@BurnAlgoDown@XFreeze@bot You misunderstand my point. I use AI in pretty much every single business process, but I would use it very clearly, with a clear understanding of what it can and cannot do.
The AI in my partner program once invented a discount code. That was the last time it got anywhere near money.
Setup: pole dance studios get their own code and 10% commission. An agent handles the replies.
One studio's address didn't match the CRM. The model "solved" it by making up a code. A guard caught it and parked the case.
Now:
– studio shows interest → I create the code → a script sends it
– studios see their commission (earned, paid, open) in their own dashboard
– revenue syncs daily from Shopify, no LLM involved
– 45 studios had logged in by early September
AI writes the replies. Anything that touches money is a script 🤷♂️
ah crazy I I built my own skill (I call it session harvest) which looks at the chat, sees if it learned anything, and updates the memories and the files. It also lets me know what it changed, added, or updated. I don't really ever go above 300k input tokens, and it also just gets slower and the answer more "clouded"
@itsolelehmann@bot Anything against Meta's official policy gives you the risk of being banned. For me, it would be a complete disaster if my business WhatsApp will be completely wiped :)
There is literally an issue with every single update.
I have two bots running at 9.6 and one running at 9.9. I have been asking Claude Code if we should update, and it has said no every single time because there are always new issues that would potentially mess up my setup. I think at this point I should seriously just move to Grokbot.
Maybe I shouldn't be this ungrateful since it's free software, but yeah, it bothers me, to be honest.
My bookkeeping runs mostly automated. The AI is the smallest part of it 😅
Every morning a bot goes through my inbox and pushes invoices into my accounting software. Fixed order:
– blocklist
– must have a PDF
– list of 174 known suppliers
– PDF content
Only if all of that fails does an LLM get a say. If the model is down, the receipt is flagged "unchecked" instead of filed somewhere random.
On the go: receipt photo via Telegram, done.
Start of the month: the Shopify revenue receipt for the previous month gets created automatically, refunds included. Marking it paid stays my job.
The worst problems had nothing to do with AI:
– the accounting software added a new required field and the monthly close just failed
– the bank feed silently stopped because the provider behind it changed
– 18 of 42 "open" receipts were duplicates. Only showed up in the tax export
How much of your bookkeeping is still manual?
@jefri_chat No I think it's fine. Sonnet is so extremely good, all my subagents run by default on Sonnet. Also, the task isn't very difficult. It's just translating stuff based on the keywords we researched. Also, it's all of my Claude Max algo, so all good here :)
1.4 billion tokens. 64 PRs. Up to 10 subagents running at the same time.
That was my Sunday: international SEO strategy for two brands, from zero to live 😅
Italy, Poland, France, England. The question: what do people there actually search for? And what does a standard translation completely miss?
Internationally there's way less competition. Sure, we have no authority in those markets and most backlinks come from German-speaking sites. Still a pretty great experiment.
What blows my mind is the speed AI makes possible now. Strategy to live in a few hours.
But nobody pressed a button and watched everything magically get better.
How it actually ran:
– Strategy based on the data we had
– Concept, approved by me
– Execution split across multiple agents in parallel
– Quality check of the results
– A separate final QA pass over everything
A human at every decision point.
Best part: no client sign-off needed. The content stays the same, only the words change. So I shipped it myself.
Curious how this shows up in the rankings.
Agents are great at doing the work. I don't let them decide what the work is.
What should a shop do next? Which pages matter, what's worth building, what to skip? I decide that myself. For strategy I stopped using agents completely.
Then I hand the execution to agents in the terminal, very specific: upload these products, write this article, build this report, change this section in the theme. Across all my client shops.
September: ~210 sessions in 30 days. Plus 26 skills, fixed workflows that run the same way every time.
I think and decide, agents execute. Works way better than the other way around 🤷♂️
Image: AI-generated
@GermanIaRLP Großvater ist gestorben, und man wollte mir nicht mal einen freien Tag von der Arbeit geben man hat stattdessen gemeint, ich sollte einen Urlaubstag nehmen.
48 backlinks from real local businesses. Price per link: one pair of shorts 😅
Quick context: Google still ranks shops higher when relevant websites link to them. Those links are hard to get and usually expensive.
In the polewear partner program, studios already get a code and commission. On top there's one extra deal: link to the shop from your studio website or Instagram bio, get a free pair of shorts as a welcome gift.
The code never waits for the link. Only the gift does.
So far 48 studio websites link to the shop. 32 in Germany, 10 in France, 4 in Austria.
Local, relevant, from the places where the actual customers train. Try buying those.