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.

By the way, if this newsletter is turning into what sounds like a programmer's manual, my recommendation to you is to soak it in anyway. Trust me, you will need this stuff in the agentic era!

In this edition (shortcuts):
🗞️ One Action. Half the Tokens💥🚀
 CLI vs MCP, the 60-Second Decision 🛠️
Seedance 2.0 for Video Production
News Worth Reading
Upcoming Events

Who here is frustrated because you frequently run out of tokens? Well, lucky you because today I am going to show you a game-changer!

If you are on Claude, open the Mode picker in Claude Code right now. (here’s a screenshot below if you have not seen it.)

A menu pops up. Five modes. Most sellers I talk to have never opened this menu. They use the default mode that came with the install and they burn tokens like the bill is free.

It is not free. And one of those five modes is sitting there waiting to save you half your spend on every complex task.

That mode is Plan Mode. And if you pair it with one extra step, you have a workflow that quietly beats the people running Claude on autopilot.

This is the workflow my podcast co-host Danny McMillan has been refining for months. He calls the second piece the push back step. We walked through the whole thing on the latest Sellers Sessions episode: Building Repeatables in Claude Skills, CLI vs MCP, and Token Discipline. Worth a listen on its own. But here is the short, hands-on version.

Quick note before we dive in. The examples below use Claude Code because that is where I run most of my work. The technique itself is universal. Cursor has it as Plan mode. Windsurf has it as Cascade plan. Any AI coding tool with a plan or agent mode plays by the same rules. Only the labels change.

Every Claude Code session runs in one mode. You cycle through them from the Mode picker (screenshot above). The exact shortcut depends on your platform and your keyboard settings, so the picker itself is the safest entry point. Here is what each one does.

🔶 Ask permissions. Claude asks before every action. Slowest. Safest. Use when you are new to a codebase or about to do something destructive.

🔶 Accept edits. Claude makes edits automatically but asks before running commands. The default for most users. Fine for routine work.

🔶 Plan mode. Claude reads, thinks, and proposes a plan. Does not edit. Does not execute. Just plans. This is the one we are here for.

🔶 Auto mode. Claude runs without permission prompts until something fails. Fast. Use only after you trust the plan.

🔶 Bypass permissions. Claude does everything without asking. Use only when you have verified the plan and you are inside a worktree or sandbox.

Most sellers I see live in Accept Edits or Bypass. They skip Plan Mode entirely. That is the leak.

Plan Mode does not edit your files. It does not run shell commands. It reads what it needs and writes a structured plan back to you.

The tokens it spends are tiny compared to a full run. Reading is cheap. Planning is cheap. The expensive parts of a Claude session are the exploration loops, the wrong edits, the rollbacks, and the regenerations that come after each mistake.

In Plan Mode, none of that happens. You get the model's best read of the task before any tokens get burned on doing it wrong.

So the question is not should I use Plan Mode. The question is what do I do with the plan once I have it.

The default answer is "approve and execute." That is fine. But it leaves the biggest savings on the table. Because the plan you got on first pass is rarely the plan you want.

Danny's contribution is the discipline of what to do between getting the plan and executing it.

Instead of clicking approve, you push back. You ask Claude to defend the plan. Question its assumptions. Find the cheap path. Show its work.

These are the four prompts I now run on every non-trivial Plan Mode output:

🔶 "Walk me through the most expensive step in this plan and tell me why it has to be that expensive."

🔶 "What assumption are you making here that, if wrong, breaks the whole plan?"

🔶 "Is there a cheaper way to get the same outcome?"

🔶 "Which step could I skip without breaking the result?"

Run them in order. Read the answers carefully. Half the time Claude will rewrite the plan on its own after the second prompt. The other half it will defend a step that you genuinely needed to know was that expensive.

A first-pass plan optimizes for getting the task done. A pushed-back plan optimizes for getting the task done with the fewest tokens.

Those are different objectives. The model will not switch from one to the other unless you tell it to.

When you run the push back prompts, you are forcing a reframe. The model has to argue with itself. Steps that survive the argument are the ones that actually mattered. Steps that get dropped were padding the model added because nobody told it to be cheap.

That is the quiet advantage. The pros are not using a different model. They are running the model with one more conversational loop before they let it touch anything.

Last week I asked Claude to mine 90 days of search term reports for one of our accounts and propose a refreshed negative keyword list.

The first-pass plan had 11 steps. Read every search term file, build a per-campaign negative list, run a final dedupe pass, then format for Amazon bulk upload. That plan would have worked. It would also have spent half an hour re-exploring patterns I had already documented in the account brief.

So I pushed back.

Claude flagged that it was assuming each campaign needed its own threshold for what counted as "underperforming." I told it the threshold was uniform across the account. The revised plan dropped from 11 steps to 6.

Claude pointed to the read loop, which was opening every search term file twice (once for filtering, once for the negative-list write). I asked for a single pass with a buffered write. Down to 5 steps.

Claude suggested running the three campaigns as parallel sub-agents instead of sequentially. Same total token spend, a third of the wall clock time.

That final plan was the one I executed. It is not the plan I would have run without the push back loop.

Honest caveat. Plan Mode is overkill for small tasks.

🔶 One-file edit. Just do it.
🔶 A quick grep or lookup. Just do it.
🔶 A task you have done five times before. You already know the plan.

