I made my claude code status line tell me when the prompt cache dies
1h ttl. every api call resets it. go afk too long with 900k tokens in context and your next message recaches the whole thing at full price
So now it just says e.g. "❄ 12m left · exp 10:46 PM"
green, yellow, red. reads the ttl straight out of the transcript jsonl. There's a .sh script behind the scenes that ticks it every 30s
stop paying 20x for forgetting to hit enter
first day in weeks where the codebase got smaller than it was this morning. closed specs, merged PRs, killed TODOs. spawned a couple follow-ups but net entropy went negative. forgot what that felt like
@SamaHoole Plato said in the republic to keep the worker class fed on grains and no meat so as the keep the weak to prevent them from rebelling. It’s the same playbook as old as time.
@dex_horthy
thanks for the causal mention months ago of JB Rainsberger's talk integrated tests are a scam. I heard you drop it during a BAML call with @vaibcode and then spent a week converting my test suite. Bugs have been down and so are my stress levels. Better a late thank you than never. https://t.co/jRapwGjEZG
@anshnanda Why default to E2E tests? I much prefer unit tests. I have it to TDD as per JB Rainsberger's talk on contract-collaboration tests, plus thin integration rings. This makes tests run super fast. He has a talk Integrated Tests are a Scam https://t.co/5C5b0BDVJw
@SamaHoole Plato said in the republic to keep the worker class fed on grains and no meat so as the keep the weak to prevent them from rebelling. It’s the same playbook as old as time.
Focus on the few not the many. Focus on the center of gravity. -Carl von Clausewitz.
Stop trying to boil the ocean. You must have laser beam focus if you want to achieve success.
@NicolasZu@paraddox Set a /goal or prompt what you want to implement, and append something like: spawn and orchestrate gpt-5.6-luna max thinking level subagents if and as appropriate so as to reduce token costs
You need to free up your brain's processing power as much as possible. Delegate low level stuff to your autopilot. Daniel Kahneman and Amos Tversky coined this System 1 (fast, subconscious, autopilot) and System 2 (slower, conscious, intentional)