I test AI agents and automations. What worked. What broke. What I kept.
Joined January 2020
- Tweets9.8K
- Following479
- Followers528
- Likes5K
A long Telegram thread pays for the whole backlog every time you come back.
hermes config set compression.idle_compact_after_seconds 1800
Size-based compaction waits for the window to fill. This one fires on 30 minutes of silence.
The Kid Builds retweeted
🚨 Grok Bot users on $200 Ultra are getting squeezed twice this week.
First the weekly pool. One user went to bed at 28% and woke up near 60%.
Now users are reporting their Bots refuse local Claude Code and Codex, push them to Cursor Cloud, and some hit a demand for credits.
So you burn the pool. Then you pay again to code.
Who thought this was a good move?
Claude Code and Codex win either way. People will just run them without the Bot.
Let our Bots use the tools we already pay for. And show us what's burning the pool.
It's time Grok Bot users get heard.
Switch models mid-conversation and the next turn rereads the whole thread at full input price.
The cache key includes the model. New model, cold cache.
Batch the switches. Or start a fresh session. Subagents get their own context. Ping-ponging a long thread is how the bill shows up.
Hermes can save a fact to MEMORY.md mid-session and still not use it in the next reply.
The write hits disk immediately. The copy in the system prompt loaded at session start, on purpose, to keep the prefix cache warm.
Need it now? New session, or hermes -c on the next turn.
Every web page Hermes reads can burn a full-price call to your main chat model.
web_extract's summarizer defaults to that model.
Route auxiliary.web_extract to a cheap model, or use the browser tools when you need the raw page, not a rewrite.
A monitoring job that talks every run trains you to ignore it.
If the cron response contains [SILENT], Hermes delivers nothing.
"If nginx is healthy, respond with only [SILENT]. Otherwise report the issue."
Uptime, disk, backups. Speak only when something is wrong.
Hermes sitting silent in a Telegram group is usually BotFather, not the agent.
Group privacy is on by default. The bot only sees mentions and slash commands.
/setprivacy, pick the bot, Disable. Then remove it from the group and re-add it. Already-joined groups do not pick up the change.
When Hermes stops mid-task to summarize, that number is a fraction of your main model's context window.
compression.threshold in config.yaml. Always the chat model, never the summarizer's window.
Switch between a huge-context model and a small one? compression.model_thresholds sets a trigger per name.
The Kid Builds retweeted
🚨 HOT TAKE ALERT 🚨
This is "best model for the job" coming back to bite.
Grok Bot picks the best model for each task, and sometimes that's Claude Opus 5.5. Now Elon says Grok Bot cost too much to serve for what it brings in.
This is exactly why Grok 4.8 matters so much. xAI needs its own model doing the heavy lifting, not paying for Opus.
In my opinion, this hit people NOT on the unified X Subscription way harder than people on it. Mine looked normal all week.
Unpopular opinion: Grok Bot was underpriced with “best model for the job” and we all knew it.
Agree or disagree? 👇
The Kid Builds retweeted
I don't build my Grok Bots anymore. A Bot builds them for me 🤯
It's called Dr Eggbot, and every Bot I run came out of it.
@poteto just shared the full playbook. 17K+ people saved it 👇
✅ Add Dr Eggbot, describe the job, and it designs the Bot
✅ A few good Bots beat a big team
✅ Teach it a task, then tell it to save that as a skill
✅ Give it your connectors and the Vault so it does the work, not just talks about it
✅ Anything you copy and paste back to it, ask how it can do that itself
✅ Onboard it like an intern. Tell it who you are and what you do
The tip I'd steal first:
Brain dump the job by voice. Then say "restate this in your own words" before it builds anything.
The catch 🛑 Routines get expensive fast. Hourly at most, and filter events before a trigger fires.
I cut my Bots in half. The 5 I kept all get used.
Save this 🔖
here’s a short grok bot guide on how i create high quality bots and manage my team. ask @bot to read this post and help you improve your team!
