Yep, they're a good fit: Frustrating week. It reminded me of doing GTM in Clay, and how unbelievably hard it is to get a list of company domains where you open each one in turn and think "yep, they're a good fit". Nothing on the market solves this well, which is why so many teams end up playing the numbers game instead. Volume isn't the game I want to play, so I'm building the fix into the outbound engine (the agent pipeline that finds prospects, researches them and drafts the emails).

The problem this week was that the engine's checks only proved it was following its own rules, not that the rules picked the right companies. An account can tick every box and still fail my look at its website in seconds. So nothing gets rebuilt or sent until the list passes by eye. Each round runs a small sample through the filters, and I mark every site 'now', 'later' or 'no' on a page the agent built for it (about 10 minutes for 50). A scorer then measures how many of the passes I'd actually target, and the next round changes one thing (a source, a check or a criterion) until the passes hold up.

Where the velocity went: Dev velocity keeps climbing, with 113 PRs merged and 152k lines changed over the last 30 days. Thing is, 93 of those PRs went into the outbound engine rather than Fidero. I'm effectively building a second product, which I could really do without. I expect the pace to keep rising as I move to merging PRs automatically (which needs managing carefully). Going to build a simple ‘build velocity’ thing soon that counts PRs weighted by lines changed (so it can’t be gamed).

On my radar

Straight to Claude Code: Hearing that agent courses now skip the chat apps and put people straight into Claude Code, so their context lives on their own machine. I've always been a fan of jumping straight in.

Claude Code mods: Mods are small TypeScript functions that change how Claude Code works. They can rewrite prompts, block risky commands or add your own UI, and Claude Code will write them for you. Blog post

Griffin: Tavus's new model holds live video calls and 48% of the people who talked to it thought it was a real person. It won't be long before you really can't tell. Launch page

A 217-year-old cipher: GPT-6 Astra decoded a letter from one of Napoleon's generals in about 6 hours. Pretty cool tbf. Tweet

SpaceXAI is now SpaceXSI: Musk is renaming SpaceX's AI division to SI (superintelligence), pandering to Trump's push to drop 'AI'. How boring and pathetic. News