TaggedGetaggt mit#tooling
10 notes.10 Notizen.
A test that reads your docs
An ordinary failing test whose subject is a sentence. The mechanism is three lines: read the fact out of the code, look for it in the document, complain when it is missing.
My biggest CLAUDE.md was 224 KB
Nine agent instruction files went from 455,530 characters to 88,584 — and the split that actually saves context is a plain markdown link, not an `@path` import.
Don't read what the agent did — read what it decided
A long AI task leaves behind a wall of output. Having it finish with a one-page report instead means you can judge the thinking before you check the work.
Stop watching the terminal — let it text you
A Telegram bot can report when a long job finishes; the useful decision is what is safe to put in the message.
One missing line, and all of GitHub can deploy to your project
Passwordless CI needs two scopes: the identity provider names the repository, and IAM limits who may impersonate the deploy account.
One tool shipped, one never ran, and neither turned CI red
The real migration numbers: 5.2×, not 35×, plus a lockfile regeneration that triggered the supply-chain scanner.
The Firefox port I budgeted for didn't exist
Chrome and Firefox can load the same extension code unchanged; permanent installation is where their paths diverge.
Coding from my phone works because none of it runs on my phone
Claude Code on the web runs each session in a cloud VM, so closing the browser doesn't stop the work — I shipped two small apps from a train with no laptop.
I told Claude Code to build its own status line
The status line is just a script that gets handed a JSON blob — git branch, files changed, usage limits — so building one is mostly describing what you want.
Four tools do three jobs, and I wait on all of them
Both my ecosystems grew a single fast binary that replaces most of the code-quality stack. Two thirds of that swap is safe; the third part has to earn it.