.sh → shell → Saint Helena, napoleon's exile island, now home to every curl | sh install script hahaha

hiring down 29% and every one of those companies still has the "we're hiring!" post pinned on linkedin from 2023 hahaha

quickbooks desktop has no API. so we built an agent that works inside it anyway receipts, bank feeds, recurring transactions, collections, P&L... a day of bookkeeping in about 90 seconds

the only two agents i run with zero hallucinations. 100% uptime, will still eat your shoes

anthropic's next model fixes your bug, then finds the whole class of bugs behind it and asks if it should fix those too. they built the staff engineer who turns a one line fix into a quarter long refactor lol

AI demos get a perfect typed PDF. real customers get a clipboard and whatever pen was in the cup lol

she said no laptops on date night. never said anything about the dashboard 😈

4 terminals of agents cooking and the first usage warning all day shows up at 3:38... for a 3:40 reset ...grinding Diablo because they don't need me anymore lmao 🤣

I'm so sick of seeing "Claude just killed x" or "GPT just destroyed y forever" posts. Sure, you can one-shot a cool demo when you have CC hooked up to API billing with no limit and you don't actually want to use whatever minecraft or FPS game you "make". Reality is, it's actually difficult to generate/iterate/test/deploy/maintain a piece of software that is successful and adopted.

I put an AI reviewer on every PR expecting the value to be in the findings. 171 of them got a written verdict. 158 accepted, 11 rejected, 2 deferred. I assumed the accepts were the cheap half — apply, merge, no trace left. Wrong. Accepted findings annotate the code too. There are 79 of those notes in src/ against 34 from rejections. Agreement leaves more evidence, not less. The difference isn't the writing. It's that an accept rides the automation and a rejection stops it — each of the 11 dropped the chain back to per-operation approval and waited on me. One waited too long. PR #69 merged before I signed off, so the rationale landed on main by itself: a 5-line comment, its own commit, message explaining why it had no PR attached. 158 findings the pipeline absorbed on its own. 11 that cost a human. The receipts were never the expensive part.

I've been treating an empty folder as our error rate on a document extraction agent. The prompt tells the model to write "" when a field is blank on the page. My code writes "" when the model didn't return the field at all. Same cell either way. I never noticed. I went looking for the code that catches a misread. Four paths lead to failed/: retries exhausted, file won't open, malformed JSON, and a bare except Exception. All four are plumbing. The only thing checked after a response is three shape tests. Is it JSON, is it a list, are the items objects. Nothing looks inside an object. So I wired a fake client into the real module and returned one empty record. {}. It became a row: twenty empty fields, plus the filename and timestamp the agent stamps on itself. The blank-row filter drops a row only when every cell is empty. The agent fills two of them. Count went up by one. TIME SAVED went up three minutes, because that tile is record count times a hardcoded 3. failed/ was never an accuracy rate. It's a list of the failures I remembered to name.