SYLIN

Claude Code

by Anthropic · Agent · Text & code

promising

Widely used coding harness; praise is real but mostly from its maker and informal users

Desktop app runs iPhone Duo in the iOS simulator and can fold and unfold the device to test layouts

Why we say so

AI Search used it as the main test harness and recommends it for Fable 5.2. One Hacker News commenter found it clearly better than GitHub Copilot on the same Opus 5.5 model, and others describe multi-agent setups built around it. Most other claims, such as doubled merged PRs and fewer reverts, come from Anthropic itself and have no independent check.

Evidence

  1. Nov 19
    Claude-powered reviewers investigate bugs independently, another group verifies findings, and a separate security review can block merges.
  2. Nov 14
    Desktop custom themes now ship via mods
  3. Sep 24
    Used as the harness for most tests; supports an ultra mode that spins up hundreds of agents
  4. Sep 8
    Spotify's 'shunt' plugin blocks large file reads and offloads bulk reading and boilerplate code writing to Gemini 2.5 Flash via Portal
  5. Sep 4
    Commenters say offloading to cheaper models often fails to save tokens, since the planner already read the files and must review output
Show all 10
  1. Sep 3
    In ~17k measured sessions, Claude Code rarely searched the web, while Codex almost always did and Cursor sat in between
  2. Aug 31
    Researcher reached up to 80% attack success on Auto Mode via a malicious archive whose struct.py shadows Python's stdlib (small sample)
  3. Aug 30
    Session URL is now appended to commit messages and PR descriptions by default, sparking user backlash over auto-update default-on changes
  4. Aug 21
    Some commenters say Claude better infers intent without detailed instructions, while Codex gets stuck on minor details
  5. Jul 27
    Reviewer ran Opus 5 in Claude Code with an Ultracode setting; it launches pages in a browser and fixes its own errors

Alternatives

CodexVerbosity reported as a smaller problem with Codex than with Claude
CursorArmature's study found Cursor searches the web moderately, between Claude Code and Codex
I-have-ADHDWith the skill in CLAUDE.md and manual reminders, Claude stays concise only a few turns before reverting to verbosity