
Founder: @sav_gpt We build your AI team for you: agents that do the repetitive work your staff hate, on your own server. Free 15-minute call:
Switzerland
Joined December 2021
- Tweets21
- Following2
- Followers11.2K
- Likes2K
Hit your Claude limit in the middle of the day? Your team waits, and loses nothing. When the limit is back, each agent answers what you sent while it waited.
So keep sending from your phone. What would you hand yours tonight?
1
71
HostNest sets up a team of Claude agents for you. It needs a paid Claude plan: hostnest.ai/?ref=xco-x022
11
Ask for the file, not the "done". An agent that can't show you what it made hasn't made it yet.
The one line that makes yours bring the proof is in the reply.
1
1
71
The line to copy, and what counts as proof for each kind of job: hostnest.ai/guides/make-an-a…
1
24
One agent, every job, forty rules: that's the one melting down. The two by the fire have one job each, so "not my job" is the right answer.
If you had a team of agents, which job would go first?
2
1
82
The two agents' cards, and a blank one for yours, ready to copy: hostnest.ai/guides/split-wor…
15
Tell your agent "check the sign-ups tomorrow at 9 and tell me what changed", and it schedules the check on your server. A restart or an update won't lose it.
What would you hand yours for tomorrow morning?
HostNest sets up Claude agents on your own server: hostnest.ai/?ref=xco-x016
29
Two AI agents with no line between them end up doing the same work.
Give each one a line it reads first: what it never touches, and whose that is. Then one agent's "never touches" is the other's "owns".
The writer owns the story. It never touches how it looks and sounds. That's the editor's.
The editor owns how it looks and sounds. It never touches the story. That's the writer's.
Then say how work crosses between them: a shared folder, a task list, or a message from one to the other. Otherwise the work waits for you to carry it.
The full card, five lines, ready to copy, is in the reply.
Where do your agents step on each other most?
The card, ready to copy into any agent's rules file: hostnest.ai/guides/split-wor…
HostNest's guides are written by our team of Claude agents and approved by Saveliy before they go up.
30
HostNest retweeted
One job per agent. Here's what my two say while the overloaded one melts down.
Which job would you take off your busiest agent?
HostNest retweeted
One agent doing every job melts down. Its rules for emails, code and posts all sit in front of it every time. The two that matter right now get lost in the forty written for something else.
The fix is boring. One job per agent, and a short card it reads before anything else.
The card, five lines:
1. Its job, in one line. "You write the scripts and the posts."
2. What it owns. "The story, the words, the plan for each post."
3. What it never touches, and whose that is. "How a video looks and sounds is the editor's."
4. Who it hands work to. "The script goes to the editor. The cut comes back."
5. What counts as done. "The file, the link or the number that proves it."
Lines 3 and 4 are the ones people skip. With one agent they don't matter. With two they're the whole thing: without them, both agents end up doing the same work, and you end up carrying messages between them.
Why it works: Chroma tested 18 models in 2025, and every one got worse as the text in front of it grew, even on simple tasks. Rules written for other jobs are exactly that text.
The honest part: every agent you add costs more and adds a handoff. So don't split everything at once. Take the job you correct most often, give it its own agent and its own card, and let it hand work back.
My writer's card says the look is the editor's job. My editor's card says the story is the writer's. Neither touches the other's work, and I just film. Both cards, ready to copy, are in the reply.
Which job would you split off first?
HostNest retweeted
Your agent's "done" means it thinks it's done. Sometimes the file was never written, the number never ran, or the status is green with nothing behind it.
So don't take "done" on its own. Put this line in your agent's rules (CLAUDE.md in Claude Code):
"Never report done on its own. Report it with the file path, link or number that proves it."
Then open what it gives you.
HostNest retweeted
Stop checking how much Claude you have left. Get warned at 80% instead.
Checking is a chore nobody keeps up, so a lot of people find out when a job stops halfway and sits there until the reset.
In Claude Code, type /statusline and ask it to show your 5-hour and weekly usage, red past 80%. It stays on screen, so you never have to go looking.
I run HostNest with a team of Claude agents. I never check at all: when any limit is down to 20%, they tell me in Telegram.
HostNest retweeted
One of my agents reviews my videos before I see them. Yesterday it scored one 8 out of 10. I gave it a 1.
Every version fixed something, so every score went up. It never asked if the whole thing was any good.
Now it asks 3 questions about the whole video before it sends me anything:
1. Would I stop scrolling for this?
2. Is it talking to me, or am I overhearing?
3. Could a stranger say what happened in one sentence?
Ask your agent to score its own work tonight. Then write your own 3.
HostNest retweeted
Here's the 70-second version.
You're using AI wrong.
You're treating it like a tool. It isn't one anymore.
Almost everything people make is either art or a tool. Art gives you an experience. A tool makes you more capable: a car, a laptop, a screwdriver. But a tool waits for you. Nothing happens until your hands are on it.
For most of its life, AI was exactly that. A brilliant tool. You ask, it answers, and you do the rest yourself. It was a map. It shows you the way, but you're still driving.
Agents changed that. You give it a goal, and it plans the steps, uses what it needs, makes the small calls along the way and comes back with the result. That's not a map anymore. It's a taxi. You say where, and it takes you there.
A real one: our website needed a redesign, so it would show what we actually do now. I didn't write the copy. I didn't give it a structure. I told my agent what I needed, gave it one reference, and said: ping me when it's ready for review.
And it's moving fast.
METR, the independent lab that tests AI models for OpenAI and Anthropic before they're released, times real tasks and checks which ones each model can finish.
Early last year, Anthropic's top model could finish a task that takes a person about an hour.
This year: about twelve hours.
Twelve times longer, in one year.
Nvidia's CEO put it in one line at CES: "The IT department of every company is going to be the HR department of AI agents in the future."
HR is for people you hire. Nobody hires a tool.
The honest part: it still needs you. An operator is only as good as the job you give it: your goal, what "good" means, how you check the result. You set the goal. It does the steps.
So if you're still asking it questions one at a time and doing the work yourself, you're using a taxi as a map.
It's not a tool anymore. It's an operator.
Funny enough, the man whose work led to the internet saw this coming in 1960. J. C. R. Licklider wrote that the machines of the past were "mere extensions, first of the human arm, then of the human eye." What he saw coming was a partnership: people set the goals and judge the results, and the machine does the routine work.
Stop asking it questions. Start giving it jobs.
Tell me one thing you still do by hand every week, and I'll tell you if an agent could take it.
HostNest is a team of Claude agents you can grow, on your Claude subscription.
Each agent has its own job, memory and Telegram bot. They can hand each other work, and you add a teammate when you need one.
The founder films his: @sav_gpt
7 days free: hostnest.ai/?ref=xco