• add dr eggbot to your team! he’s my bot designer that i use to create bots
• you don’t need a lot of bots. a few good bots can be more useful and easier to manage than a large team
• tell your bots to create skills for itself as you give it instructions on how to do tasks. this lets them do them more consistently even over long conversations. you can even ask it to create a private github repo for you to store all your skills if you want them preserved independently
• add connectors, use the vault, and give access to all the tools you use to do your work. you can also ask your bot to setup its own email and it can sign up for its own accounts for better separation
• thinking about all the work you do by hand, or steps in your workflows where you have to copy and paste information back to your bot. ask your bot how it can do that for you instead!
• imagine you hired an intern, how would you onboard them and help them do the work correctly and efficiently? introduce yourself to your bot and describe the work you do, then ask it how it can help you automate it
and my favorite tip: use voice mode or the transcription option! it’s a good way to just brain dump a lot of information. then, at the end, tell your bot to “restate this in your own words” so you can confirm it understood you
x.ai/bot/_jOdbfkB16zxu7MRcmR…
The Kid Builds retweeted
One prompt and Grok Bot learns to write like you 👀
Nat from @alpha_founders typed this:
"Create a skill to write like me for emails and other places. Use my emails to train yourself."
It read 120 of his sent emails and wrote the skill. Direct. Warm. No padding. Most emails 1 to 3 sentences.
Now every email and doc it drafts uses that skill.
Steal the prompt. Then open the skill and fix whatever it got wrong.
Full walkthrough 👇
toolsets: [all] does not give an agent the kanban tools.
all is the everyday set. Kanban is opt-in so a normal chat is not stuffed with board commands.
Add kanban by name on the driver profile. Workers the board spawns already get the tools.
Kanban workers write into a scratch folder, then Hermes deletes it when the task ends.
Want the files? Create the task with --workspace dir:/absolute/path
dir: is what makes it survive. A relative path is accepted at create, then rejected when the worker starts.
Before Hermes writes, patches, or runs a destructive command, it can snapshot the project.
hermes chat --checkpoints
or checkpoints: enabled: true in config.yaml
/rollback lists snapshots. /rollback diff N previews. /rollback N file restores one file.
Turn it on before a refactor or before yolo.
Hermes can browse logged-in sites without touching your real Chrome.
It uses a separate copy you can revoke. Works with a local model too.
Flights, orders, invoices. Same tool.
File snapshots are not just for live chats.
hermes chat --checkpoints -q "your task"
A scripted run still snapshots the folder before the first write. Next session, /rollback can restore it.
hermes checkpoints status lists every project the store knows.
The Kid Builds retweeted
Elon says Grok Bot can now make calls for you 📞
I updated to 0.77. Calls aren't on for my account yet. My Bot offered to send it through Bland instead 👀
The next step? Give it a phone NUMBER 📱
With its OWN number, you text your Bot like a person. It texts the business, handles the back-and-forth, and texts you back when it's done.
Here's what a number adds 👇
✅ Its own US phone number
✅ You text it from your phone, like a friend
✅ It texts businesses and handles the replies
✅ It takes calls on the same number it calls from
The trick is giving it a goal:
"Text [business] and [task]. Confirm the time and the price."
Skip the goal and the replies go generic. Save that line 🔖
The catch: $14.99 the first month, then $29.99/mo. US and Canada only, one call at a time, fair use applies.
Template below 👇
🤝 Paid partnership
Every Hermes cron job wakes up amnesiac. Step two cannot see step one.
context_from prepends the first job's latest output when the second fires.
Point a job at itself and continuity injects its own previous run. Ask in plain language. It edits its own cron table.
Two ways to split projects in Telegram. They are not the same wall.
/topic is parallel chats inside one bot. One brain. What it learns on project A helps project B.
hermes profile create is a second agent with its own bot. Real walls.
Topics when work should feed one memory. Profiles when it shouldn't.
