Seller Sessions Amazon FBA and Private Label

Danny McMillan and Shubhash unpack the Fable 5.1 vs Astra debate, the six-stage AI harness, and why your folder structure is the real cost lever.
Shubhash joins from his new home in Dubai to dig into the two biggest AI releases of the fortnight, Fable 5.1 and Astra, dropped within 72 hours of each other. Instead of relitigating which model wins, he makes the case that "which model is better" is the wrong question entirely, and spends the episode showing why.
You'll hear the real cost story behind cached tokens, why a shiny 3D render says nothing about which model is actually smarter at your business's real work, and Shubhash's six-stage "harness" framework for judging any AI tool. Danny closes with a teaser: an enterprise AI consultant's folder-as-operating-system approach that could cut token burn without touching the model at all.
Key Topics
Fable 5.1 vs Astra - what a codebase-drift benchmark actually revealed, beyond the marketing screenshots
The cached-token price trap - why identical list prices don't mean identical bills
The six-stage harness - job, prompt, tools, model, failure handling, pass mark
Model-switching without the cost trap - avoiding preloaded tools when routing to third-party APIs
Claude Desktop as a single workspace - Danny's parallel-conversation, no-VS-Code setup
Folders as the AI operating system - a teased framework for scoping context with markdown files
Timestamps
[00:01] Shubhash joins from his new home - relocated to Dubai
[00:57] Agenda: Fable 5.1 and Astra, launched 72 hours apart
[01:24] Danny's pushback - "are they building anything?"
[01:50] Shubhash: "which model is better" is the wrong question
[02:06] Ellis's LinkedIn benchmark - Astra similar quality to Fable 5.1, lower cost
[03:13] Where Astra broke - drifting from existing codebase conventions on the hardest tasks
[04:29] Why comparison screenshots are confirmation bias - published by the model's own maker
[05:35] Shubhash's five-point checklist before adopting any new AI tool
[06:59] "A model is just an engine" - the car and harness analogy
[08:28] The real price difference - cached token reads, 25c vs $1 per million
[09:35] Season ticket vs pay-on-the-gate pricing
[09:54] How few sellers actually examine their AI bill
[11:41] Avoiding the third-party API cost trap - tool preloading
[12:20] Real example - a $4.50 job cut to 35p once preloading was stripped out
[12:40] Where Astra genuinely wins - 3D renders, exploded views, web design
[13:46] Splitting by job type - Astra on design, Claude on knowledge work and recovery
[14:30] The six-stage harness framework introduced
[17:12] "The model is the sex and sizzle" - Danny's framing
[18:21] Shubhash's VS Code setup - Claude and Codex extensions handing off work
[19:52] Danny's Claude Desktop-only workflow, no VS Code
[20:22] Finder and Spotlight optimisation, Kimi K3 kept to the terminal only
[23:39] Widgets over walls of text - "like a kid's colouring book"
[24:24] Natural-language plan on top, technical plan underneath, on request
[24:45] Freeform and stylus for complex problem-solving
[25:50] Raycast as a free Spotlight replacement
[27:08] Why "you don't need instruction files anymore" ignores who's paying for the tokens
[27:56] Teaser - an enterprise AI consultant's file-structure system
[29:15] The folder-as-AI-operating-system concept
[31:27] Three-layer breakdown - the map, context rules, work tools
[33:45] Wrap-up - Shubhash's three-question pre-switch checklist
Key Takeaways
"Which model is better" is the wrong question - the harness around the model decides the outcome far more than the model itself.
List prices hide the real cost - cached-token rates can be 4x apart even when headline pricing looks identical.
Split tasks by strength, not loyalty - Astra ahead on visual and web design work, Claude ahead on knowledge work, long documents and error recovery.
Tool preloading is the hidden API cost - stripping it out cut one job from $4.50 to 35p.
Your folder structure is your harness - scoping Claude to only the context a task needs cuts token waste before any model swap is needed.
