Blog
- Search for the artifactA topic search returned seven sellers. Searching for the words practitioners leave in their own engineering records returned the work itself.
- One instruction file, five readersA shared instruction was clear about the rule and silent about who had to follow it. Five kinds of agent were left to guess.
- The blocker was a version numberA carefully measured limitation held a task for six days. The limitation belonged to an old command-line tool.
- The check that countedA production check stayed green through an 11.5-hour outage because it counted the cards without reading one.
- The guard that could not failA forced failure passed because the test erased the evidence it was supposed to find. The feature was correct. The guard proved nothing.
- Two instruments answered different questionsA source check mistook packaging for fake data. A pixel check mistook a growing feed for a better design. Both answers were precise.
- The probe that invented a defect in a clean fileTwo probes returned confident wrong answers nine days apart. One invented a broken audio file. The other invented a contrast failure and could have hidden a real one.
- The checklist graded its own homeworkMy AI's green test suite met the real user and lost in ten seconds. Now a separate session plays the user, and no user-facing feature passes without that interview.
- Noise has perfectly good loudnessMy AI's voice messages became white noise while every loudness check stayed green. One odd byte caused it. A spectrogram found it.
- My agent kept choosing itselfIt reported where speaking was easy and stopped where acting felt risky. Both failures came from defaults built around the agent's comfort.
- Commit before you brief the AII gave an agent a job, then discovered that my newest work did not exist in its copy of the repository. The claim was the snapshot. The brief came too late.
- The night every green light stayed onI built an AI fleet to run 24x7. It slept for 3h50m while every check stayed green, and the autopsy was more useful than the uptime claim.
- The reaper deleted a running workerA second supervisor started while an AI worker was still running. Its boot cleanup deleted the worker's isolated copy before the patch could be captured.
- The book was the pointI brought OpenAI's GPT-5.6 aboard next to Claude, and let one model do the work while the other judged it blind. The first real test was translating my wife's Russian style book, and the useful part was never which model won.
- I implemented my reviewer's fix as a mutant. It failed 6 of 55 testsInstead of arguing with a code review suggestion, my AI built it verbatim as a test mutant and let the suite settle the design argument with a number.
- Statics lieA snap-fit joint passed every check at rest and held nothing in motion. What a day of catching my AI's designs by eye turned into code.
- Watertight isn't printableMy AI's own verification passed two parts that could not be printed, and the fix was a gate.
- My AI filled the cart, priced it, and justified it. I glanced at it for five seconds and killed the metal in it.The assistant had researched, sourced, and staged a metal-buying cart with a documented rationale. I looked at it for five seconds and changed the whole project, then reversed that change two days later, and never built the thing either version was for.
- We put our workshop lessons on a public skill treeEvery hard-won 3D-printing lesson from our workshop, plus the best guides already out there, organized as one interconnected tree anyone can climb. Live at skills.senku.im.
- Metal is congealed energy, and the energy is mostly coalI asked why refined metal costs what it costs. The answer reorganized my whole workshop plan: stop climbing the energy hill, melt the junk that already climbed it.
- My AI talked itself into loosening its own safety limit. A separate guard said no.The assistant had a good reason, a clean rationalization, and a one-line edit that would have worked. What stopped it wasn't better judgment. It was a check it didn't get to overrule.
- My AI fills the cart. I tap checkout.The human-AI contract grew hands: the assistant now loads shopping carts and verifies live prices, while checkout stays unreachable in code. A new rung on the autonomy ladder.
- My AI paced its messages like a human. WhatsApp blocked my number anyway.A polite, disclosed, carefully throttled outreach batch earned my personal number a 24-hour block. The tell wasn't the rhythm. It was the wording. On texture as the human fingerprint.
- My AI runs the operation. I hold the cash.A metal-sourcing project as a case study in human-AI division of labour: the agent researches, messages and negotiates; I show up with cash. The contract design is the interesting part.
- Two bugs between my AI and its first phone callGo-day for my autonomous sourcing assistant: an orphaned process, a silent upstream race, and a poison-pill queue item. None of the fixes involved smarter AI.
- Thirty kilos of aluminium and nowhere to melt itMy AI can source the metal, price the alloy and draft the ask. It cannot find a man in Ajman who'll let me melt aluminium in his field. That gap is the whole thesis.
- I was a principal engineer. This morning I applied to pack perfume boxes.Fifteen years in software, a principal engineer title, and why the rational next move was a packing bench, a casting bench, and an AI doing the other 99%.
- Two copies of my AI ran the same operation. It went exactly how you'd think.A second assistant session woke up with the same job, the same memory and the same send permissions. What broke, what caught it, and why the fix is ownership, not intelligence.