If you’re new, welcome! If you’re a regular, thanks for being here as we navigate this AI shift in eCommerce. I share what I’m testing and what’s actually moving the needle. Where’s the vault you ask? Every previous edition is saved here.

In this edition (shortcuts):
You Do Not Have to Watch It Work💥
Cool Tools GPT Image 2 vs 2.5 🧰
Your Claude Code Has a Doctor 🩺
Go with the Flow: Signal to Noise 🎤
News and Events Worth Following

Who here has sat at their desk watching Claude work, waiting for it to ask you one question?

Well, lucky you, because there is a one command fix and it has been sitting in your terminal the whole time.

Claude Code is great when you are at your computer. But some of the best use cases are the ones where you do not want to sit there watching it work. Maybe it is halfway through a build. Maybe it is cleaning up a system, working through files, or handling something that is going to take a while.

You want it to keep going. You also want to leave.

If Claude is already running in your terminal, type:

Then open the Claude app on your phone and tap Code. Your session is sitting there in the list. Tap it.

That is the entire setup. No install, no account to create, no bridge to configure.

One small thing that trips people up: there is no separate Claude Code app to download. Claude Code lives inside the regular Claude app, in the Code tab. If you do not have the app yet, type /mobile in your terminal and it will show you a QR code to grab it.

Your work is still happening on your laptop.

Same files. Same tools. Same MCP servers. Same project configuration you spent a weekend wiring up. Type @ on your phone and it autocompletes file paths from your local project, because it is reading your actual disk.

This is not a copy of your session. It is not a cloud sandbox with a fresh checkout of your repo. It is a window into the machine on your desk. The phone is the remote. The laptop is the thing doing the work.

That distinction sounds pedantic until you need it. Every workflow I care about depends on something local: a credential, a CLI tool, a database connection, a folder full of client exports that is never going anywhere near a cloud runner. A remote that gives me a clean sandbox is useless for that. A remote that gives me my machine is the whole point.

I was building a system to track all the AI workflows my team has created. They were scattered across a dozen different places and nobody could find anything. One of my teammates had just told me the catalog was clogged with things they would never use, and asked for it to be split up by the job in front of them: troubleshooting, growing sales, account analysis.

I had to leave halfway through. It was time for my walk.

So I turned on remote control and walked out. Claude kept working on my laptop.

The first question came a few minutes in. Claude had noticed the catalog sheet was updated by hand and had drifted out of sync three times this year. It wanted to know whether to do a one time cleanup or build the sync into the pipeline, so the sheet regenerates itself from the source file every time. I picked the pipeline. That is not a decision I would want an agent making on its own, and it turns out it is not one I needed to be at my desk for either.

The second question came about forty minutes later. The dry run matched the plan, 80 skills across 8 tabs, and Claude wanted a yes before it wrote to the live sheet the whole team uses. I said yes and kept walking.

By the time I got home, the sheet was rebuilt and the sync script was in the repo. Under an hour, start to finish.

That is it. That is the whole story, and I keep coming back to how unremarkable it felt. I did not configure anything. I did not think about it again until my phone buzzed.

The command is the easy part. These are what make it something you actually use instead of something you tried once.

🔶 Stop having to remember the command. Run /config in your terminal and turn on Enable Remote Control for all sessions. Every interactive session now registers itself automatically. You never type /remote-control again, and you never have that moment of walking to the car and realizing you forgot.

🔶 Turn on the right notifications. Also in /config, there are two separate toggles. Push when actions required fires when Claude hits a permission prompt or a question, which is the one that matters if you want to leave. Push when Claude decides is the proactive one, for things like a long task finishing. Turn on both and your phone becomes the thing that tells you when you are needed, instead of something you keep checking.

🔶 Ask for the notification you want. You can just say it in your prompt. Something like notify me when the tests finish works. It is a small thing that makes the whole feature feel less like a setting and more like a conversation.

And a detail I did not expect to appreciate as much as I do: Claude Code skips the mobile pushes while you are actively typing in the connected terminal. So turning notifications on does not mean your phone lights up every ninety seconds while you are sitting right there. It only reaches for your pocket when you are not at the keyboard.

For what it is worth, mine are all on. Auto connect is on, so every session I open is already reachable from my phone, and the push notifications are on. I set them once and have not touched them since.

This is the part that made me smile.

If a turn runs long, Claude Code puts a Still working notice in your terminal with a Check in from your phone link on it. If you have answered several permission prompts in a row, it offers you Approve tool calls from your phone with the session link attached.

The tool is watching you sit there and suggesting you stop. It knows that a long build and a human staring at a progress indicator is a waste of the human.

Here is the part I have not seen anyone talk about, and I think it is the sleeper feature for anyone who is not a full time developer.

