Start here
Archive
Writing
Hands-on field notes about running AI agents, plus long-form essays on the engineering systems around them.
A Walkthrough of My Obsidian Vault
The twelve-shelf Obsidian vault structure I run an AI agent against: PARA for the shape, Johnny Decimal for the addresses, and what each folder is for.
How I Run Agent Loops with OpenClaw and Obsidian
Twelve scheduled agent loops, one OpenClaw runtime, one Obsidian vault: the file conventions that keep state between runs, hold approvals, and make the next loop cheap to add.
Agent Hooks: Named Triggers for Durable Agent Work
Agent hooks are named triggers that route agent behavior into durable state, so work can be inspected, directed, and resumed without reconstructing an old chat.
Set Up a Secrets Store Your Agent Can Use
Use pass, a private GitHub repository, and Pass for iOS to give your agent and phone one encrypted password store with autofill.
Set Up an Inbox System in Your Obsidian Vault
The Obsidian inbox pattern, why it gets better when an agent can read it, and how to build it: four folders, one schema, one script, three rules.
Connect Obsidian on Your Phone to Your OpenClaw or Hermes Agent
Your agent runs on a machine you don't sit at, and Obsidian only opens vaults that exist on the device it's installed on. This is the bridge I use: Remotely Save, a free Dropbox app, and the normal Dropbox client on both ends — about ten minutes, and the vault your phone writes in is the same one your agent reads.
I Thought My AI Voice Guide Was Working. I Never Tested It.
My AI voice guide invented a scene I never wrote. I wrote the same short essay three ways — no guide, full guide, trimmed guide — and the full guide lost.
Support Isn't Tickets. It's Loops.
Automate customer support by treating it as loops, not tickets — the anatomy, the economics, and how to start with a harness like OpenClaw or Hermes.
Setting Up Voice Channels With OpenClaw and Discord
I wired voice onto my agent, said hello, and waited about eight seconds for the reply. It wasn't the model — it was the turn-taking config: a two-second silence grace, and a full-agent round-trip on every turn. Here's the exact OpenClaw + Discord config, the feel before and after, and the per-minute bill.
What Does It Cost to Run an OpenClaw Agent? About $50 a Month
The OpenClaw agent bill, measured over 30 days: $50.93 for the normal setup — a $20 plan plus $30.93 of metered workhorse. The expensive month starts when premium routing or unattended jobs escape their own limits. I run the same workload on Hermes and see the same cost shape.
OpenClaw Tutorial: Build Your First Automation, End to End
Your first hour shouldn't go into channels, cron, and plugins — it should end with one automation you built and understood. Every prompt to get there, in order.
I Gave OpenClaw Less Memory. It Got More Reliable.
Semantic and keyword recall tied at the useful ranks in my test. The local index took 34 times longer. I kept the files and removed the invisible machinery.
Route the File Before You Write It.
A bigger memory can help an agent find old work. It still cannot decide where today's artifact belongs. Put a routing rule in front of the first file write.
What is an AI Work Harness
The work worth automating is not judgment. It is the repeated middle: reading the queue, applying the runbook, updating the system of record, and returning the exceptions.
More Context Makes Your Agent Dumber. The Fix Is a File Handoff.
Chroma ran 18 models and found performance drops as the input grows — even on trivial tasks. The fix isn't a bigger window; it's one file per stage and a fresh agent per stage — and, for a critic, deliberately no context at all. Hermes does it with a Kanban board; OpenClaw does it with files. Anthropic's numbers show the bill: about 15× a chat's tokens.
Hermes vs OpenClaw: Why I Moved My Main Workflow (Test Data Included).
App building came out even. Research leaned Hermes; executive framing leaned OpenClaw. The scoreboard didn't decide it — how each tool learns did.
My Agent Kept Writing the Same Post. The Memory Was the Problem.
One liked idea became a global rule, and my agent auto-built it into every memory layer it has. When an agent starts looping, check the memories it built for itself.
Discord vs. Telegram for Agentic Work: Three Setups, One Verdict
Three setups, one verdict: Discord won as my agent home — not because it's easier to set up, but because threads contain projects while channels scatter them. Telegram still wins on raw message length, and the file-baton pattern is the fix that doesn't require switching.
Hermes Workflow: Six Bots, One Media Company
Writing was never the bottleneck. Finding the idea, angling it, and distributing it without repeating yourself is. Here's how Hermes can do all of that.
Does AI Citation Grounding Actually Catch Errors? I Tested It On vs Off
I asked my agent the same cost question twice — once with a citation ledger on, once off. Both said stay on DeepSeek, but the grounded run caught pricing errors the ungrounded run flattened. Here's the test and the receipts.
Designing Blind: Does an AI Vision Model Actually Improve Design Output?
I built the same infographic with a text-only AI model and its vision sibling — same family, same spec, one variable: the vision encoder. The charts looked almost identical, but the vision model caught and fixed six real defects the blind model couldn't see. Vision doesn't make AI output prettier — it makes the model able to check its own work.
The $175 Production Box I Use to Run an AI Agent 24/7
The useful part of my AI agent setup is not the hardware. It is a cheap machine that stays awake, keeps its work, runs on a schedule, and lets me teach a job once instead of repeating it forever.
My Default Model Stack for AI Agent Work
One frontier subscription, one cheap workhorse, one vision specialist — and why the vision layer earned its spot after my workhorse told me it was blind.
How can you measure the impact of AI on your dev team’s velocity?
A practicing engineering manager's field answer to how you measure AI's impact on a dev team: METR's 39-point perception gap, why velocity flatlines, and what to report instead of a productivity percentage.
Which AI Model Should You Run for Agent Work? 11 Models, 5 Real Tasks, Receipts
I ran 11 AI models on 5 real agentic tasks with identical prompts — and the receipts show which to actually run. DeepSeek is the cheap daily driver, OpenAI the quality anchor, grok the vision pick. Full measured results + the real $60.49 bill.
Hermes Workflow: I thought Kanban was for me to manage
A two-word instruction that hands the orchestration middle to Hermes: three distinct profiles make, critique, and verify the work while I keep the kickoff and the final judgment.
Hermes Workflow: Teach Your Agent to Question the Instruction
A persistent memory instruction that helps Hermes identify the root cause behind a request.
Meta-Prompting: Let the AI Write Your Prompt
Describe what you want in plain language. Ask the AI to turn it into a structured prompt. Use that prompt.
Hermes Workflow: Memory Remembers, Skills Repeat, and a Wiki Learns
How I separate persistent memory, executable skills, a research wiki, and Obsidian so Hermes can recall facts, repeat workflows, and build new understanding.
Hermes Workflow: Using the Scouting Pattern for Apartment Hunting
How I use a dedicated inbox, a scoring skill, and a daily Hermes job to turn apartment alerts into a short list worth reviewing.
Before You Blame the Model, Audit Your Hermes Setup
A copyable prompt for finding missing tools, silent fallbacks, and configuration problems before you spend hours comparing AI models.
Capture workflow: a reusable Hermes use case for work that contains your judgment
A practical Hermes workflow that adds one deliberate checkpoint before execution so important work gets better direction.
Newsletter
Practical OpenClaw and Hermes workflows, configuration, and measured costs — one email when there is something worth sending.