Drop it into your agent. It triggers itself whenever substantial output has to be trusted: before a pull request, before you ship, any time you are deciding whether something is actually done.
A routine is just a Claude Code session that triggers itself. Field notes on the three decisions every routine comes down to, with a work order builder you fill in and copy.
The field rule. Decide three things and the rest is infrastructure: when it wakes up, what context it gets, and how far it can steer on its own. What you stop owning is the point.
Hi, I'm TCash. I ship with coding agents for a living and use this place to think out loud about the parts that don't make it into anyone's docs: where the loop stalls, which corrections actually unstick it, and why the artifact that survives the session is the one that lets the work compound.
Every dispatch here ships as a single, self-contained HTML file you can open on a double-click. Which is the whole argument, really. The nine surfaces where HTML beats markdown come from Thariq's “The Unreasonable Effectiveness of HTML”; the beats below are just where my own work keeps landing.
No hot takes I can't run in production. If a dispatch didn't survive contact with a real repo, it doesn't post.
Most of what your agent hands back wants to be a page, not a paragraph.
The agent can produce more than you can responsibly inspect. The next skill is keeping verification cheaper than generation, so the loop creates leverage instead of review debt.
Six months of defaulting to HTML for plans, reviews, and diagrams: what changed, what broke, and the one carve-out I keep. A reply to Thariq and Kurtis Redux.
Twenty-one rules went in. Five survived contact with real work, and one of them was “write less here.”
When the loop stalls, I reach for the same four corrections, in order, before I touch the prompt.
An HTML diff with the reviewer’s notes pinned in the margin beats a comment thread every single time.
Where the cheaper agent holds, where it folds, and how its prompting wants to be shaped differently.
Treat the agent as a study partner, not a vending machine. A reading method that survives long sessions.
Notes from how a shop that ships every day delegates to its agents, and where the humans stay.
Ask the model to write the prompt before you write the task. Why the extra hop earns its keep.
Seven points where a human still has to step into the agent’s loop, and what each one costs you.
Skill compounds where the work happens, not in the docs. A case for letting the agent watch you work.
No digest, no drip, no “7 prompts that will change your life.” You get the piece when it ships, in plain text, with a link to the HTML version. Roughly twice a month.
Or pull the feed: tcash.ca/feed.xml · no tracking, unsubscribe in one click.