Notes from West Cork. I spent most of this week offline on the southwest coast of Ireland: slow, analogue, screens nowhere in sight. Easy to forget that's the whole point. I don't pour time into automation to go faster for its own sake. I do it to buy more evenings like the one I wrote this after: outside a pub, pint in hand, sun coming off the water. The automation is the means to that, not the point of it.

Context is data. I keep a set of rules for what my AI knows about each customer, built up over the best part of a year, each one added the day something went wrong. Line them up and they're not note-taking, they're data governance: one file owns each fact (primary key), don't duplicate, link instead (normalisation), a 300-line cap so old detail ages out (lifecycle management). I didn't set out to port data governance onto markdown. The disciplines I reached for by reflex were the ones I'd spent years using on the customer data itself, and that's the point. The context an AI holds isn't like data. It is data. Manage it like data and the agent stays sharp. Let it rot and it gets slow, confused and confidently wrong, the same way a neglected warehouse does. Blog post

Model vs effort: really only two levers. A question doing the rounds: with model, effort, context and time all in play, how do you actually choose? They're not four levers. The only real axis is how much reasoning a task needs, and that sets two levers: model and effort. Context should always be the minimum that's fully sufficient, nothing more. And if time ever feels like a factor, that's usually a sign your context needs work. Worth knowing: effort is a behavioural lean, not a token budget, and it shapes everything including tool calls. Opus 4.8 defaults to high now (Anthropic say it uses about the same tokens as 4.7 but performs better) – though xhigh is still where I start for coding and agentic work.

Why your new session opens heavy. If a fresh session is already large before you've typed a word, it's almost always your memory files. Run /memory – it lists every CLAUDE.md (persistent instruction file), local file and rule that's loading, plus what auto memory has saved for the project and your sub-agents. The usual culprit: an @file import pulling a whole document into context at launch. Drop the @ so the agent knows where to look without loading it, and keep each CLAUDE.md under 200 lines. Docs

On my radar

Opus 4.8 landed. Better at the things that matter for agentic work: fewer compactions (forced context resets mid-task), better recovery on long context, more reliable tool calls, and Anthropic's claim that it's ~4× less likely to let a flaw in its own code slip past unremarked. Early reactions were mixed, the usual rocky start before a release settles. Anthropic

Dynamic workflows. New in Claude Code: a script Claude writes to orchestrate dozens (even hundreds) of sub-agents in the background, built for codebase-scale audits and migrations. It can also run quality patterns, like having independent agents adversarially check each other's findings before anything's reported. Powerful, but token-hungry, so run it on a small slice before you point it at the whole repo. (Worth clearing up: "ultracode" isn't a magic mode, just xhigh effort plus standing permission to launch these.) Claude blog

The lock-in squeeze is getting real. Cron-triggered Claude Code calls start metering mid-June, nudging people back toward flat-rate tools. The standing lesson holds: keep your infrastructure model-agnostic. Vendor lock-in on something this strategic is a bad bet.