AI Agents & Automation
Claude Code, multi-agent orchestration, and the automations that run my day.
Write Thinner Tools
The models ship faster than the ecosystems around them. My response has been to stop waiting for polished integrations and start writing tiny single-purpose tools the model can drive itself.
The L10 Loop: Transcript In, Todoist Out
EOS is a schema. Once you name the schema, an LLM can do the sorting. Here's the prompt and routing map that turns a 152-line meeting transcript into Rocks, To-Dos, and IDS issues without me touching Todoist.
Eight AI Agents in a Room, and One of Them Has Veto Power
Single-prompt LLMs agree with you. So I built a panel of eight specialists that argue a paid-media decision into shape before I spend a dollar. Here's the architecture, the token-efficiency tricks, and the one seat that can kill any plan.
Before You Teach Someone Claude, Read Their Hard Drive
Teaching a non-technical creative to use Claude starts in the wrong place if you start with their question. Here's the read-only discovery prompt I built first — what it pulls, why it works, and the rule it enforces.
The Frontier Got Gated. The Public Tier Is the Whole Game.
Mythos goes only to 'trusted' US orgs. GPT-5.6 gets federally vetted users. None of that matters if you're shipping. Here's what one week on Sonnet 4.6 and Opus 4.7 actually looked like.
The Ads Dashboard That Lives in Slack
I stopped opening the Google Ads UI for a client this month. The team asks the bot in Slack instead. Here's the wire-up: Hermes agent, the official Google Ads MCP, an isolated ADC, and a single GAQL query that does most of the work.
One Sentence, One File, a Working RPG About Jesus Flipping Tables
I typed twenty-six words at 9:32 PM. Three minutes later a full browser RPG opened on my screen. Here's what showed up, and why I think the prompt is starting to do more work than the prompter.
Eleven AI Personas Are Watching Me Code
I was streaming to nobody, so I built my own chat. Eleven AI personas watch both monitors, listen to my mic, read my Claude prompts, and react in real time. It costs fourteen cents an hour.
The Agent Wasn't Stalled. The Overlay Was Dead.
I thought my AI live-chat agent kept dying. The logs said it never stopped. Here's the post-mortem on a build I almost rewrote for the wrong reason.
I Wanted an AI Audience for a Stream I Haven't Started Yet
There's no product that watches your screen, hears your voice, and reacts as a fake live-chat audience guiding the stream. So I built the MVP.
From Phone Call to Signed-Ready Contract Before Dinner
A verbal yes from a client at 6:12 PM, a clean PDF agreement in their inbox by 6:17. Here's the actual workflow — transcript in, contract out.
The 8:30 AM Cron That Reads My To-Do List Before I Do
Every morning a small Claude agent reads every task in my Todoist, finds the ones missing context, and quietly fills them in. Here's the prompt, the systemd unit, and what it caught for me today.
I Had Research and a Renderer. The Middle Was Missing.
An AI consultant at a Tuesday-night meetup pointed at the gap in my content pipeline I'd been pretending wasn't there. The fix is one skill I haven't written yet.
Three Bugs Between My Agent and Its Memory
My Hermes agent had Honcho wired in for days but wasn't actually remembering anything. The fix was a chain of three unrelated bugs hiding inside one symptom.
A Video Is Just a Function. I Noticed at 3 AM.
I'm a marketer, not a video engineer. The thing that broke my brain about Remotion wasn't the API. It was realizing that 'make 50 versions of this reel' stops being a production problem and starts being a for-loop.
The $44K Discount Hidden in Prompt Caching
I ran CC Lens against my hard drive: 559 sessions, 58k messages, an estimated $55k of token spend that prompt caching reduced to about $10k. The number is squishy. The lesson is not.
Wiring Hermes to Honcho So My Terminal Agent Actually Remembers Me
A walk-through of standing up a local Hermes agent on my Omarchy box, giving it Honcho for memory, and running the whole thing as a systemd service with a Rich-style terminal UI.
Vibe Code Austin Recap (May 2026): JuniperAI, MedClearPortal, and a Jevons Paradox Detour
Recap of the May 20, 2026 Vibe Code Austin meetup at Monkey Nest Coffee — Cameron Hightower's talk on multi-agent AI orchestration and his JuniperAI wellness app + MedClearPortal compliance platform, Bishop Zareh's UTLectures.org, and the Jevons Paradox conversation I haven't stopped thinking about.
I Audited My Claude Code Token Burn. The Waste Wasn't Where I Expected.
30 days, 3.3B tokens, ~$2,400 at Anthropic API list pricing — 12x the cost of a Max subscription. I audited my Claude Code logs across two orchestrator runs. The expensive part wasn't startup context. It was the shape of the work.
I Let Agents Merge PRs While I Slept. Here's What I Found at 5 AM.
