|
In this issue
|
Happy Wednesday. Last week my website got rebuilt, and the strange part is how little of the typing I did myself.
Over the last month, I asked people I trust for feedback about my site. I got it. One of them put it in a sentence I have not stopped thinking about: “I don't understand the website. I'm scared of this.” That stung, and it was the most useful thing anyone has said about my business in months.
This issue walks you through what happened next: the five-step workflow that turned a folder of hard feedback into a redesign plan, and the plan into a new site, in ONE day. You can run the same play on your own business this week.
Also in here: what a marketing assistant is made of, four things worth your time, one AI phrase decoded, and the new site itself.
From a folder of honest feedback to a new site Five steps from honest customer feedback to a live redesign, and what my assistant was allowed to finish alone. |
You already know something is off with your website and the way your business presents itself. The evidence arrives in fragments: a customer email that misreads your offer, a comment after a meeting, a friend who admits she never understood your homepage. Each fragment stings for a day, then disappears into a text thread or a call nobody wrote down. And fixing it feels like a six-week agency project, so it stays on the someday list.
That was me in August. Over four weeks, nine people gave me verdicts on my site: customers, readers, people who know me well. The scared line above came from one of them. So did the one that cut deepest for someone who sells AI work: “Did AI build this? Looks like something AI would build.”
Here is what matters for you: I did not sit down and manually rewrite and redesign my website. I sat down with Claude and wrote a plan for rewriting it, built from the findings, and then let my assistant run the plan while I slept. One day later I had a new website, designed and written, waiting only for my go to publish.
The workflow · five steps 1. Put every piece of feedback in one place. Recorded calls, text messages, emails, exactly as they arrived. Nothing cleaned up, nothing summarized. One folder. 2. Ask for the verdicts word for word. My instruction: go through all of it and pull every verdict about the site, exactly as the person said it, with who said it and where. No summaries. A summary softens exactly the sentence you most need to hear. 3. Findings before fixes. The assistant wrote nine findings and zero recommendations, on purpose. Mine boiled down to one sentence: the writing was clear to me and frightening to a reader who does not already know the material. Sit with the findings before you let anyone, human or AI, pitch you the fix. 4. Turn the findings into a plan. This is the part you do together, and it takes the longest. I spent about three hours one evening with Claude going through the nine findings one at a time and deciding what changes: which pages, what each page says, what it looks like. When the first design came back I told it that it looked like a machine made it, and we went again. By the end I had a written plan for every page, and I could read it and say yes, that is the site I want. 5. Decide what it may finish alone. At 9:35 that night I handed my assistant the finished website copy & design pass. By 7:56 the next morning all 37 pages were done; a reviewer had checked the work and approved it to present it to me for publishing. It didn't publish, per my plan instructions. Each page went live only on my go. |
What makes a plan something an assistant can run on its own? Mine had five things in it. None of them are technical. You could write all five on one page.
The plan · five parts 1. Describe the work in order as a checklist. Not “redesign the website.” A list: first read the current site and touch nothing, then build the home page, then the class page, and so on, seven steps. Each step is one box to tick. In the morning I opened the plan and saw which boxes were ticked. That is how I knew where it had got to. 2. Each step gets a label. One of three labels on every step. Run on its own. Stop and wait for my go. Leave it for me. Building a page was run on its own. Putting a page live was stop and wait. Choosing the photos was leave it for me. 3. How to check its own work. After it built a page, it had to open that page the way a stranger would, logged out, and compare what it saw with what it had built. Not a glance. Word for word. If anything was different, that page was not done, and it was not allowed to patch it and move on. 4. When to stop. If a check failed twice, stop everything and write down what happened. Do not guess. Do not start the next page. I would rather find one question waiting for me in the morning than five pages built on top of a mistake. 5. A second reader before it ran. Before I set it off, I had a second Claude session read the plan cold, with none of our conversation, and tell me what was missing or wrong. It found five things. One of them: the plan said two animations were approved, and the note on the design board still said they were waiting on me. I had said yes earlier that day and never updated the note. Small, and exactly what stalls an assistant at 2 AM. |
The overnight leg makes the best story, but the honest timeline reads: a month to collect the feedback, one morning to sort it, three hours one evening to turn it into a plan, one night for the build. It still wasn't a straight line. The next morning I spent close to four hours reviewing, answering its questions, and calling one bigger change it had stopped to ask about. Then I put the pages live myself, one at a time.
I packed steps 2, 3 and 5 into a starter you can run this week: the exact instruction that turns a feedback folder into a one-page findings list, plus the short checklist I use to decide what my assistant may finish without me. Download the feedback starter here.
The whole system in one sentence: a place where every piece of customer feedback lands, an instruction that turns it into findings you can read on one page, and a written line between what the assistant finishes alone and what waits for you.
One honest limit: a clearer website does not bring anyone to it. Every verdict in my folder came from a conversation I had personally. The redesign fixed what visitors see. Getting more visitors is my next problem, and no overnight run solves that one.
What a marketing assistant is made of Three kinds of files, all on your computer, all yours: the feedback, the instructions, and the work. |
When I say I built a marketing assistant, I mean something plainer than the phrase suggests. Mine is three kinds of files in folders on my computer. The feedback: what customers and readers actually said, collected as it happens. The instructions: how I write, who I write for, what my offers are, saved once and read every time the assistant works. The outputs: the plans, drafts and pages it produces, filed where I can find them again.
The AI is the engine. The assistant is the files. I'm the driver. That distinction matters when you pick tools, because engines are swappable and files are yours. I use Claude; if I moved to a different AI tomorrow, my assistant would survive the move. Same feedback folder, same instructions, same outputs, new engine.
If you read Issue 15 on giving your AI work a home, this is that idea grown up. A project holds the things you are tired of repeating. An assistant adds two flows: the feedback coming in and the work going out. Starting yours takes twenty minutes. Make the folder, write one page of instructions, and drop in this week's feedback as it arrives.
Four things worth your time Claude gets hands in your browser, Meta builds you a marketing assistant, AI seat prices split in two, and a new Claude I have been running since yesterday. |
Claude can now work inside your browser, on every paid plan. Claude in Chrome left beta: the assistant can open pages, click, type and fill forms using your existing logins, and take routine steps on its own with a safety check on each action. The decision this hands you is not whether to try it, it is which sites it may touch. Start where a mistake is reversible, and keep it away from anything that moves money. Anthropic, August 26 · claude.com
Meta built a marketing assistant for small businesses. Meta AI now connects to your Instagram and Facebook accounts, your Meta ad campaigns, and Google Workspace. It reads 90 days of ad performance, compares your reach against similar brands, and writes the reports. Free to start. One caution: an assistant that lives inside Meta will tend to recommend Meta's ad products. Take the reading, keep the spending decision yours. Meta, August 19 · facebook.com
ChatGPT business seats now come in light and heavy. OpenAI added Premium seats at $125 a month next to the $25 Standard seat: about five times the usage, no session cap, meant for the person on your team who leans on the tool hardest. The move worth copying into your own planning: pay by who actually delegates work to the assistant. One heavy seat plus cheap seats usually beats upgrading everyone. PYMNTS, August 11 · pymnts.com
A new Claude, and this issue was built on it. Anthropic released Claude Fable 5.1 yesterday, in the Claude app, in Claude Code and for developers, at the same price as the version before it. I moved my own work over the same day. The gain I notice is on long jobs: it holds a plan over hours without drifting, which is exactly the skill the Build above depends on. One of Anthropic's own testers said it the way this whole issue does: “I would wake up in the morning to the next phase finished.” Same rules apply. Give it a plan with a stop line in it, then let it run. Anthropic, September 1 · anthropic.com
The phrase: “AI employee” What the pitch means, what it hides, and the question that cuts through it. |
You will hear it in sales calls this fall: “hire an AI employee,” sometimes “digital worker” or “AI workforce.” Underneath, it usually means a chat assistant with a job title, wired into a few of your tools. The word employee is doing the selling. It moves your question from “what does this do for me” to “who could this replace,” and it makes a software subscription feel cheap next to a salary. My own build last week would get pitched as an AI marketing employee. What it actually is: my taste, my design, files I own, one job it does well, and a hard stop at every decision that counts. Replace nobody. Augment one person. Measure that. And the question to ask any vendor using the phrase: which single task does it finish on its own, and where does the work land when it is done?
Go and look at the new site The site this whole issue is about is live. Click around, then tell me what still confuses you. |
Everything above is now live at manual.praxisx.co. One headline, one photo, one button on the home page, where there used to be dense menus and a wall of text. Go and click around. If anything on it still confuses you, reply to this email and tell me in one sentence. It goes in the folder, and the next fix starts from it.
Want to build yours with me walking you through it? The four-week class starts Wednesday 30 September, with a Thursday group from 1 October. Ten seats a run, $1,295 for this first class. Reserve a seat, or read what the four weeks cover first.
Talk Wednesday Here is the one thing to do this week: make the feedback folder. One place where every hard sentence about your business lands the day you hear it. Ten minutes to set up, and the next rewrite of anything, your homepage, your offer, your pitch, starts from what customers actually said instead of from a blank page. |
See you next Wednesday.