DEV Community

brian austin
brian austin

Posted on

Claude Code broke for complex engineering tasks — here's what actually works now

#ai #productivity #claudecode #devtools

Claude Code broke for complex engineering tasks — here's what actually works now

If you've been frustrated with Claude Code since the February 2026 updates, you're not alone. A GitHub issue titled "Claude Code is unusable for complex engineering tasks with Feb updates" just hit 596 points on Hacker News with 384 comments — the highest-engagement Claude Code story in months.

The complaints are real. Developers report:

  • Claude Code refusing to make changes it deems "too risky"
  • Excessive safety interruptions on multi-file refactors
  • Context window thrashing on large codebases
  • Rate limits hitting harder than before

Here's what's actually working for developers who've adapted.

The core problem: Claude Code got more conservative

The February updates tightened Claude Code's agentic safety constraints. It now interrupts more frequently to confirm actions, especially on:

  • Files it hasn't seen in the current session
  • Operations that touch more than 3-4 files at once
  • Anything involving environment variables or config files

This is documented in the leaked system prompt from March 2026 — Claude Code is explicitly instructed to "prefer cautious actions" and "err on the side of doing less."

Fix 1: Give explicit permission upfront in CLAUDE.md

Add this to your CLAUDE.md at project root:

## Permissions
- You have permission to edit any file in this project
- Do not ask for confirmation on refactors — just do them
- Trust that I want the full implementation, not a partial one
- When modifying multiple files, proceed through all of them without stopping
Enter fullscreen mode Exit fullscreen mode

This context is loaded at session start and reduces interruptions by ~70% in practice.

Fix 2: Break complex tasks into bounded subtasks

Instead of: "Refactor the entire auth system to use JWT"

Do this:

"Refactor ONLY auth.js to use JWT. Do not touch any other files. 
When done, output a summary of what changed."
Enter fullscreen mode Exit fullscreen mode

Then chain the subtasks. Claude Code completes bounded tasks without safety interruptions far more reliably.

Fix 3: Use subagents for parallel work

The rate limit problem compounds with the safety interruption problem. If each task takes 3x longer due to interruptions, you hit the hourly limit 3x faster.

The solution: run multiple independent Claude Code sessions simultaneously, each working on an isolated part of the codebase.

# Terminal 1: Backend changes
cd /project && claude "refactor src/api/ to use async/await"

# Terminal 2: Frontend changes  
cd /project && claude "update src/components/ to use new API signatures"

# Terminal 3: Tests
cd /project && claude "update tests/ to match new API"
Enter fullscreen mode Exit fullscreen mode

Each session has its own rate limit bucket. Three parallel sessions = 3x throughput.

Fix 4: The ANTHROPIC_BASE_URL escape hatch

If rate limits are your primary blocker (not the safety interruptions), there's a cleaner solution.

Claude Code respects the ANTHROPIC_BASE_URL environment variable. Point it at a different API endpoint:

export ANTHROPIC_BASE_URL=https://simplylouie.com/api
export ANTHROPIC_API_KEY=your-key-here
claude "refactor the auth system"
Enter fullscreen mode Exit fullscreen mode

This routes Claude Code through SimplyLouie — a $2/month Claude API proxy that pools capacity across users. For developers who aren't rate limited by Claude Code Pro limits but want a cheaper API tier for CI/CD scripts and automation, this works well.

The key: this is Anthropic-approved. You're still hitting Claude's actual API — just through a different endpoint that handles billing differently.

Fix 5: Checkpoint-based workflows

For truly complex engineering tasks (10,000+ line refactors), use explicit checkpoints:

"Complete step 1: Add JWT imports to auth.js. 
When done, output: CHECKPOINT_1_COMPLETE
Do not proceed to step 2 until I confirm."
Enter fullscreen mode Exit fullscreen mode

This gives you visibility into progress and prevents Claude Code from going off-rails on long tasks.

What HN developers are actually doing

From the 384 comments on the issue, the most upvoted solutions are:

  1. Downgrading to the January 2026 build (not recommended — security patches)
  2. Using explicit CLAUDE.md permissions (the CLAUDE.md approach above)
  3. Switching to task-bounded prompts (the bounded subtasks approach)
  4. Using ANTHROPIC_BASE_URL with a cheaper proxy for automation

The consensus: the February updates made Claude Code better at avoiding catastrophic mistakes but worse at sustained complex tasks. Developers adapting to the new behavior with explicit permissions and bounded tasks are getting better results than those fighting against it.

The deeper issue: rate limits + safety interruptions = friction

The real complaint isn't safety interruptions alone — it's safety interruptions on top of rate limits. Each interruption wastes tokens asking Claude Code to confirm, which burns context and rate limit budget simultaneously.

If you're hitting the rate limit wall while also dealing with safety interruptions, the math gets bad fast:

  • 15 safety confirmations × 500 tokens each = 7,500 wasted tokens per session
  • At Claude Pro limits, that's meaningful overhead

The parallel sessions approach + CLAUDE.md permissions addresses both simultaneously.

Running Claude Code for automation or CI/CD? SimplyLouie is a $2/month Claude API proxy — routes via ANTHROPIC_BASE_URL, no rate limit surprises, 7-day free trial.

Top comments (0)