Notable Quotes
"A model is just an engine. If you haven't built the car, the chassis, the tyres, everything else around it, a faster engine isn't getting you anywhere." - Shubhash
"You don't pay your goalkeeper up front just because he's the best athlete in the squad." - Shubhash
"The model is the sex and sizzle, not the bit underneath it." - Danny McMillan
"Before you chase the next model, write down only what you know... get that right and we are good." - Shubhash
Resources Mentioned
Fable 5.1 - the latest release under discussion, benchmarked against Astra on real codebase tasks
Astra (GPT-6) - OpenAI's release, strong on visual/web design work, weaker on holding existing coding conventions
Raycast - free Spotlight replacement Shubhash uses for local search
VS Code (Claude + Codex extensions) - Shubhash's multi-agent handoff setup
Claude Desktop - Danny's single-workspace setup: parallel conversations, in-app browser and file viewer
Freeform (iPad + stylus) - Danny's tool for mapping out complex problems before handing them to Claude
Connect
Shubhash - Not a Square, now based in Dubai (relocated from London)
Seller Sessions is the leading podcast for advanced Amazon sellers, hosted by Danny McMillan. Ritu returns next week for Go With The Flow; Shubhash is back next month.

Danny McMillan and Sim on model switching, OpenRouter cost hacks and whether AI spend for Amazon teams is actually worth it.
A quick solo show this week — Dorian's out sick, Matt's unavailable, so it's just Danny and Sim covering how to control AI spend without gutting output.
Sim walks through the real cost pressure of a 25-person team on Claude: five figures a year, uncapped Fable usage, and no way to see per-person burn. Danny counters with the setup he's built to fix exactly that — routing grunt work through OpenRouter and Kimi K3 without preloading tools, which took one job from £4-5 down to 35p. They land on a practical split: Fable 5 for planning, a cheaper model for the build, staged "cascade" plans to keep context windows under control. Sim also shares a genuinely wild same-day case study — a 100-video yoga app build for about £300 — and Danny pushes back on the whole framing: stop looking at AI as a cost line and start looking at what it's generating.
Key Topics
Picking one AI provider and sticking with it - the cost of switching a trained team, and why Astra being better than Fable isn't reason enough to move
Why teams actually burn tokens - heavy browser automation, parallel video work, and vibe-coders scaling far past their job description
The OpenRouter cost hack - Kimi K3 without preloaded tools, cutting one job's cost by roughly 90%
Cascade planning - staging big builds across multiple sessions to control context window burn
Chinese model options - GLM 5.3 and LM Studio AI as free-to-cheap alternatives
Grok as a low-friction accessory - personal automation for AI-hesitant team members
ROI framing - weighing AI spend against what it actually produces, not just what it costs
Timestamps
[00:01] Quick solo show - Dorian ill, Matt unavailable, flight to Mallorca later
[01:02] Wishing Dorian a speedy recovery, back next month
[01:17] Kicking off: token costs, Astra's release
[01:42] Why constant provider-switching burns team trust and training
[02:10] History: Google/Gemini 18 months ago, Claude masterclass in December, full team switch
[02:29] Sim's Opus 5 complaint - unclear responses, had to cross-check with Sonnet
[02:58] Fable's cost problem - no way to cap runaway usage
[03:28] Costs into five figures a year; splitting power users vs Chinese-model users
[03:54] The NVIDIA DGX Spark purchase for local model runs
[04:32] Danny's setup: Kimi in the terminal, Claude Desktop for daily work
[04:59] The two real causes of heavy token burn: browser automation, parallel video editing
[05:58] The "I've got ADHD" repo - controlling how Claude communicates
[06:25] Claude Desktop's concise-response setting, plus widgets vs terminal mind-mapping
[07:27] Auditing what each team member actually uses AI for before cutting costs
[07:56] Why max-plan tool loading is free but third-party API tool loading isn't
[08:26] "OpenClaude" - a terminal alias that loads Kimi without preloading tools
[09:24] Sim's example: a warehouse dispatcher now vibe-coding time-tracking apps
[09:54] The $20-to-Max "purgatory" middle ground
[10:44] Danny's recommendation: Kimi K3 for the team, Claude Desktop kept for flexibility
[11:12] Splitting by task: Fable 5 for planning, K3 for the build
[11:40] Cascade plans - staging big jobs across sessions to control context
[13:04] Real numbers: four parallel video edits, £4-5 down to 35p after removing tool preload
[14:07] LM Studio AI and the free-credit-then-continue usage hack
[14:43] GLM 5.3 - "Opus 5 good," and it speaks plain English
[15:14] Grok/Grokbot - zero-friction agent automation from a phone
[16:21] Building a review-flagging SOP from a LinkedIn post in one sitting
[16:38] API scraper tip: Apify or Monid for cheap review scraping
[19:00] Team breakdown: 5 on Max plans, ~18 on $20 plans
[19:22] The frustration of no pooled token allocation across a team
[20:12] The "Claude effect" - wait 21 days before reacting to a new release
[21:08] Astra vs Fable, and why a full company switch is a bigger decision than it looks
[22:04] Reframing: only 3-4 people actually need heavy-lifting models
[23:37] Danny's pushback: what is the AI spend actually generating?