But the moment a task touches three or more files, or you would struggle to write the plan yourself in two minutes, Plan Mode pays for itself. The push back loop pays for itself again on top.

The rule I use: if the task feels complicated enough that I want to think before I start, Claude probably does too.

The technique transfers cleanly out of pure code work. Every one of these is a Plan-Mode-plus-push-back candidate.

🔶 A bulk campaign restructure. Plan first. Push back on which campaigns to touch, in what order, with what rollback if performance drops.

🔶 A multi-account performance audit. Plan first. Push back on which signals are actually predictive vs which are noise.

🔶 An ad copy refresh across 50 ASINs. Plan first. Push back on whether you really need new copy on all 50 or whether the bottom 10 by CTR carry the whole gain.

🔶 "Build me a Claude skill that does X." Plan Mode is the entire conversation. The skill is the artifact of the conversation. The push back loop is how the skill ends up tight instead of bloated.

The pattern is always the same. Plan first. Push back. Then run.

Plan Mode has been sitting in Claude Code the whole time. Cycle through it once and you understand why most users skip it. It looks like a slower version of normal.

It is not slower. It is the version that does not waste your tokens on the wrong path.

The sellers who learn the Mode picker and the push back loop are going to look like they are running tighter, cheaper sessions than the rest of us for the next year. They are not on a different plan. They are just on a different mode.

Have you been running Plan Mode already? What is the push back prompt that has saved you the most? Hit reply and let me know. If you want the four prompts as a copy-paste block plus the Mode picker how-to, hit reply with "plan mode" and I will send it over.

Know someone who is still running Claude Code on Auto and complaining about the bill? Forward this to them and earn Karma points.

PPC Ninja is helping brands future proof their listings for AI, helping you build RUFUS enabled, stunning images and videos with AI. Hit reply on this to chat with us. Explore how we can scale your content production across Social media, Amazon ads, Amazon Posts efficiently and affordably.

Every week somebody asks me whether they should install a new tool as a CLI command or as an MCP server. Both work. They are for different things. Here is the rule that takes 60 seconds to apply.

🔶 Use the CLI route when the tool has a local binary you can run from a terminal. Examples: gws for Google Workspace, gh for GitHub, bq for BigQuery. Fastest setup. Lowest moving parts. Survives an MCP outage.

🔶 Use the MCP route when the tool only lives in the cloud, has no local binary, or you want one integration usable across multiple AI clients (Claude Desktop, Claude Code, Cursor, etc.). Build it once. All your AI surfaces talk to it.

🔶 Use both for the same tool when: never. Pick one and document the choice in your project notes. Installing both is the fastest way to get your AI to call the wrong one when you are not looking.

For sellers specifically:

🔶 Pulling BigQuery data, sending email, reading sheets, listing GitHub PRs → CLI route. These tools already have rock-solid command line versions.
🔶 Figma reads, Canva pulls, Notion writes, Linear tickets → MCP route. These are cloud-native services where MCP servers shine.
🔶 Anything specific to a single niche workflow that you only use in one AI client → just a CLI is fine.

The quiet advantage. Pros pick once and move on. Amateurs install both, run into the duplicate-command-path bug, and lose half a day debugging it. Pick the rule. Apply it for 60 seconds. Move on.

We unpacked this in more detail on the same Danny episode Building Repeatables in Claude Skills, CLI vs MCP, and Token Discipline. Worth a listen if you are about to install your next batch of tools.

The PPC Ninja team has been busy creating videos for our clients. And we cannot stop using Seedance 2.0. It is a game changer.

If you want the deep walkthrough that pulled it all together for us, watch this one: Seedance 2.0 is CRAZY for AI Filmmaking, Full Course. Twenty minutes. Worth every one of them. Below is the short version with the techniques that translate cleanly to Amazon product video.

Most video models give you one or two reference images and a prompt. Seedance 2.0 takes 9 images, 3 video clips, and 3 audio clips in a single generation. That is fifteen reference assets shaping one clip.

For sellers, that is the whole game. You can stack the product hero, a lifestyle scene, brand color and lighting plates, a motion reference, and your recorded voiceover. All into one prompt. The model resolves them into a coherent clip with native stereo audio baked in.

Under the hood, ByteDance's Temporal Anchor tech is the anti-morphing layer. It keeps your product shape stable across the clip. Most AI product videos drift halfway through. This one does not.

Know anyone who needs an AI-first agency to build cool, creative videos for E-commerce? Forward them this newsletter and let them connect with us at [email protected].

🔶 RUFUS is now Alexa for shopping!
https://www.linkedin.com/feed/update/urn:li:activity:7460329139674198016/

🔶 Anthropic New Monthly Limits: Starting June 15, paid Claude plans will transition to a "metered" agentic era, which includes specific monthly credits for programmatic usage across tools like the Claude Agent SDK and Claude Code GitHub Actions.

🔶 Codex 5.5 is Now Better Than Ever (I am using both Opus 4.7 and Codex 5.5 equally for coding projects)

May 27, 2026, Seoul
I am honored and excited to take the stage once again at the Amazon Ads Prime Day Summit in Seoul. I will be sharing the stage with Bradley Sutton, Bopyo Park, and Ha Eun-young. My session: Getting Ready for Conversational AI: How Rufus Is Reshaping Product Discovery.

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