Stack · Published 2026-06-23 · 2 min read
How I work with neural networks. Part 1 of 4
I run five client ad accounts. Predictions, creatives, daily reports, campaign decisions — Claude Code does most of that work.
I spend 90% of my own time in Telegram
Inside Telegram I:
- receive reports
- build predictions and judge performance
- launch campaigns
- upload creatives to the ad account, which creates the campaigns and ad sets straight away
- generate new creatives
- and brainstorm in the spirit of "so what do you reckon?"
The next three posts go into detail — predictions, creatives, the daily report. This one is about the infrastructure as a whole.
A separate server
Claude Code runs on its own Windows VPS, around the clock. I would rather work on some flavour of Ubuntu, but MCP servers such as Pipeboard and Adjust need an open browser at the moment of authorisation, so for now Windows is the simplest place to host Claude.
Scripts run on a schedule: reports for every client are assembled at 9 in the morning.
One Google Drive holding every client
Google Drive is installed on the server and synced with every client. Creatives, exports, configs — everything sits in one file system that Claude can reach directly. At the start we agree with clients, designers and editors to work through Google Drive and nothing else. The moment a new creative lands in a folder it is available to the system, with no forwarding and no manual upload.
Access to the accounts and the tools
Claude has access to every ad account (Meta and Google Ads through Pipeboard) and to Adjust, Adapty and RevenueCat through MCP and API. A single process sees the spend from the ad account, the attribution and the revenue from the revenue tool at the same time.
Skills
What is convenient is that experience gained on one project spreads to all the others through skills. If a creative uploaded badly on one project, it does not happen again on another. A prediction carefully worked out on one project improves the predictions everywhere.
Security
Each client has access only to their own folder in Google Drive. Nobody but me has access to the bot or to the remote machine. Cross-project leakage is not possible.
Telegram
Claude on the remote Windows box is tunnelled into Telegram, so Telegram is effectively the interface for talking to Claude.
Next post: predictions on Mondays, in detail.
If you want to assemble the same workflow yourself, setting it up from scratch is exactly what I cover in the AI module of User Acquisition — from getting the tokens to the first campaign launched through Claude Code. The module can be bought on its own.
Questions by direct message: @SiberianPaaanda