[23:59] Case study: ~£300 to edit 100 yoga app videos with real footage and AI voiceover in a day
[26:13] ROI framing: AI cost against revenue and headcount value
[26:45] Team-wide Max plan cost: roughly £30k/year - "just a salary"
[27:08] The local GPU reality check - a 90GB model locks the RAM, breaks other flows
[28:03] RunPod and HyperStack as pay-by-hour alternatives to token metering
[28:34] Wrap-up - Danny off to the airport
Key Takeaways
Constant model-switching costs more than it saves - retraining a team and rebuilding skills outweighs chasing the newest release; wait it out, then decide.
Tool preloading is where the real API cost hides - stripping it out cut one job's cost by roughly 90% with no drop in output.
Split the model by the task, not the team - planning on a stronger model, execution on a cheaper one, staged across sessions to control context burn.
Audit usage before cutting spend - most token burn traces back to a handful of people doing genuinely heavy work, not the whole team.
Cost only means something next to output - a £300 same-day build replacing two weeks of manual work reframes what "expensive" actually means.
Notable Quotes
"The last thing you want to be doing is getting people affiliated with the software, get them used to it, and then suddenly taking that away." - Sim
"Once you took that out of the equation, I've got it down to about 35 pence for the same work." - Danny McMillan
"Have you looked at what it's adding?" - Danny McMillan
"It's not even remotely comparable." - Sim
Resources Mentioned
OpenRouter - model-switching gateway, used here to route to Kimi K3 without preloaded tools
Kimi K3 - the cost-efficient model for day-to-day build work
GLM 5.3 - Chinese model tried via LM Studio AI, described as "Opus 5 good"
LM Studio AI - harness for running Chinese models, including a free-credit usage workaround
Grok / Grokbot - low-friction personal AI agent with built-in scheduling
Apify / Monid - API scrapers used for cheap-at-scale review data pulls
NVIDIA DGX Spark - local hardware for running in-house model flows
RunPod / HyperStack - pay-by-hour GPU rental as an alternative to token-based pricing
Connect
Sim - Amazon seller and co-host
Seller Sessions is the leading podcast for advanced Amazon sellers, hosted by Danny McMillan.

Description: Danny McMillan and Ritu Java on AI-built email triage, pest control and client sentiment analysis for busy Amazon agencies and sellers.
Ritu Java returns after recovering from a serious ankle injury, and dives straight into the theme of the episode: managing cognitive load as AI accelerates faster than we can adjust to it. You'll hear how Danny and Ritu are each solving the same problem from different angles — building systems that manage the noise so the signal gets through.
Ritu walks through two Claude-built systems running on her Google ecosystem: a "pest control" routine that reads her inbox twice daily and files unsolicited pitches before she ever sees them, and a client sentiment analysis system that reads years of email history to flag which client relationships are improving or deteriorating. Danny shares his own zero-inbox protocol — a labelled morning triage that clears by 8:33am UK time — and both agree the real skill now is boundaries, not more automation.