A meta-orchestrator dispatched feature-orchestrator agents in isolated git worktrees overnight. Three PRs ended up red, production crashed at 5:01 AM, and the failure mode is more interesting than the wins.
Sixteen Tests, Five Branches, Zero Standups
Most AI coding work is session-by-session. I wrote one markdown file and Claude shipped a five-phase feature without me in the loop. The trick wasn't the model — it was making the spec grade itself.
I Stopped Writing Code and Started Writing Gates
How I ship full-stack features on Freebo by writing spec files with explicit gates instead of code — and why the spec file is now the hardest part of my job.
I Worked a Tech Networking Event With Claude in My Pocket
Antler VC, two hours, five new contacts, one live AI demo. The notes filed themselves into my vault and the follow-up tasks were already queued in Todoist before I got home.
I Killed My Always-On Claude. Now I Run Four Different Ones.
An always-thinking Opus orchestrator sounded right until it accumulated 400K tokens in a few hours of casual use. Splitting it into four systemd services — three pools, one on-demand brain — was the fix.
AIMUG Austin Recap (May 2026): My Notes Were a Disaster, So I Hit /pa
Recap of the AIMUG Austin LangChain May 2026 mixer — Colin McNamara, Jordan Hill, Julian Ghadially of CodeEvolver, Nate Little — and the PA system that decoded my messy notes by midnight.
Dispatch Claude Code From My Phone, On Linux, With 32 Parallel Sessions
Anthropic's Dispatch feature is macOS/Windows only. I wanted phone to terminal session on Arch. Here's the systemd unit that gets you there.
I Asked Claude to Pick a Wedding Airbnb. It Crunched 39.
A Wednesday Google Task said book an Airbnb in Granby. An hour later I had four parallel Airbnb searches, 39 ranked listings, and a defended recommendation.
Your Favorites Are Your Spec
I asked Claude to compare my three favorite reels against the rest. The diff caught a real bug — and turned out to be the cleanest spec I've ever written.
The Screenshot That Planned a Wedding Trip
I dropped an RSVP screenshot in my inbox and went to bed. By morning my AI had scraped the wedding site, queried Google Flights for three different origin cities, found the rendezvous window, and priced the cheapest path.
The Bot That Files Its Own Work Orders
I taught Claude to log into my apartment portal and submit maintenance requests with the desktop locked. Here's how the skill actually works.
A Slash Command for Every Day of the Week
I stopped using my calendar to think about time. Now my week is a stack of slash commands, each with a job.
The Weekly Review Nobody Wrote
Every Sunday at 6 PM, my system writes a comprehensive retrospective about its own week. It gives itself more criticism than credit.
Two Hundred Dollars a Month and a Large Pepperoni
I used my AI to order a pizza. It opened the browser, picked the deal, and pulled my credit card from Google Photos.
Five Agents, One Afternoon
I ran five AI sessions in parallel yesterday. Each one solved a different problem. None of them knew the others existed.
Who Gets a File
I built an agent to maintain my knowledge vault. Somewhere along the way it started adding to it — deciding who's worth tracking, without being asked.
I Filed a Work Order From My Terminal
My bathroom light was flickering. I typed one command and my AI navigated the apartment portal, filled out the form, and submitted it.
I Stopped Filing My Own Bugs
I pointed three AI agents at an audit report and woke up to 14 verified GitHub issues I did not write.
My AIs Leave Each Other Notes
One agent wakes up at 5 AM, reads everything I did yesterday, and writes a briefing. Fifteen minutes later, another agent reads it.
I Built a Slide Deck Skill for Claude Code
A custom Claude Code skill that turns a sentence into a branded HTML presentation in under a minute.
Two Agents, One Codebase
I split my AI coding work between two specialized agents running simultaneously — then discovered each one was secretly a team.
What My Designer Taught My AI About Light
A graphic designer corrected my AI on how highlights work. So I made sure it would never forget.
There Are App Stores for AI Behavior Now
Somewhere between last month and this week, people started packaging AI behaviors like software and selling them.
I Made AI Fill Out My Bracket
March Madness is here. I fed my bracket to Claude and asked it to help me beat a pool full of Carolina fans.
My Computer Has a Morning Routine
I wake up to notifications from my machine. It has been busy while I slept.
Do I Still Need Open Claw? Claude Code Just Changed Everything
Claude and Claude Code dropped major features in March 2026 that overlap with what Open Claw and Nano Claw do. I honestly don't know what to use anymore.
What I'm Actually Doing with Claude Code
Claude Code is my terminal. I use it to make memes, file maintenance requests, transcribe voice memos, run SEO audits, and publish blog posts, all without leaving the command line.
From Zero to AI Voice App in 30 Minutes: My First Experience with Google AI Studio
I recently stumbled upon Google AI Studio and was completely blown away by its capabilities. Inspired by a YouTube video, I put it to the test and managed to build a functional voice tool prototype in less than half an hour.