Studio Toriumi / Observatory
Field notes on Claude Code and other coding agents, collected every day from Zenn, Qiita, Hacker News and Reddit. Filtered for the things that aren't in the documentation — the configuration people actually use, the workflows that survived contact, and what went wrong. About two thirds of what lands here is in Japanese, which is the part English-language aggregators miss entirely.
New posts scroll past, but the principles underneath them repeat. Reading these first makes everything below absorb faster.
What decides an agent's output is less the model than what you let it read. Load it with irrelevant files and the important part gets diluted until instructions start getting dropped.
Two things work: cut the unit of work smaller and reset the context at the boundary, and push long investigations into a subagent that returns only its conclusion.
Typing "and write tests" every session isn't instruction, it's re-invention.
Put it in a project convention file like CLAUDE.md and it holds without being asked.
The trick is not writing too much. A three-hundred-line convention gets skimmed. Twenty to thirty lines covering only what actually breaks if ignored is the practical size.
Tests, type checks, linters, a build command — anything the agent can use to verify itself changes the quality of what comes out. In areas with no verification available, agents write something plausible and stop.
Which means: the better your test coverage, the stronger your agent. The real investment for AI-assisted development turns out to be test infrastructure.
Once you get used to the speed, it is tempting to ship code because it ran. But nearly every failure story that surfaces comes from a diff that was merged unread.
If reading takes too long, make the change smaller instead. Not generating changes larger than you can review turns out to be the faster path overall.
The urge to let it run unattended is understandable, but reversible actions
(editing files, running tests) and irreversible ones (push, deploys, deletions,
anything that leaves the machine) deserve different treatment.
Keep the second group off auto-approve. There is no undo after the fact.
Correcting the same mistake in conversation treats the symptom. Fix the convention file, the script, the type, or the test and the mistake becomes impossible to make again.
"Getting good at prompting" pays less, over time, than "building an environment where the wrong thing can't happen."
Headlines and original links only — read the posts at the source. Ordering combines recency, engagement, and how much a title smells like actual practice (configuration, workflow, lessons learned).
Nothing matches this filter.
No reposting and no translation of anyone's article. Headlines and original links only — you read it at the source. The people who wrote it should get the reader.
Being listed is not an endorsement. Collection is automated, so accuracy isn't vouched for. The six principles above were written here.