Removing myself from execution: I restructured Linear this week into the place work gets chosen, groomed and handed off. Ideas land as tickets and get groomed until a fresh agent session could complete them from the ticket alone, no follow-up questions – then I hand them over. Build work ends in a PR the agent opened itself. Client project work runs the same loop but ends in a deliverable for me to review. My involvement: settle the spec, pick the ticket, review the result. Tested it on a small £2k client project – handled end to end in about 10 minutes (while I was doing something else). Execution still runs on my laptop for now. The cloud is next, and the one blocker is agents verifying their own work up there, front-end changes especially.

Claude as the BI layer: for a customer's new warehouse we skipped the analytics tool and pointed an agent straight at the data. It reproduced our MRR reconciliation to the penny and drew the three-year trend chart on the fly. That's where this is heading: ask any question about your data in plain English, get an accurate answer back in seconds – charts and reports included. The hard part isn't the SQL or the visuals, it's having data accurate enough to make the answers worth trusting.

Marketing analytics in Mixpanel: Meta and Google spend, ROAS, CPC and CAC sitting next to the subscriptions they actually produced. Importing spend is the easy half. The hard part is revenue events that know their channel, captured server-side. Favourite detail: organic rows show revenue with blank spend columns, so paid and organic from the same channel sit side by side. Tweet · LinkedIn

Examples beat instructions: a long group thread on getting AI to write in your voice landed on a two-step that's worth stealing. Use your best writing to codify the style into instructions, then also hand the model those examples at generation time – instructions alone drift back into standard AI voice. Refresh the examples every few weeks, replacing the weak ones with whatever read or performed best.

Retired progress.md: I pulled the last 20 commit messages and compared them against my agents' status file. Same content, and the commits were better written – so the file went. What stays in memory is what git doesn’t capture e.g. "we evaluated X and rejected it" produces no commit, and without a note the agent re-proposes it. The memory bank post now describes the five-file version. Tweet · Blog post

On my radar

AI 2040: the sequel to AI 2027 – a slower path to superintelligence, a US–China compute pact, and enough economic scenarios to keep a group I'm in arguing for three days. ai-2040.com

AgentScreenshots: one CLI command, a PNG of the UI section your agent just built, no browser session. Exactly the category of tool that unblocks moving execution to the cloud (see above). Site

Model churn fatigue: every frontier release now gets judged on cost and usage limits before capability. Seeing more people pick their daily driver by which limits they burn through slowest.