Welcome to the monthly paid member Q&A for September 2026, where I answer your AI questions live on Substack.
This time, Dheeraj Sharma joined me for the first hour, so a few answers have his take next to mine. He also questioned whether the new always-on agents are worth learning, which I’ll get to in the second half.
Like last month, the session had two parts. First, I answered four member questions. Then I walked through What’s New in AI, which included OpenAI’s Dots & Spaces, Typesafe’s Jev, Opus 5.5 for generating animated videos, and some workflow updates on my Grok Bot, plus a few new features in Anthropic and Cursor that I’m only watching for now.
Before we start, here are some related posts you may want to read this month:
How to Make Your AI Agent Review Its Work Before You Do: the eval process I mention in Vijay’s answer.
I Built a Seven Grok Bot AI Team I Can Run From My Phone: the setup behind the Grok Bot updates in What’s New.
Turn Your Website or Slides Into a Video With HyperFrames: how to install HyperFrames before you try the Opus 5.5 video workflow.
How I Use Claude Code Routines to Run Work While My Laptop Is Closed: useful for Alex’s and Chuck’s answers, since both depend on work that runs on a schedule.
10 ChatGPT Work Features That Made Me Use Claude Code Less: background for Des’s question and the new ChatGPT launches.
Here are this month’s questions:
Vijay has a voice-writing problem: he can’t get Claude to follow his writing style, and Claude keeps making mistakes even though he already built a QA check from his own edits, so the AI drafts still don’t sound like him.
Alex pulls information from more than 500 bank websites and wants a way to track any changes and extract what matters.
Chuck trades growth stocks, is thinking about starting a newsletter, and wants a screener that tells him when the market turns bad.
Des likes Codex but burns through a day of usage in an hour on the Plus plan, and he is wary of giving an agent access to his browser and email.
We are going to tackle all of these questions in this post.
What you’ll take from this month’s Q&A
I didn’t expect this month’s Q&A to turn into such a long one. The whole session ran almost 90 minutes: I spent the first half answering questions in detail, and the rest walking you through what happened in AI over the last 30 days, how I use these tools, and what other people are using them for.
So you’ll get even more inspiration from this post.
Here’s what’s in the rest of the post:
A prompt that turns your past writing into a writing guideline: it pulls structure, rhythm, sentence patterns, and the AI habits you keep deleting from your own files, and quotes your writing as evidence for each rule.
A daily website monitor built from a Google Sheet and a log file: a scheduled task that checks every site, saves today’s results, compares them with yesterday’s, and only contacts you when something changed.
A read-only market dashboard prompt: connect a market data MCP, refresh it on a schedule, and get a buy, sell, or look-closer signal, with you making every call.
Some ways to give your agent its own email and a short list of sites it’s allowed to visit through a browser use: This is how you can manage risk if you don’t want to share your agent email and want to prevent your agents from browsing dangerous websites.
The always-on agent decision: what dots and Grok Bot did for me this month, and why Dheeraj thinks Claude Code on an always-on Mac does the same job.
What’s new that I’d actually try: Opus 5.5 with HyperFrames for product videos, ChatGPT Images 2.5 for newsletter infographics, Jev for sorting large piles of text, and ChatGPT’s new Space and Pages.
By the end, you’ll have a prompt that creates your writing guideline, plus a way to run evals so another agent reviews your drafts before you do. You’ll also learn how to:
Set up agents that check websites and market data on a schedule and only tell you when something changes
Give your agents their own email
Use OpenAI’s new dots and Space, which just launched
Create animated videos with Opus 5.5 and HyperFrames
Learn how to use Typesafe AI’s Jev to classify and categorize massive amounts of data
Let’s dive in. This is going to be a long post—buckle up!
Vijay: My AI writing still doesn’t sound like me
Vijay uses Claude to edit his newsletter. He has built a voice profile, and he even asked Claude to analyze all the edits he makes and turn them into a QA check. He’s been using Fable 5.1. The drafts still don’t sound like him.
My answer had three parts.
Build a detailed writing guideline from your past writing
The first thing I showed Vijay was my own newsletter writing guideline. It’s a file in my project that Claude reads whenever it writes or edits a newsletter, and it’s long, around 700 lines.
It covers the rhythm of how I explain things (what I expected, what happened, what I now recommend), my long-form structure, sentence-level patterns pulled from my best posts, and a list of sentence-level anti-patterns. These are the things AI drafts do that I always end up removing.
It’s that detailed because the agent needs specific rules to follow: how a piece should move from the opening to the end, when to use a pattern and when to avoid it, and which habits count as mistakes. An agent can only follow a standard it has in writing.
You don’t need to write it by hand. You can ask the agent to extract the patterns from your past writing, which is the prompt I promised during the session. I wrote it after the call, and it’s below.
I also mentioned a context folder: a few files that describe your voice, your audience, and examples of your writing, so the agent knows who you’re writing for and what good looks like in your archive.
Check this post below to explore them further.
Try Opus 5.5 for drafting
The second part is the model.
Before Opus 5.5, I drafted with GPT-6 Astra a lot, and I liked it. Since Opus 5.5 came out on September 22, I’ve been using it for long-form drafts because it follows my writing style better than Astra does in my experience.
I still use Fable 5.1, but for deep research, not writing. So my suggestion for Vijay was to try Opus 5.5 for drafting instead of Fable.
Models will keep changing, though. The guideline is the part that carries over to whichever model you use next.
Make the agent check its work before you see it
The third part is an eval process, which I wrote about in How to Make Your AI Agent Review Its Work Before You Do.
Even with a guideline, you need something that checks whether the agent actually followed it. When the agent finishes a draft, it reviews its own work against your requirements before coming back to you. You can also send the draft to a separate sub-agent that reviews what the writing agent produced.
Dheeraj added how he does this.
He has a large collection of his own writing: personal emails, direct messages, travel blog posts, and thousands of YouTube and WordPress comments. He used that collection to define the DNA of his voice. His evaluator compares each draft against that DNA and scores it. If the score is below eight, the agent goes back and writes it again.
He was also honest about how long this takes. He still sees drafts that ignore what he asked for, and he keeps feeding those findings back into the system. In his words, it won’t get fixed in one, two, or three tries. Expect to keep improving it over a few months before it starts to sound like the way you would actually write or narrate something.
I agree with him. If your QA check comes from your edits, the guideline below gives it something more specific to grade against.
The prompt: turn your past writing into a writing guideline
Before you run this, set up a folder with three subfolders:
past-writing: pieces you wrote yourself and still like. This is the main evidence.
edits: AI drafts next to the version you published after editing. This is where your anti-patterns come from, so Vijay already has most of this.
other-writing (optional): emails, DMs, comments, or quick posts where you wrote naturally. This is Dheeraj’s approach.
Then open the folder in Claude Code (or Codex) and paste this:
