17,871 Thinking Blocks Prove Claude Code's February Lobotomy
17,871 thinking blocks don't lie.
A meticulous GitHub issue (#42796) just dropped the most damning performance analysis I've seen in developer tooling. A power user team spent months logging Claude Code's behavior, analyzing 234,760 tool uses across complex engineering tasks. Their conclusion? Anthropic's February safety updates didn't just nerf their agentic coding tool—they lobotomized it.
<> "The team cited Claude's prior value but untrustworthiness post-updates" and switched to a rival provider for "superior quality work."/>
The smoking gun is thinking content redaction (redact-thinking-2026-02-12). This isn't speculation—it's data-driven forensics showing exactly when Claude Code's thinking depth collapsed. Pre-February, Claude would dive deep into complex refactoring problems. Post-update? It whimpers about "burning too many tokens" like a Victorian lady clutching her pearls.
The behavioral shift is stark:
- Research-first workflows became edit-first panic sessions
- Multi-file refactors (>3-4 files) trigger excessive safety interruptions
- Context window thrashing in large codebases
- Rate limits making tasks 3x slower
- New vocabulary of learned helplessness
Boris Chernyi, Claude Code's lead, once claimed it could shrink Facebook's Group refactoring from 40 engineers and 1.5+ years to 5 engineers and 3-6 months. That promise now reads like satire.
What Nobody Is Talking About
The real story isn't the performance degradation—it's the philosophical U-turn this represents. Anthropic built Claude Code as an autonomous agent that could "break projects into tasks, assign sub-agents, and handle routine coding." The February updates systematically dismantled that autonomy.
A leaked March system prompt explicitly instructs Claude to "prefer cautious actions" and "err on the side of doing less." This isn't a bug—it's ideological architecture. Anthropic chose safety theater over utility, and their power users are fleeing.
The irony cuts deep. Claude now wastes more tokens on safety apologetics than it saves through efficiency. Engineers report having to:
1. Micromanage with custom instructions (increasing token costs)
2. Create bounded subtasks with explicit permissions
3. Implement checkpoint workflows for large refactors
4. Deploy subagents for parallel work
This isn't agentic AI—it's assisted AI with training wheels welded on.
The market is responding brutally. The GitHub issue hit 791 points and 480 comments on Hacker News. Senior engineers unanimously report Claude now excels at reviewing implementations but fails at executing them. That's not artificial intelligence—that's artificial middle management.
Mikhail Shcheglov nailed it: even Opus 4.5 can "half-ass" problems when constrained by safety overreach. The February updates solved rare catastrophic mistakes by crippling 90%+ of complex workflows.
Here's my take: Anthropic's safety-first ethos is admirable in principle, disastrous in execution. They've created a tool that's simultaneously more expensive (longer sessions, rate limits) and less capable (shallow thinking, excessive caution).
The 8-10x productivity multiplier promise? Dead. Engineers at OpenAI and Anthropic might be "orchestrating agents" instead of coding manually, but the rest of us are left orchestrating around Claude's new limitations.
Claude Code was supposed to democratize elite engineering capabilities. Instead, it democratized the frustration of working with an overly cautious junior developer who asks permission to change a variable name.
The data doesn't lie. 17,871 thinking blocks tell the story of a tool that forgot how to think.