You can attach things from your phone, and they reach the local session properly. I have done this. The photos merged into the session nicely, as if I had dropped them in at my desk, which is not what I expected from something arriving over a phone.

🔶 Photos go straight into the conversation, and Claude Code also saves each one into a folder on your machine and hands Claude the file path. So Claude can look at the photo and copy the actual file into whatever it is building.

🔶 Other files get downloaded to your machine and passed in as normal file references.

Four things, and they are the reason most people's first attempt does not work.

🔶 You have to be signed in through claude.ai. Run claude and use /login. An API key will not do it. Neither will Bedrock, Vertex, or Foundry. This is a subscription feature, not an API feature.

🔶 Start it from a project directory, not your home folder. Claude Code needs you to have accepted the workspace trust dialog for that directory at least once, and the startup dialog never saves trust for your home directory. Run claude inside the project once, accept, then you are fine forever.

🔶 Your machine has to stay on. This is the real constraint. Not the network, the machine. Close the lid and put it in a bag and you have ended the session, not paused it.

🔶 One remote session per Claude Code instance. If you have three terminals open, each gets its own. You just cannot run two remotes off one process.

One more thing worth being straight about, because I would want to know: while Remote Control is connected, the conversation transcript is stored on Anthropic's servers. That is what keeps your terminal, your browser and your phone in sync and what lets the session recover after a network drop. Your code execution and your file access stay on your machine. If you work somewhere with Zero Data Retention requirements, that organization cannot turn this on at all.

There is a lot of attention right now on fully autonomous agents. OpenClaw, Hermes, and more arriving every month. Those systems are interesting and the people building them are solving a hard problem.

They are also a bigger commitment. More setup. More moving parts. More to understand before you are comfortable letting them run on your behalf.

So what happens to most sharp operators is not that they adopt one. It is that they look at it, decide they will get to it properly later, and do nothing at all. The gap between "my current workflow" and "autonomous agent" is wide enough that people stand at the edge of it and go back to their desk.

Remote control is the rung in the middle, and almost everyone steps over it.

You are not handing everything over. You are letting Claude keep working on your machine while you get on with your day, and staying reachable for the moments where it needs a human. And here is the part I find genuinely reassuring: you could not hand over the wheel completely from your phone even if you wanted to. From the app you can choose Manual, Accept edits, or Plan mode. You cannot select Bypass permissions from a connected device at all, and you cannot put a Remote Control session into Auto mode.

I used to read that as a limitation. I have changed my mind. It is the shape of the tool. This is not the surface where you let something run unsupervised, and it has been built so that you cannot accidentally turn it into one from a phone in your pocket while you are distracted.

That is the quiet advantage here, and it is not really about the feature. Everyone is waiting for the day they can trust an agent completely. Meanwhile there is a version of this available right now that does not require trusting anything completely, and it is one command away.

You keep your hand on the wheel. You just stop having to be in the chair.

So the next time Claude Code is running and you need to leave, type /remote-control. Open Claude on your phone. Tap Code.

Then go.

Have you tried running a build from your phone yet, or does the idea still make you nervous? Hit reply and tell me which, because I suspect the split is about fifty fifty and I would like to know.

~Ritu

Know anyone who wants to work with an agency that has AI and tech in their DNA and are pros at managing advertising and creatives for E-commerce? Forward them this newsletter and let them connect with us at [email protected].

OpenAI has a new image model! GPT Image 2.5. And it doesn't come as one model. It comes as two, called flare and sunburst.

I was ready to switch every image tool I run over to it. But a new version number is not a promise, so I ran a proper test first.

Same product. Same prompts. Two rounds: first with no reference image, then with reference images.

Most of the differences were subtle. One of them was not.

Here's exactly what I ran, so you can repeat it.

Read the prompt again. It never asks for a brand name, a label, or ingredients. Just a plain amber bottle. And yet…GPT Image 2 added all of it anyway.

A brand called Eva Naturals. A leaf logo. "Skin Clearing Serum." An ingredient list. Even benefit copy. A whole label, invented from a prompt that never mentioned one.

Now see what GPT 2.5 did…

GPT Image 2.5 ran the exact same prompt and returned a plain bottle. No label, no brand. Its flare style and its sunburst style both did this.

Why this matters: if you prompt for product shots before your own photo is loaded, GPT Image 2 can hand you a brand name and label copy you did not write.

I ran this round three times: GPT Image 2, GPT Image 2.5 flare, and GPT Image 2.5 sunburst.

Each one got two reference images. Here is the catch: both were the same Eva Naturals bottle from round 1. So the correct result was two identical bottles.

GPT Image 2.5 flare and sunburst also delivered two bottles on a wooden tray, labels intact.

However, GPT Image 2.5 sunburst made one bottle smaller. Yes, AI still hallucinates.

