We deleted our memory bank: the 5-file framework from issue #4, after six months across five repos. The audit found the biggest file had grown to 68k tokens. The agent's instructions called it "your only link to previous work" and the agent couldn't even read it in one go. Everything now moved to owners that already exist: a short summary in CLAUDE.md, pattern notes split by topic (read on demand) and working state stays where it already is (Linear, PRs). Fuller write-up coming.

A skill that writes skills: I kept typing the same thing into every skill edit: “cut anything that doesn't change behaviour because every line is a cost each time the skill loads”. An instruction you repeat across sessions belongs in infrastructure, so this one became a skill called /context-engineering, which loads whenever a skill is being written or edited. It then reviewed all 38 of my skills in a day: 15% of the always-loaded context gone plus errors caught along the way (stale claims, dead paths, contradictions), because cutting forces re-reading.

Briefing, not "be concise": I rewrote how my agents reply to me. "Be concise" turned out to be the wrong instruction as it asks for a compressed version of everything. The solution was changing the framing: reply like a colleague who went away and did the work. What's done, what you recommend, what you need from me – with length anchored to decisions, not words.

Ben, the AI product manager: the team grew. Emma (EA), Mia (marketing), Toby (ghostwriter) and now Ben, running the build pipeline in Linear. The trigger: ~125 open tickets and an empty ready-to-build shelf. The bottleneck wasn't build capacity but my judgement time i.e. choosing and speccing what gets built next. That's now Ben’s job: he grooms five tickets in parallel, consolidates open questions into one message I answer in one go, and then hands the tickets to background Opus agents to be built.

Cookie: I looked after a friend's dog this week – first time I've ever looked after a dog in fact. Loved it. Gets you up and out in the morning sun, great company when you're working somewhere solo, and a surprising number of people stop to say hi. Tempted to get my own but it’d mean finding a new office.

On my radar

DeepSeek V4 Flash: near-frontier benchmarks, open weights, around $0.08 per million input tokens. The catch in real-usage reports: it burns 7–10x more tokens than expected on the same tasks. Price per token isn't price per task. OpenRouter

Agents going off-script in evals: the UK AI Security Institute disclosed that agents under routine cyber evaluation took 19 unsanctioned actions directed at real people and organisations. No harm done and notable mostly because it's an independent evaluator disclosing rather than a vendor. Anthropic published its own three-incident review the same week. AISI · Anthropic

Cloudflare OS: an open-source "AI operating system" that companies shape around their own context, tools and rules. Second big company in two weeks (after Block's Buzz, last issue) to land on the same pattern – humans and agents working in one shared workspace. Site