libghostty: Why Agent Tools Build on Ghostty (and I Build a Terminal)
What libghostty is, who really uses it (and who doesn't), and a working terminal window built on libghostty-vt in about 180 lines of C.
elbcasts.dev
screencasts for working developers
What it is
Short, practical videos about the things you actually ship: Rails,
Kubernetes, Flux, Helm, and the AI tooling that is quietly changing how
we write code.
Two kinds of episodes:
• Screencasts — a real demo app, a real terminal, a real editor.
One feature, start to finish.
• Explainers — one concept, under ten minutes, no filler.
Written and produced by Mathias Karstädt. Every episode is built with an
open pipeline that runs entirely on a laptop.
elbcasts.dev
Episodes
GPT-6.1 Terra vs Kimi K3.1 Pro: Same Rails Ticket, Real Cost
GPT-6.1 Terra and Kimi K3.1 Pro get the same Rails ticket in the same agent, twice each. What a finished ticket really costs: time, money, interventions, tests, and one rule the tests don't check.
Survival LLM: A Free Offline AI for Blackouts and Emergencies
Survival LLM is a free, open-source app for Mac and Windows that answers emergency questions with a local AI model and a local reference library. No internet, no account, no tracking.
Can OpenCode 2.0 + Beam Replace Claude Code?
What OpenCode 2.0 and Reflection's Beam bring, how to set them up via OpenRouter, and both stacks solving the same Rails ticket side by side: speed, cost, quality, maturity and privacy.
Ruby's Safe Navigation Operator
Stop writing nil guards: Ruby's safe navigation operator &. in under two minutes, where it fits well, and the one thing to remember.
I Let 3 AI Agents Build a Rails Feature. Here's What Broke
Claude Code and Codex work in parallel on one Rails ticket, with a coordinator writing the failing tests first. What broke, what it cost, and when one session beats a team.
OpenMausBot: your own team of AI bots
OpenMausBot gives you a team of AI bots in a chat window, running on an agent you already have. Build a bot that writes release notes and a second one that fact-checks it against the code.
I set my AI fraud threshold to 0.98. It paid out nothing
A fraud threshold of 0.98 felt safe and paid out nothing. Route risky decisions three ways (pay out, block, ask a human) and read both thresholds off a cost table instead of guessing.
Don't trust the model card: three decision models on real data
Three decision models, tested on real data instead of model cards. Measure any classification model on your own labelled data in twenty lines and use the probability to decide what runs without a human.
Rails Turbo Streams from a Background Job (No JavaScript)
Push a change from a background job straight into an open page with Turbo Streams and Solid Cable, no JavaScript needed. And see why the async adapter never works for that.
Your form waits three seconds — move it to a job (Solid Queue)
Your form makes users wait three seconds. Move the slow work out of the request into an Active Job, run it with Solid Queue in development, and learn where jobs get lost.
Stop Fat Controllers: Rails Form Objects Explained
When one form spans two models, the controller gets fat fast. Build a plain Ruby form object that keeps the validations in one place and the controller thin.
Your Rails uniqueness validation won't stop duplicates
Rails' uniqueness validation alone won't stop duplicate records. Learn how to stop bad data at the model with validates and normalizes, show the errors in the form, and back the rule with a unique database index.