I turn complex ideas
into carousels
people finish.

THE SOURCE

been thinking about why AGENTS.md feels like the finish line when it isn't. prompt engineering optimises what you say. context engineering optimises what the model sees. neither covers what happens when the thing runs for six hours with nobody watching. that's the harness. tools, sandboxes, permissions, recovery paths, feedback loops. guides vs sensors. most people build guides and forget sensors, which is why review still eats their whole day. need to work out the ordering, what actually pays off first...

WHAT I DID

Eleven slides built on a three-layer argument: prompt, context, harness. Yellow marks an insight, pink a warning, green an action, so the colour tells you what kind of slide you are on.

THE SOURCE

"People are trusting their agents to do more work without supervision, but how do you decide when to trust them?… The real answer has nothing to do with the model, and everything to do with the task." Source: posthog.com/newsletter/agent-autonomy. Full post, ~1,200 words.

WHAT I DID

Built in PostHog's own brand tokens rather than my house style. Three full-bleed colour slides break up the rhythm so a long post does not read as a wall.

THE SOURCE

ok so the p99 thing. turns out the pooler had idleTimeoutMillis set to 1000, nobody remembers setting it. under burst traffic basically every request was opening a fresh connection and eating the full TLS handshake. db cpu was flat at 22% the whole time which is why the dashboards looked fine. deleted the line. p99 went 890 → 550. connection churn down ~40%. zero infra change, zero cost. writing this up before i forget.

WHAT I DID

A Slack dump restructured into hook, symptom, diagnosis, fix, result. The diagram and the metrics stack vertically because the frame is portrait.