Caveman changes how Claude talks. Cuts articles, filler, hedging. ~75% fewer output tokens, same technical accuracy. Code and errors stay exact. Ponytail changes what Claude builds. A lazy senior dev who asks "does this even need to exist?" before writing anything. Stdlib over deps, one line over fifty, shortest diff that works. One governs the mouth, one governs the hands. Stack them = terse explanation of a minimal solution. Which are modes do you find the most success with?

Anyone else having issues with remote control? I get used it up until I need to approve an action and it just decides to get stuck

I'm on the Claude Pro plan, and I've been hitting my session limit pretty consistently and I'm trying to squeeze more out of each one. Curious what's working for people. What I'm getting Claude to do right now: - Model routing — drop to a cheaper model for routine turns - Session handoffs instead of /compact - CLIs over MCPs where I can - Read each file once and reference it from then on Anyone got other tips for keeping token usage down / getting more out of a session before hitting the wall? Always looking to learn.

Most health practitioners I work with have no idea what lands on 19 June. The Data (Use and Access) Act quietly reshaped UK data and cookie rules back in February, and from the 19th every business needs a proper route for people to raise data complaints, plus the process behind it. PECR fines now sit on the same scale as GDPR, so "too small to matter" stopped being a defence. Rather than work through each site by hand, I built a function that does two things: - Audits my practitioner clients websites against the new rules (cookies, consent, banners, the lot) and brings it into line. - Drafts the policies they are now expected to hold, structured so they only complete the parts that are genuinely theirs. The interesting part was deciding what to automate and what to leave to the human. The site fixes can be handled in full. The policies cannot, not honestly, because the practitioner has to own the decisions inside them. So the function does the heavy lifting and hands back something most of the way there, rather than pretending to finish a job that needs their judgement. Built solo, AI-assisted, and it now runs across every client site instead of being a one-off scramble before the deadline. Curious what the rest of you have been building. What is the most efficient function or automation you have put together lately?

Token-saving move that nobody talks about: don't auto-load your skills. I keep 50+ marketing/design/writing skills out of the default session. ~8.7k tokens that aren't burned every time. A lean catalogue file points to them. Claude reads the one it needs, when it needs it. Lazy-load your skills. Your context window thanks you.