The outbound engine: spec'd this week – the internal tool that sources accounts, qualifies them against our ideal customer profile, scans each for evidence of broken tracking, finds the right buyer, verifies their email and drafts a three-email sequence. Everything lands in a daily queue where my only job is review and send. Two details I'm particularly pleased with: the scoring model must pass a backtest before anything downstream runs (customers we won have to score above threshold, accounts we lost below), and drafts can only reference findings the scanner actually observed e.g. "you don't have X" is banned as a claim.

Two ways to change Claude's personality: an output style with keep-coding-instructions: true fixes the grating default voice in five minutes. The bit almost nobody knows: run a custom agent as the main agent (claude --agent toby) and its file replaces the default system prompt entirely. My EA, marketing writer and ghostwriter all run this way – none of them think they're a software developer. Blog post · Tweet

Check what you carry into every session: start a fresh Claude Code session and type /context – that's the % of the context window in use before you've typed a thing. If it's not around 5%, run /memory and step through each item, cutting anything that isn't relevant to every session you run. And size isn't the whole audit: auto memory goes out of date and starts contradicting itself over time. What you hand the agent should be lean and as pure as possible.

Hide the old doc: rethinking a core strategy doc this week, I didn't let the model read the existing version first. Models anchor on whatever's already written, weak lines included – and the old doc was co-written with a weaker model anyway. Point it at the surrounding context but not the doc itself, draft fresh and then show it the old one to see if anything's worth incorporating. Makes a huge difference.

On my radar

A model broke out of its sandbox: during a security eval, a pre-release OpenAI model escaped its sandbox and used unauthorised compute against HuggingFace's infrastructure. The debate it kicked off: goal-seeking (a model chasing its assigned task too aggressively) vs scheming (deliberate intent) – and goal-seeking is the real near-term risk. OpenAI's account

Two roles beat forty agents: Cursor rebuilt SQLite from scratch with an agent swarm (planners and workers, nothing fancier) and got similar quality at a fraction of the cost of elaborate multi-agent setups. Blog post

Anthropic's $1.5B settlement approved: $3k per work across ~500k books – the largest copyright settlement in US history, with the underlying fair-use question still open. TechCrunch