The prompt said to keep both products exactly as shown. A smaller bottle is a different product. The label was still right, which makes the change easy to miss. I could have repeated the prompt and it would have fixed the mistake.

Why this matters: size is part of what a shopper is buying. An image that shrinks your bottle shows a product you do not sell.

Same price per token. What you pay per image depends on how much detail each model decides to render, so round 2 ranged from $0.031 to $0.052. GPT Image 2.5 was faster in both rounds.

Claude Code picks up baggage.

Every plugin, hook and instruction file you add stays loaded, whether you still use it or not. Over time it slows everything down, and nothing tells you.

At an AI hackathon run by Andrew Erickson today, I learned about one command that checks all of it. I tried it on the spot.

It used to be a quick install check. Now it reads your whole setup, plugins, hooks, instruction files, skills, and reports back before touching anything. It helps you clean up any messes that are slowly building up in the background adding latency to your Claude operations.

My doctor diagnosis was humbling:

🔶 A forgotten plugin was timing out on every message. About 34 seconds each, before Claude even saw what I typed.

🔶 My team sync hook was registered twice. Every session started with the same 14 second run, back to back.

🔶 14 skill files with broken headers, some by a single stray colon. A skill with a broken header is invisible when you ask for it.

🔶 About 125 lines of team instructions that were repeated, contradictory, or obvious. Roughly 2,500 tokens loaded into every session for nothing.

The part to steal is not the command. It is how I answered it.

It offered one big “clean up everything” button. I said no. I asked for each item one at a time: what it is, what changes, how to undo it. Twelve items, twelve short questions.

Then it applied the lot, left the shared team edits uncommitted for my review, and kept a dated backup of my settings file.

Since then: no pause after each message, one sync at startup, fewer permission prompts. Nothing was broken. It was slow in ways I had stopped noticing.

Your setup drifts a little every week, and nothing tells you. Now something does.

I joined Danny McMillan on Go with the Flow, and the whole conversation came back to one idea: signal versus noise.

AI is speeding everything up. More tools, more messages, more output to review. Our ability to make decisions has not sped up with it.

"We're slower than the AI now in terms of how fast things are moving."

So the answer is not to read more. It is to build systems that sift out the noise, so the signal gets through.

We came at it from different angles. I walked through a routine that files unsolicited pitches before I ever see them, and a sentiment check that reads years of client email to flag a relationship drifting. Danny shared his zero inbox protocol, and why he reviews AI output in stages instead of all at once.

His line stuck with me:

"AI is not an enabler for us to think on our behalf. It's our job to verify it."

Listen here: Signal to Noise: AI Inbox Systems for Amazon Sellers

PPC Ninja is touching down at Amazon Accelerate 2026 in Seattle, September 22–24. I am thrilled to share that I’ll be joining the Amazon Ads team at their booth for a featured live interview! 🎙️✨ We will deep dive into the future of marketplace search, AI workflows, and next-gen ad performance. If you are there I would love to chat, just hit reply on this email and let me know.

🔶 Prime Big Deal Days deal deadline just moved Amazon pushed the deal submission window back from September 8 to September 22, so there is still time to load lightning deals and coupons into the October event.

🔶 Liability insurance now applies below the old threshold Starting November 2, enhanced safety category products need one million dollars in commercial liability coverage at any sales volume, removing the ten thousand dollar a month exemption low volume sellers relied on.

🔶 Compliance test requests can now be filed in bulk A new tool lets you submit test request forms across multiple products and stores at once instead of one at a time, which matters because an unresolved request suppresses the listing.

🔶 FTC filing says advertisers paid their own top bid The FTC and state attorneys general allege Amazon charged Sponsored Products advertisers close to their winning bid roughly eighty percent of the time. Amazon disputes the framing and points to falling real CPCs since 2019.

🔶 Alexa for Shopping now watches for things you might want A new Update Me When feature tracks a brand, a show, an artist or an author and alerts the shopper when something new drops, which is a fresh surface worth knowing your listing can appear inside.

🔶 The Q4 surcharge stack is now complete Peak shipping surcharges run October 25 to January 16, FBA peak fees start October 15, and a 3.5 percent fuel and logistics surcharge sits on top of both. Any Q4 price you set on September fee math is already out of date.

🔶 Codex just got a new default brain OpenAI's GPT-6 Astra, the model already being called Codex Astra since it is now the bundled default in Codex when no model is set, went live September 3. It succeeds GPT-5.5 as the flagship and is the first OpenAI model to hit the top tier on the company's own cybersecurity risk framework.

We hope you liked this edition of the AI for E-Commerce Newsletter! Hit reply and let us know what you think! Thank you for being a subscriber! Know anyone who might be interested to receive this newsletter? Share it with them and they will thank you for it! 😃 Ritu