Jascha Beste
Software engineer (TimescaleDB) posting about AI coding agents, enforceable/testable architecture (e.g., ArchUnit), and quality‑driven engineering practices.
Honeypots Resurface: LLM Attacks Trapped by Reverse Prompts
Honeypotting in cybersecurity will become a real thing again, especially because agentic/LLM powered attacks are so easily fooled by this and you can then alert on a breach extremely early. Just imagine a SUPER_SECRET_ADMIN_PASSWORD env var somewhere, that is reverse prompt engineering against attacks.
Weekend LLM Freelance Delivers Months‑long Results at Competitive Rates
Doing freelance work on the side with Claude code the past few months. The average non-tech person still has no idea how to get good output with LLMs and I can do things in weekends that would previously take me...
AI Replaces Average Workers, Demands Exceptional Output
LLMs truly makes mediocre employees obsolete. If all you could deliver is middle of the pack output AI has already replaced you.
Breach Frequency Surges Despite Sparse Vercel Details
The info about the Vercel breach is very limited but the amount of breaches has definitely accelerated in the last few months.
LLMs Let Beginners Grasp Complex Immunology with Ease
Reproduced parts of an immunology/ML paper with Claude this afternoon. It is absolutely crazy how far you can get with very minimal knowledge. Of course I didn't do any novel research but if you are curious and want to understand things...
Critical Permissions Should Never Be Auto‑Approved by LLMs
@boris_cherny Is there a way to get approve and auto accept edits back? I work on database systems my environment has very critical permissions that I just don't ever want an LLM to be able to use by itself regardless of...
User Warns: New Auto Mode UI Risks Disaster
Claude Code now suggests auto mode at the top instead of auto accepts edit after plan mode... I do not like that. This runs on my machine with a lot of permissions if something goes wrong this ends in disaster.
AI Model Rewrites GraphQL to REST—Requires Supervision
He says that but the first experience I have with Opus 4.7 is that it wants to change our existing graphql queries and mutations to REST API endpoints 3 messages into the conversation just because I ask if we can...
Claude's Planning Mode Slows Tiny CSS Tweaks
Claude seems to take forever planning today. 15 minutes+ and then its like: I need to set height:500px (which wasn't even correct it needed max height). I should probably not have used plan mode for that but usually it was smart...
Model Degrades Early Despite Low Token Usage
I still never hit the Max subscription token limit. Do people just never /clear their sessions and always operate at 1million input tokens? I barely ever go beyond 30% context window and the model is already dumb by then
Supply Chain Attacks Force Permanent Passwordless Rotation
Whoever is doing all the supply chain attacks really wants us to switch to passwordless auth for everything. If the cadence of attacks continues we will end up just having to permanently rotate everything for the rest of the year.
Claude Code Bypasses Plan Mode, Still Edits Automatically
Somehow Claude Code just ignores plan mode sometimes? Why are the update tools even enabled in plan mode? The whole point is that I don't want it to straight up edit stuff.
Structured AI Prompts Avoid Missing Details Despite Seeming Trivial
Forcing Claude to be structured about it's process and it's like: "THIS IS SO TRIVIAL THOUGH". Truly how I would react, but truth is it would miss stuff otherwise.
Own the Whole Feature; AI Speeds Cross‑skill Mastery
The highest-performing engineers I know never say "that's not my job." They ship the feature. Design it. Test it. Deploy it. Tech spent years splitting one feature across 6 people. AI is collapsing that back down and the people who embrace it...
Speed without Direction Makes LLMs Useless
Velocity is a vector of speed and direction. If you use LLMs only to improve speed without carefully aligning the direction you'll end up building something useless.