. It also runs scheduled tasks, so a recurring report can happen without you starting it.
Why this matters: The gap between "AI helped me draft this" and "AI produced the finished file" is closing on the work small teams do least well, which is the recurring report nobody wants to own. Worth testing on one report you already produce every month.
───
🛠️ THIS WEEK'S TRICK
Run a Weekly Review That Argues With You
Most weekly reviews just restate what you already believe. You look at the numbers, you find the story you expected, and you move on. The version below is built to catch the two mistakes that actually cost money: reading noise as a trend, and judging something before it has had time to produce a result.
Step 1: Dump the raw week, not your summary of it
Don't write "traffic was down." Paste the actual numbers. Whatever you already track is enough: revenue, leads, calls booked, hours spent, quotes sent, site visits. If you track almost nothing, that's fine, use what you have.
That matters because the moment you summarize, you've already decided what the week meant. You want the AI looking at the same raw material you saw, not your conclusion about it.
Step 2: Make it separate the signal from the noise
Paste this into ChatGPT or Claude:
I'm reviewing last week for my business. Below is raw data, not my interpretation of it.
The numbers: [paste whatever you track: revenue, leads, calls booked, hours spent, quotes sent, site visits]
What I did differently this week: [any change you made, or "nothing"]
How long this kind of result usually takes to show up: [e.g. "leads take 2-3 weeks to close", or "I don't know"]
Do 4 things, in this order:
1. Tell me what actually changed versus what looks like normal week-to-week variation. Say plainly which of these numbers are too small a sample to mean anything.
2. Name the single number most likely to reflect something real.
3. Tell me what I don't have enough data to judge yet, and how long I'd need to wait before it means something.
4. Ask me the 2 questions you'd need answered to be more confident.
Don't congratulate me and don't soften bad news. If the honest answer is "nothing meaningful happened this week", say that.
Step 3: Write down the date you're allowed to judge it
This is the step that saves you. When the answer to item 3 is "you need 6 more weeks," put that date on your calendar and stop evaluating it until then.
I skipped this step and nearly threw away the best-performing thing I make. I'd checked that number on my schedule instead of its own.
What this works best for:
Weekly or monthly revenue and lead numbers
A new service, offer, or price you're testing
Ad spend and marketing channels
A new hire in their first few months
Any change where the result shows up long after the effort does
One honest limitation: the AI can't tell you how long your particular thing takes to produce a result. You have to supply that, and it will guess badly if you leave it blank. If you genuinely don't know, that's the real finding of this week's review, and it's worth more than any number on the list.
───
🔧 TOOL OF THE WEEK
ChatGPT Projects (https://chatgpt.com)
A Project is a folder inside ChatGPT that keeps its own files, instructions, and chat history. Anything you put in it stays available to every conversation you start there.
For this issue's trick, that's the whole game. Run your review in a regular chat and every week starts from zero. Run it in a Project and week 6 can look back at what you said in week 1, notice that the thing you called dead in March is now your best channel, and tell you so.
Set one up called "Weekly Review," paste the prompt above into its instructions, and do the review there every time.
Price: Included on the paid plans, starting at $8/month for Go and $20/month for Plus.
Best for: Anyone who is going to run this review more than twice. If it's a one-off, a normal chat window is fine.
Worth trying? Yes, but only if you commit to the habit. The value is entirely in the accumulation, and a Project with 2 entries in it is just a folder.
───
💡 PROMPT OF THE WEEK
Use this the moment you catch yourself about to quit something.
I'm about to stop doing [the thing you're considering quitting].
What I know:
- How long it's been running: [time]
- What I expected by now: [expectation]
- What actually happened: [result]
- What it costs me to keep going: [time, money, or both]
Before I decide, tell me:
1. How long does this kind of thing usually take to show a real result?
2. Based on that, is my sample big enough to make this call yet?
3. What would I need to see in the next [timeframe] to justify continuing?
4. If I stop now, what is the most likely thing I'd turn out to be wrong about?
Be direct. If I have enough information to quit, tell me to quit.
A tip on getting a straight answer: add one line about what you want to be true. Something like "I'm hoping you tell me to keep going." Naming the bias up front makes the model likelier to push back on it instead of agreeing with you.
───
👋 THAT'S A WRAP
Two weeks ago I asked a 3-day-old number to tell me something it would take 3 weeks to know. It gave me an answer, and I believed it.
Most things in a small business pay off later than they cost. Ads, hires, new services, referral relationships, all of them look like failures for a while, because that's what the early part of working looks like.
Pick the thing you've been quietly thinking about killing. Run the Prompt of the Week on it. Then either quit it on purpose or write the date on your calendar when you're actually allowed to decide.
If this was useful, forward it to someone who pulled the plug on something too early and still wonders about it.
See you next Friday.