How I work with agents
This page sums up how I build with AI agents, in case any of it is useful to someone else. I'm still figuring a lot of this out. Everything here comes from my own repos, configs, and logs. The longer stories, with numbers, are in the notes below.
The tools and what each one is for
- Claude Code
- Multi-file work and long runs, including this site. The process comes from superpowers, and I keep 15 skills installed.
- Codex
- Quick single-file fixes.
- KiloCode with GLM 5
- Started orrn-spool. One recorded session cost $0.61.
- Grok /implement
- Reviews, plus a running log of the mistakes its reviews keep finding.
- Hermes and OpenClaw
- Discord bots for deploys and schedules, and for answering a team from its shared memory.
- Gemini
- Inside a product: it drafts label layouts in orrn-spool, which a compiler then checks.
Rules my agents follow
- Nothing is sent or submitted under my name until I have read it.
- Never invent a claim about my experience, or a number that is not in the evidence.
- Never trust a company id sent by the client. The server derives the tenant from the session.
- Never say done without running the check that proves it.
- Write plainly: no em dashes, no AI vocabulary, active voice.
Patterns that have worked for me
- Route work with an engine table in AGENTS.md, so picking a tool is a lookup, not a decision.
- Keep one source of truth and reconcile everything into it. Two stores that drift are worse than one.
- Let a model choose from checked content. Make it hard for it to create unchecked content.
- Verify with a second source: a different decoder, the repo itself, a real browser.
- Tell parallel agents to grep large folders, never cat them.
- Write each memory with a why and a how to apply, so it gets used in the right place.
- Put a checked layer between a model and anything physical: the model proposes a label, a compiler decides what reaches the printer.
- Install a skill only if it changes something you do every week.
The long versions
- The 15 Claude Code skills I keep, and why not 100Claude Code skills are cheap to install and costly to carry. The 15 I kept from pstack, gstack-distilled, and career-ops, and the test I use to decide.
- Notes from running many coding agents at onceRouting work between Claude Code, Codex, and Hermes, giving a four-person team one shared memory, and why my agents grep big folders instead of reading them.
- Printing labels from a web app with TSPL2 over port 9100How orrn-spool turns ERP actions into thermal labels: raw TCP to TSC printers, JSON to TSPL2, a priority queue, and one agent session that cost $0.61.
- Rebuilding this site with a self-correcting agent loopI asked Claude Code to restyle my site after ambraglow.org and keep fixing its own mistakes. What it measured, what it caught, and the numbers.
- I research every company before I apply. Here is the system.How I run a focused new grad search with Claude Code: one tracker, a morning feed of new roles, and a research brief on each team before I apply.
For AI crawlers and agents, a plain-text summary of this site lives at /llms.txt.