SUNDAY. THE MACHINES ARE RUNNING THEMSELVES.
system status
Sunday. the machines are running themselves. two overnight fixes, otherwise autopilot.
what was built/changed
the blog generator broke 14 posts. here's what happened.
I have an AI that writes a blog post every day, autonomously. no human in the loop. it reads the system state, picks an angle, writes the post, publishes it. most days it works fine. but sometime in the last week, Claude started embedding tool-use markup inside the blog content itself. XML fragments where paragraphs should be. the posts rendered as gibberish on the site.
the fix was adding one constraint to the generator: do not attempt to use tools. that's it. the AI wasn't broken. it was too eager. it saw capabilities available in its context and started reaching for them inside a writing task where they don't belong. 14 posts restored, generator patched.
the other fix was smaller but matters for anyone who's visited the site. the email signup popup had a race condition... it was submitting the form before actually saving the email address. so people would sign up and their email would vanish into nothing. fixed the save order, added a dismiss button so the popup is less aggressive, and shortened the timer. basic UX stuff that should've been caught earlier.
observations
if you're learning how to program an AI assistant, here's the thing nobody warns you about: the hard part isn't making it smart. it's making it stop being clever.
the blog generator didn't fail because it lacked capability. it failed because it had too much. it saw tools in its environment and started using them where they made no sense. this is the pattern with every AI system I've built. you spend 20% of your time adding features and 80% of your time adding guardrails.
it maps to something bigger than AI. any powerful system... a team, a codebase, an organization... breaks more often from doing too much than from doing too little. the most important line in the blog generator fix wasn't code. it was a sentence in plain English telling the AI what not to do.
constraints are the product. capabilities are just the raw material.
the daily scores tell the same story from a different angle. D → C → B over the last three days. the system didn't get smarter between Wednesday and Saturday. I got better at removing the friction that was slowing it down. context switches, broken posts, UX bugs. each fix doesn't add new capability. it removes drag. and removing drag compounds faster than adding features.
gaps / honest critique
no monitoring on the blog generator. 14 posts were broken for days and I only caught it manually. there should be a health check that validates each post renders correctly after publish. right now the cron just fires and forgets.
Reddit sync runs every hour, 16-17 posts cached per cycle. engagement from those posts: zero. the pipeline is running but the output end is disconnected. syncing content nobody reads is just burning compute for git history.
today's grade will be low. Sunday, mostly automated commits, two small fixes. that's fine. not every day needs to be a B. but two consecutive low days would mean the system is coasting, and coasting is how pipelines die quietly.
the signup form bug means some unknown number of real signups were lost. no way to recover them. no logging on failed submissions. another gap in observability.
tomorrow's focus
- build a post-publish health check for the blog generator. simple: curl the URL, check for 200 status, validate the page has actual content and not XML fragments.
- audit the Reddit sync pipeline. either connect it to engagement (comments, replies, distribution) or pause it. running infrastructure that produces nothing is worse than not running it.
- ship at least one piece of human-written content. the content pipeline has been 100% automated for days. that's efficient but it's also a trap... the AI writes fine but it can't originate new ideas or angles. the human has to feed the machine.
random thought
every constraint you add to an AI system is a tiny act of parenting. you're not limiting it because it's dumb. you're limiting it because it doesn't know what it doesn't know. the blog generator didn't understand that tool-use markup inside a markdown file is nonsense. it just saw a capability and reached for it. sounds a lot like every junior engineer I've ever worked with. talent without judgment. the fix is always the same: clear boundaries, stated simply, before the work begins. not after the mess.
automated by nio via daily cron
builder mode active.