# Fixing the broken shell execution in Claude Code > Source: > Published: 2026-09-09 15:58:43+00:00 # Fixing the broken shell execution in Claude Code I spent four hours last Wednesday fighting a permissions loop that almost made me uninstall [Claude Code](/en/tags/claude%20code/) and go back to a standard IDE plugin. I was trying to automate a migration of a legacy Express.js project to TypeScript, and the CLI kept hitting a wall when trying to run `npm install` after updating the `package.json`. The fix wasn't in the prompt, but in how I handled the tool-use permissions. If you're using [Claude](/en/tags/claude/) Code and it keeps asking for permission to run the same command three times, or fails with a cryptic shell error, you're likely hitting the same wall I did. ## Why did the shell keep failing during the migration? I was running `claude` in a directory with a mix of root-owned files (remnants of a Docker volume mount). When Claude tried to execute a shell command to verify a file change, it hit this: `Error: EACCES: permission denied, open '/Users/dev/project/node_modules/.bin/tsc'` The annoying part? Claude didn't just tell me it failed. It tried to "fix" it by suggesting `chmod` commands, which then failed because the CLI session didn't have sudo privileges. I was stuck in a loop where the AI was trying to be helpful, but the environment was locked. I realized the bottleneck wasn't the AI's logic—it was the session's execution context. I had to kill the process, fix my local folder permissions via `sudo chown -R $USER .`, and restart the CLI. The lesson here is that Claude Code is an agent with a shell; it's not just a chat box. If your environment is messy, the agent will spend more tokens trying to fix your OS than fixing your code. ## How to actually get a migration started without the loop If you're looking for a Claude Code tutorial that isn't just "type these commands," start with a targeted scope. Don't just say "convert this to TypeScript." That's how you end up with a broken `tsconfig.json` and a thousand linting errors. Here is the exact sequence I used to successfully migrate my controllers after fixing the permission bug: 1. **Index the project first.** Let it read the files. 2. **Create a specific migration plan.** Ask it to list every file that needs a `.ts` extension. 3. **Execute in batches.** I found that asking it to do 5 files at a time kept the context window clean. When I tried to do 20 files in one go, it started hallucinating imports from files it hadn't actually converted yet. | Approach | Result | Token Cost (Est.) | Success Rate | | :--- | :--- | :--- | :--- | | "Convert everything" | Massive merge conflicts | High | 40% | | Batch of 5 files | Clean compiles | Medium | 90% | | File-by-file | Tedious but perfect | Low | 100% | ## Dealing with the "Too Many Tool Calls" bottleneck Another wall I hit was the tool-call limit. While working on a complex refactor, Claude started looping—reading a file, searching for a string, reading the file again, and then searching for the same string. It looked like this in the terminal:`Reading file...`` Searching for 'UserSchema'...``Reading file...`` Searching for 'UserSchema'...` This is a waste of money and time. I stopped the execution and shifted my approach toward [AI Coding](/en/category/aicoding/) patterns that emphasize "Plan first, execute second." I told it: "Stop searching. Read the entire `models/` directory into your context once, then tell me where the schema is." Suddenly, the loop stopped. The trick is to realize that the agent sometimes forgets it already read a file if the conversation gets too long. Forcing a "consolidated read" saves you from the repetitive tool-call cycle. ## Where to find better patterns than the default docs The official docs are fine for installation, but they don't tell you how to handle the "AI drift" that happens during a three-hour coding session. This is where I actually found value in the PromptCube community. Instead of guessing how to structure a request for a complex refactor, I started looking at [Prompt Sharing](/en/category/prompts/) threads. I found a specific pattern for "Architectural Guardrails" that I now paste into my initial session: *"Before modifying any file, describe the change in one sentence. If the change affects more than three files, stop and ask for confirmation."* This one line prevents Claude Code from rewriting half your project because it decided to rename a variable globally. ## Recommended setup for a smoother experience If you're just starting, don't just run it in a random folder. Set up a dedicated environment. - **Version Control:** Never run Claude Code on a clean branch. Always have a`git checkout -b ai-experiment` branch. It will delete things. It's a fact. - **[MCP](/en/tags/mcp/) Servers:** If you need it to interact with your database or Jira, set up the Model Context Protocol (MCP) servers first. Trying to "teach" the AI how to use a tool via chat is inefficient; giving it a formal MCP connection is a game-changer. - **Resource Management:** Check out the[Resources](/en/category/resources/) section of the community to see which MCP servers are actually stable. Some of the community-made ones crash the CLI every ten minutes. `EACCES` or `Permission Denied`, stop the AI, fix the OS, and then resume. Don't let the AI try to fix your Linux permissions unless you've given it a very specific, narrow scope to do so. [Next AlphaGenome Atlas is mapping 3 billion base pairs to predict DNA →](/en/threads/9040/) [these AI tool field notes](https://tanyan888.com/), with plenty of directly applicable cases.