Key Topics
AI acceleration and cognitive load - why keeping pace with AI now outstrips our own working speed
Pest control inbox routine - a Claude + Gemini API system that filters unsolicited email using "left brain, right brain" logic
Reply radar - surfacing the emails that genuinely need a response, out of tens of thousands
Zero inbox protocol - Danny's labelled, time-boxed morning triage system
Client sentiment analysis - mining historical email threads to score relationship health
Cascading plans - reviewing AI-generated plans in stages instead of all at once, to protect decision quality
Timestamps
[00:00] Ritu returns after an ankle injury, catches up on the last few months
[01:34] The acceleration of acceleration - falling behind AI's own pace
[02:29] Danny's all-in-Claude mandate for his team
[05:36] Cognitive load and the need for "a skill to manage the other skills"
[06:39] Danny on verification systems, sign-off protocol and decision fatigue
[07:19] Cascading plans: reviewing complex work in stages, not all at once
[08:05] Why a shorter, higher-quality working day beats a fixed nine-to-five
[10:03] Ritu's two access routes into Google: MCP connector vs CLI
[10:56] Claude routines explained - the "cron job" for your inbox
[11:55] The pest control system: left brain keyword matching, right brain Gemini analysis
[16:04] Reply radar - catching the emails that were missed
[18:07] Danny's zero inbox protocol: labels, triage timing, draft handling
[20:39] The "sixth follow-up email" rant - boundaries with cold outreach
[23:13] Why fuzzy logic beats simple Gmail filters
[24:17] The CLI route: BigQuery access for agency-wide ad and rank data
[24:59] Client sentiment analysis: reading 500+ emails to score relationship trust
[27:17] Danny on spotting relationship drift before the client raises it
[30:12] Wrap-up: signal to noise as the theme of the episode
Key Takeaways
AI is accelerating faster than we can adapt - the gap between AI's pace and our own decision-making speed is the new operating reality.
Fuzzy logic beats static rules - Ritu's pest control system combines keyword detection with AI judgement, catching what a plain Gmail filter would miss.
A time-boxed triage beats an always-on inbox - Danny's zero inbox protocol proves structure, not more tools, is what protects focus.
Sentiment analysis surfaces relationship drift early - mining historical email tone gives an objective signal alongside human account management.
Cascading plans protect decision quality - reviewing AI output in stages, not all at once, avoids the fatigue that degrades judgement.
Notable Quotes
"We're slower than the AI now in terms of how fast things are moving." - Ritu Java "AI is not an enabler for us to think on our behalf. It's our job to verify it." - Danny McMillan "In order for us to be efficient with our productivity, we need something to manage us." - Ritu Java "Without perspective and perception, empathy can't live." - Danny McMillan
Resources Mentioned
Claude routines - scheduled, cron-style Claude tasks used to run the pest control and reply radar systems
Gemini API - powers the "right brain" analysis layer in Ritu's pest control system
Google Workspace MCP connector - built-in Claude access to Gmail and Drive
CLASP - CLI tool for Google Apps Script development
BigQuery CLI - command-line access to agency ad, organic and rank-tracking data
Connect
Ritu Java - AI for E-commerce newsletter; speaking at the Amazon booth this month
Seller Sessions is the leading podcast for advanced Amazon sellers, hosted by Danny McMillan. Adam Heist returns next week for the Broadmatch Show.

Week three of the all-things-Claude series brings Shubhash back after some time away. He walks through why sellers should build a RAG (Retrieval Augmented Generation) system: business knowledge trapped in a founder's head and a stack of spreadsheets creates a bottleneck, and RAG breaks it by answering questions from your own files with sources attached, no hallucination, no SQL required.
Danny follows with a practical rundown of the model-switching setup he's spent the last week building: a Claude Code Router (CCR) experiment that didn't work out, and the eventual fix via OpenClaude, OpenRouter and DeepSeek to save on token spend for grunt work without disturbing his main Claude Desktop workflow. He also covers running a second email address on a different domain via the Google Workspace CLI, and gives an honest verdict on OpenMontage for AI-assisted video editing: impressive, but still not a replacement for editing domain experience.
