I use Claude every day for this newsletter and for client work, and until I sat down and actually mapped it out, I couldn’t have told you which parts of that were genuinely load-bearing versus which ones I’d tried once, thought “huh, neat,” and never opened again. So here’s everything that’s earned a repeat spot in an actual marketing week, and what I stopped bothering with.
The workhorses
Spreadsheets, from a description or a messy pile of numbers. Budget trackers, campaign trackers, a media plan across five markets. Describe what you need or paste in the mess, and you get back a working file, not just a screenshot of one. No more rebuilding the same tracker template for the fourth campaign in a row.
Getting it out as a file, in whatever shape it needs to be. Word doc, PowerPoint, a filled-in PDF, a merged or watermarked one — the output most marketing tasks actually requires isn’t a chat reply, it’s a file someone who doesn’t use Claude can open.
Research with a confidence check, not just an answer. Ask it to research something and rate how confident it actually is in what it found, rather than just handing you a flat, confident-sounding answer. It doesn’t stop hallucination, nothing does, but it gives you a place to start fact-checking instead of starting from zero.
Making it ask before it guesses. Underrated. Before you hand over a brief, ask it to ask you clarifying questions first instead of guessing and running with the first interpretation. Five minutes of back-and-forth upfront against a rewrite later makes a huge difference.
Reading your inbox and calendar so you don’t have to. The unglamorous one. Half of campaign ops is “what did the client actually agree to in that email thread” and “when is this actually due.” Having Claude read the thread instead of you re-reading it adds up across a week faster than anything flashier.
The regional wins nobody mentions
Multi-currency calculators, if you work across markets — an SGD/THB event budget converter or a rate card you can hand to someone else is a real unlock, it just doesn’t sound like one unless your work happens to be regional.
Visualisations: timelines, process maps, run-of-show — for the specific case of a roadmap a client needs to see, not just read. Not weekly tools for me, but nothing else in my stack produces either this fast when I need them.
The two nobody talks about
Feeding it your own material, not just the open web. Web research gets all the attention. Working from your material is a different move: past newsletter issues, a stack of call transcripts, a competitor’s deck, fed in so it’s synthesising against your own archive and matching your own voice instead of researching from scratch.
Building something that gets reopened, not sent once. The bigger version of “build a tool” isn’t a one-off calculator, it’s something a client or teammate reopens and updates themselves: a living tracker, a one-pager, a small tool, rather than a file you rebuild from scratch next time. This is the same shift that changed how I run this newsletter — giving it a persistent project with the running context of what I’m working on, instead of re-explaining the brand and six weeks of decisions every time I open a new chat.
The current limitations of LLMs
Claude, like all other LLMs (be it GPT, Gemini, Grok, DeepSeek), is still a language model and it will still say wrong things with total confidence if you let it. Confidence ratings help; they don’t solve it.
I fact-check every number that goes in front of you here, and I’d do the same with anything client-facing that comes out of a spreadsheet or deck it builds for you, especially figures, dates, and anything with a name attached.
Reply and tell me: what’s the one Claude habit that’s actually changed how you work, versus the one you tried dropped?
The AI;DR
Elsewhere in the AIverse
ChatGPT’s Chrome extension can now act across your open tabs. In regular Chat mode, ChatGPT can read and compare what’s open across your tabs without you copying and pasting between them. In ChatGPT Work, with permission, it goes further — filling forms, sending messages, downloading files, organising tabs, and using information from one page to complete a task on another before stopping for your review. Separately, GPT-6 Pro users got faster responses and the ability to split a task across multiple agents working in parallel, with visibility into their progress.
Anthropic’s IPO filing shows a $42B loss. The IPO is now targeted for mid-November, pushed back from October, with some reporting the $2 trillion valuation talk slipping toward $1.5 trillion after a Q3 growth slowdown. 2025’s net loss was about $42 billion against $4.6 billion in revenue, mostly a non-cash charge tied to convertible financing. But Q3 2026 revenue hit roughly $16.7 billion, up 50% on Q2, and the company says it’s now adjusted-profitable and cash-flow positive for the second straight quarter. OpenAI, meanwhile, is reportedly chasing its own $30B+ bridge round at a $1.4 trillion valuation — which, per Bloomberg, pushes its IPO to 2027.
Krea 3 lands in second place on a major image-editing benchmark. Krea’s new model edits images in about 8 seconds for roughly $0.06 a run, and ranks second on CPI-Bench — an Alibaba-run benchmark for real-world editing tasks — behind only OpenAI’s GPT Image 2.5.
Separately, ElevenLabs shipped Eleven v4 and v4 Turbo, text-to-speech models tuned for emotional delivery, with Turbo built for real-time voice agents and live calls at roughly 100–150ms latency.