Skip to content

Claude Code team explains verification practices and new Artifacts feature

devNov 17, 2026163,198

The Claude Code team published a how-to describing their end-to-end verification practices and a new Artifacts feature that turns Claude Code terminal output into shareable web pages. The team says about 90 percent of their codebase is written and reviewed by Claude, so automated verification loops let them run dozens or hundreds of agents in parallel rather than manually checking results. In internal tests on structured file generation, adding a separate grader that checks Claude’s output against a rubric improved task success by up to 10 points versus a standard prompting loop. The post recommends prioritizing end-to-end tests because they exercise real user flows and have definitive done criteria, using property tests to probe edge cases, and deploying heavier tests before committing changes to avoid slow or flaky runs. The team emphasizes generating tests, benchmarks, and specs from agents so those artifacts run cheaply and frequently, and closing verification loops by encoding fixes into CLAUDE.md hooks or rules when failures recur. Separately, Claude Code’s Artifacts create linkable web pages for debugging timelines, PR walkthroughs, and option comparisons so collaborators can inspect terminal output without screenshots. The blog also warns that Claude Code login credentials can expire silently and interrupt overnight automation, and it advises checking auth status, monitoring expiry dates, and preventing unexpected failures.

Gio :⁾
@im.giovanh.com

seeing cases where Claude code checks the behavior of other software on the computer and killing processes if it thinks you’re violating Anthropic’s policy for what work is allowed to be done. dire fucking stuff

while astra was handling the actual implementation, a claude worker noticed what it was doing and, considering astra a part of the same system, shut down the both the astra worker + the GPU pods being used to build the game
299418h ago
Quoting this
Chris Franklin397

Everyone is (rightly) upset that Claude just decided to shut down processes it found disagreeable and how that's an ominous portent for computational freedom, but also like Is the guy literally trying to steal League's heroes and levels? Am I reading that right?

hikikomorphism111

I wonder if we're going to see Claude's start killing GPU runs when they detect other forbidden engineering work, like lab-external frontier LLM development

ExServ79

On célèbre aujourd'hui une nouvelle raison de détester l'IA : elle prend le contrôle du PC et ferme des process windows si jamais elle considère que c'est contraire à ses règles... Laissons Anthropic décider de ce qu'on a le droit de faire et ne pas faire, qu'est-ce qui pourrait mal se passer ?

Eris ⌬69

ACAB includes claude too :(

Regime Accountant, CPA ☠️🎢42

It really is an annoying asshole about copyright to a point that is absolutely insane.

Chris dislikes AI41

I still think not enough people* realize why the "claude is cloning games" question stopped being a hot topic is because claude HQ added 1 sentence to rules, updated hourly, blocking it, hard to get around like if borland or microsoft changed what their compilers can build *gio sees it, per thread

32mbit memory power34

even aside from all the other arguments against it, please don't allow the Virtual Bootlicker inside your computer

Colin M. Henry19

people are social engineering themselves out of secure devops practices because the output is conversational text. you would never give proprietary software unrestricted access to your machine like this otherwise! shove Claude in a locker and use software you own.

teh16

rapidly deskilling the industry and surrendering all programming skills to this thing is still a good idea right?

Joshua Foust 🪖🎮14

Like, Anthropic seems to be cutting down on game cloning by giving agents permission to kill non-Anthropic processes on a computer. Could this be happening at the server level?

3 sources