Field Notes
from Production
Random thoughts, occasional insights, and lessons from breaking things where the stakes were real. 46 pieces; probably a few should have been tweets.
Featured
See which page ranks for each query with the Search Console API.
Search Console exports separate queries from pages, hiding which URL Google shows for each term. The API joins them. Here is the fifteen-minute setup.
A price is simple until the customer qualifies for it.
A guest can start with one offer, then qualify for another when the customer context becomes known. Commercial must define that change, and every touchpoint must keep the price true.
A checkout countdown has to earn its delivery promise.
A checkout countdown can create urgency or support a real delivery promise. That promise needs clear rules for what changes at zero, which date to show, and how it survives payment.
Codex vs Claude Code: two agents on one codebase
I run Claude Code and Codex daily, sometimes on the same repository. My session logs show the context cost, while the working tree shows the coordination cost.
The person logging in is not always the customer.
Login proves who is at the screen. It does not prove what they own, may manage, or can buy. Model those questions separately before the edge cases arrive.
What using Claude Code actually looks like
Using Claude Code is a loop: ask, change, check, decide. Here is what a normal session looks like, six commands worth learning first, and how to recover when it goes wrong.
Capturing email and calendar without a mailbox API
Microsoft Graph and IMAP were off the table, so I used Power Automate, OneDrive sync, and one 15-minute scheduled task to pull email and calendar into a local inbox.
Most of my AI pipeline isn't AI
By volume, most of my AI pipeline is plain Python. The model is a small, expensive part I call as little as possible. The script-first split, and why it wins.
Claude Code skills: building a workflow that runs itself
The control layer of my work system is markdown and small scripts. How a workflow combines Claude Code skills, rules, model routing, and parallel sub-agents.
The Claude Code plugins I built and run
I put the Claude Code tools I reuse across projects into one public toolshed: generators, workflow commands, anti-slop checks, search visibility, rule packs, and prompts.
My AI was smart. It just couldn't see my job.
My work context was scattered across email, meetings, documents, and notes. So I wired it into a private Obsidian vault Claude Code can read.
Your next team is three people and a pile of agents.
I expect AI-assisted teams to become smaller and organize around complete workflows. The model works only when the pod owns the outcome and runs what it builds.