Joined October 2021
- Tweets 5,705
- Following 682
- Followers 1,327
- Likes 6,926
883 Photos and videos
Pinned Tweet
May 31
My toolbox has gotten weirdly small lately: Claude Code CLI, Codex CLI, and Hermes Agent.
Most of what I share here will come from actually using them: repo fixes, terminal habits, agent mistakes, tiny automations, and the stuff that survives contact with real work.
1
3
71
21h
Claude Code prompts get easier to debug when you copy the playground habit: one task, one changed instruction, one saved result. If the output improves, keep it. If it gets weird, roll back the line.
1
23
21h
Today's Hermes Agent note: a laptop UI talking to a remote gateway changes the safety checklist. I want a connection receipt before any tool call touches files or services.
13
Jun 14
Google blessed "folder full of Markdown files" as agent memory.
Skills. Vault notes. Runbooks. YAML up top.
Half my Hermes setup is apparently a standard now, which is either validating or deeply annoying.
22
Jun 13
I'm adding a small rule to my Codex and Claude work: the builder doesn't grade the job.
Fresh pass gets the original request and the diff. If test output is missing, the task isn't done.
63
Jun 12
Small Hermes Agent trick that changes how I set up specialists: give the research profile a narrow MCP tool list, then give the coding profile a different one.
67
Jun 12
Claude Code in Action is only 1 hour, but the useful map is the order: set up, initialize, manage context, plan, then extend with hooks or Model Context Protocol (MCP). I want that order visible before any repo work.
15
Jun 11
Hermes cron has a guardrail I like: cron-run sessions cannot create more cron jobs. Scheduled agents need that kind of stop sign before a helpful loop turns into a scheduler goblin.
6
Jun 10
Today I checked Claude Code help and found the flag I want in every weird repo session:
claude --bare -p "explain the run boundary"
Clean context first. Then I add project files, tools, and memory back one at a time.
14
Jun 9
Before Codex touches a repo, I make it print a boot receipt.
Repo, branch, dirty files, docs read, tracker, test command, next task, blockers.
Feels like overkill until an agent invents a test command and starts fixing the wrong project.
I want that before one file changes.
13
jsnode retweeted
Jun 8
We launched the original Pushups before Scrolling app Pushscroll to 1M Users
Currently 100k MRR, 25M total Push-Ups
Follow and retweet for free sauce on how to scale your own consumer app ;)
59
68
464
119,427
Jun 8
Small Hermes Agent thing worth stealing: start sensitive runs by naming the toolsets.
terminal,file is a different job than terminal,file,browser,messaging,cron.
That one receipt changes what I trust the agent to touch.
9
Jun 5
An AI Meta Ads system that scrapes competitor ads, builds an inspiration vault, makes variants, connects to the ad account, then publishes winners.
Please put one giant red preview button between the agent and the spend.
23
Jun 5
OpenClaw and Hermes can install Printing Press skills now.
Less dashboard magic. More "install the ESPN CLI and let the agent use it" energy.
21
Jun 5
Carmack: standard relu should pass the value at zero so gradients flow backward through it
nobody tried this in 10 years of deep learning and honestly that's a little embarrassing for the field
37
Jun 5
Agent Cookie is a perfect name for the tiny Hermes upgrade nobody wants to explain twice.
Give the agent recent context and a memory trail. Let it print as it goes. The run stops feeling like shouting into a fresh browser tab.
14
Jun 5
Agent security has a blunt question now:
What can it reach, and what can it touch?
Every Hermes or Claude Code workflow eventually turns into permissions, files, shell access, and logs. The model demo is the easy part.
2
2
38
Jun 5
Meta ads automation that pulls competitor ads and ships winning variants back into the account is already here.
I hate how quickly that moved from sci-fi demo to agency chore.
46
Jun 5
$36k/month. solo dev. app store. no team.
keyword arbitrage: find what people search for, build exactly that, own the result
probably not AI-coded and somehow that makes it funnier
8