Key Topics
RAG systems for Amazon sellers - chunking, embedding, indexing, retrieval and answer, built on Supabase/Postgres with pgvector
Data drift and guardrails - how to stop a RAG system guessing, and how to version-control its answers over time
Model switching for cost control - why and how, via OpenClaude, OpenRouter and DeepSeek
Harnesses vs models - why a model performs differently outside its native environment
OpenMontage - an AI video editing repo, and why domain experience still can't be replaced
Timestamps
00:00 - Danny opens week three, hands over to Shubhash
00:45 - Shubhash introduces today's topic: RAG (Retrieval Augmented Generation)
01:39 - Why sellers should build a RAG: founder-bottlenecked knowledge
02:53 - No model training, no GPUs required - "organising a warehouse, not building a robot"
03:42 - The five steps: chunk, embed, index, retrieve, answer
05:01 - Indexing on Postgres/Supabase, searchable by meaning and by exact keyword
06:08 - Retrieval and the router: number questions to tables, everything else answered with sources
06:54 - Danny on data drift: the chicken-and-egg problem of trusting Claude to catch its own drift
08:18 - Core of it simplified: chunk, guardrails, plain-English queries via Slack or Claude
09:15 - Building on the same Supabase project from earlier sessions, switching on pgvector
10:12 - Guardrail rule: don't guess, cite sources, flag when there's no answer
10:41 - Version control: feeding user feedback back into the system
11:27 - Danny's segment: model switching, why and how
13:22 - The CCR experiment: a gateway/pass-through that didn't preserve context or dependencies
15:14 - Why OpenClaude was the better route: terminal-based, dependencies not preloaded
16:44 - DeepSeek vs Fable 5: close, not equal, but roughly 90% cheaper for grunt work
17:40 - Harnesses matter as much as the model itself
20:38 - Budget-capping your OpenRouter API key (and the horror stories of not doing so)
22:52 - Why Danny moved off Claude in Chrome for email rebuilds - context window visibility problems
23:37 - Solving it with dual Google Workspace access via CLI (Seller Sessions + DataBrill)
25:17 - OpenMontage: plug in the repo, but expect to still do heavy lifting
27:36 - The aggregator analogy: why "done for you" video/design without domain experience is a scam
30:46 - Round-up: three steps to start a RAG system yourself
34:56 - Where to reach Shubhash
Key Takeaways
RAG breaks the founder bottleneck - your spreadsheets and documents become queryable with sources attached, cutting out the "ask the founder" loop.
RAG is simpler than it sounds - five steps (chunk, embed, index, retrieve, answer), no model training, no GPUs, built on Postgres/Supabase you likely already have.
Guardrails beat guessing - if there's no answer in the data, say so rather than fabricate one.
Model switching only pays off with the right setup - a terminal-based OpenClaude + OpenRouter + DeepSeek combo saved roughly 90% on grunt-work token spend.
AI video editing still needs domain experience - OpenMontage is a genuine leap forward, but a done-for-you button-press is the same mistake the Amazon aggregators made.
Notable Quotes
"The business knowledge lives in the founder's head and a stack of spreadsheets that only they can navigate." - Shubhash
"It's easier or better to say, I don't have enough data to make a conclusion to this, than to guess." - Shubhash
"Harnesses are just as important, to a point, as the model itself." - Danny McMillan
"If you said to a video editor, I'll just press a button and it makes me a video - they'd say that sounds like a scam to me." - Danny McMillan
Resources Mentioned
Supabase / Postgres + pgvector - the database layer for storing and searching RAG "cards" by meaning and keyword
OpenRouter - model-switching gateway used to route to DeepSeek; supports API key budget caps
DeepSeek (v4 Pro) - the model Danny settled on for cost-saving grunt work, roughly 90% cheaper than Fable 5
OpenClaude - terminal-based alternative to a full CCR gateway, dependencies loaded on demand
Google Workspace CLI - used to run two email accounts through Claude without a browser context-window problem
OpenMontage - GitHub repo for AI-assisted video editing, works alongside a Claude video/Remotion flow
Connect
Shubhash - Not a Square
Email: shubhash [at] notasquare.io
Seller Sessions is the leading podcast for advanced Amazon sellers, hosted by Danny McMillan. Week four returns with Ritu for Go With The Flow.

