Hand 80% of your work to AI. Pay for the result 

How I work: grandpa-style context engineering

Andy Avramenko 12 October 2026 2 min read 1 view

One folder, one CLAUDE.md, a fresh context for every task. How I kept that habit when my agents moved into one terminal: tmux and /clear.

I'm used to starting Claude Code from a particular folder that holds the necessary and sufficient context (CLAUDE.md + documentation). If I needed something from a neighbouring project, I asked for a prompt and simply pasted it into the next window.

Nobody has worked like this for a long time. Slava told me I'm a grandpa of the AI world. People at Anthropic laugh at me. But I still try to keep every iteration to a fresh context window (/clear) and to work inside its own project.

Then I decided to do my morning tasks by voice, away from the computer (the morning ritual, playing with the cat, matters more than all those AI agents of yours), and found I can no longer live the old way: the task list itself is shared, and the agents will have to sort out who gets to talk into my ear first...

The cat on the balcony reaching up at the vine leaves

I had to switch to working from a single terminal. But like every grandpa I'm stubborn: I don't want some abstract subagents being spun up, I want it to work the way it used to. From a specific folder, fresh, it reads CLAUDE.md and off it goes...

Agents have recently learnt to talk to each other. So at first I started a separate Claude Code session in each folder and asked the main session to do nothing itself, only pass commands on to them.

But that doesn't work the way I want: there's no way to pass on a /clear. And I don't trust auto-compact. So I moved all the sessions into tmux (a virtual terminal) and told the main session to start new sessions itself, send them tasks and, once they're done, type /clear or close them.

It turned out perfect. I can even tell the main session, by voice, to type /clear to itself. Just to be really, really safe.

I know it isn't a swarm of agents like the ones all the rich influencers run, but I'm happy, and we'll live with this setup for now. The next step is playing with the cat outdoors. There are a few nuances there that aren't as easy to build as they seem.

On Wednesday Slava Sazhin will show at a masterclass how he organises multi-agent development. For subscribers only (from €10 a month).

Was this article worth your time?

A new masterclass every week

On camera, a practitioner walks through a setup that works and hands over every file in it. $11 a month.

Latest recording · 30 September 2026 Claude apologised and promised never to do it again? Do not trust promises, set up hooks Slava Sazhin
All masterclasses

Contact the Combobulating team