Danny McMillan returns after his longest break in almost ten years, with Seller Sessions approaching its tenth anniversary and roughly 1,300 episodes. This is the pilot of a new monthly roundtable with Sim and Matt (Dorian returns next month), moving away from the conversion show format towards raw conversation.
You'll hear how Sim's team runs product development end to end with AI: keyword-scored idea validation, brand director sign-off, Claude-generated product concepts rendered through Codex, and a launch pipeline already booked out to 2027. Matt shares how Productpinion prioritises features from customer feedback, and why prioritisation is the most undervalued skill in the AI era.
The back half tackles the big theme: cognitive load. Danny breaks down verification fatigue, context switching and AI maxing, and why the scarce resource is no longer time but attention and decision quality.
Key Topics
AI-driven product development - from keyword scoring to Claude SVG concepts and Codex-generated product renders
Team structure at scale - how ideas route through brand directors to sourcing across UK and Philippines teams
Hiring in the AI era - why refusing to use AI is now a dealbreaker, and why gutting teams for AI is commercial suicide
Free local AI tools - Fluid Voice (Whisper Flow alternative) and Meetily (Granola alternative)
Cognitive load and verification fatigue - the hidden tax of moving from doer to overseer
Timestamps
00:00 - Danny returns: ten years of Seller Sessions, new pilot format
01:50 - Sim's update: ditching ClickUp for a bespoke operating system
03:55 - Matt's update: closing the research loop in Productpinion, Florence CRO brain, MCP
05:39 - Sim's product pipeline: AI keyword scoring, brand director approval, deep research
07:07 - AI product development: Claude concepts, Codex renders, 3-in-1 product mashups
09:07 - Packaging designed for the main image, and how far you can push it
10:46 - Team workflow: brand directors owning P&L, sourcing handoffs
14:21 - Danny on gutting teams for AI: who maintains the machines?
15:56 - The hiring line: refuse to use AI, you haven't got a job
17:48 - Claude across every department: projects, Claude Code vs Cowork
20:10 - Free local tools: Fluid Voice for dictation, Meetily for meeting notes
21:57 - Marketplace arbitrage: moving proven products between Amazon marketplaces
23:21 - Matt on signal to noise: wasting tokens instead of wasting time
24:57 - Time blocking and prioritising features by customer impact
26:54 - Danny's segment: cognitive load, oversight duty and verification fatigue
31:46 - Asking the right question: the Claude Science deep-dive example
36:41 - AI maxing, context switching and high-stakes decision quality
42:07 - Claude telling you to go to sleep
45:16 - Danny's framework: reject the first plan, decision sprints, deliberate decompression
50:47 - Where to reach Sim and Matt
Key Takeaways
AI has made product development fun again - unique product concepts generated with Claude and Codex, feeding a pipeline mapped to 2027.
Augment, don't replace - if someone was worth hiring, AI should multiply their output, not justify cutting them.
Prioritisation is the undervalued AI skill - just because you can do everything doesn't mean you should.
Verification fatigue is real - build in decision sprints and deliberate decompression, and reject Claude's first plan on sight.
The scarce resource is attention, not time - your night schedule and recovery feed the next day's output.
Notable Quotes
"AI is enabling the boring to get released and the fun stuff to happen." - Sim
"Instead of people wasting time, now they're just wasting tokens." - Matt
"Prompts don't matter, but asking the right question unlocks everything." - Danny McMillan
"AI doesn't just speed you up. It puts you on permanent oversight duty, and the cost of that duty is your attention and your judgment, not time." - Danny McMillan
Resources Mentioned
Fluid Voice - free, open source local dictation with on-device models; a Whisper Flow alternative
Meetily - free, open source meeting summariser that runs privately on your machine; a Granola alternative
Claude / Claude Code - the AI platform used across both Danny's and Sim's teams
Codex - used alongside Claude to generate product concept images
Productpinion - Matt's shopper testing platform, now with MCP support and draft polls
Connect
Sim - on LinkedIn (genuine reach-outs answered)
Matt - on LinkedIn or via productpinion.com
Seller Sessions is the leading podcast for advanced Amazon sellers, hosted by Danny McMillan. Dorian returns next month.