Read, Write, Edit, Bash, Grep and Glob: which one for what

Claude Code and the Agent SDK come with built-in tools for working with files and commands. The exam doesn’t ask about their parameter names, but about which one to pick in each situation and how to combine them.

What each one does

Tool	What for	Example
Grep	Content search: what the file says	Who calls processLegacyOrder, where an error message appears, a key matching AKIA[0-9A-Z]{16}
Glob	File path pattern matching: what the file is called	Every test with **/*.test.tsx
Read	Read a whole file	Open the file that registers a route
Edit	Targeted change replacing a unique text	const PORT = 3000; to const PORT = 8080;
Write	Write a whole file	Create a file or rewrite it after reading it
Bash	Run commands	Run the tests or the build

The short rule: Grep is what it says; Glob is what it’s called. Searching for tests with Grep returns any file that mentions “test”; searching for a function’s callers with Glob only finds files with that name.

Edit and its plan B

Edit needs unique anchor text: it replaces a fragment that appears only once. If timeout: 30 appears in three sections and only the payments one must change, Edit fails. The plan B the exam guide names is Read + Write: load the whole file, change only what’s needed and write it back in full. In practice you can also widen the text with more context, but in the exam the answer is Read + Write.

The reverse is a trap too: if the text is unique, Edit is the right call. Rewriting 900 lines to change one risks accidental changes.

Exploring without reading everything

To understand a 47-file module, reading them all first fills the context with irrelevant code. Incremental exploration starts from an anchor:

1. Grep for something concrete: a route string (/api/v2/invoices), a class name, an error message.
2. Read the file that comes up.
3. Follow the imports and the flow, reading only what you need as you go.

Finding every usage

A single Grep for the original name isn’t enough when there are re-exports under another name or import aliases. If a “barrel” file exports processOrder as submitOrder, consumers that import submitOrder never contain the text processOrder. The search takes several steps:

1. Grep for the definition and its re-exports.
2. Read those files to collect every exported name (fmtMoney, fc).
3. A new Grep for each name.

The usual traps

• Grep to find files by name, or Glob to search content.
• Reading the whole repository “for context”: the main anti-pattern.
• sed or replace_all when Edit fails on repeated text: they change every occurrence, including the ones that had to stay.
• Modifying the file to make the text unique, or restructuring code to dodge a tool limitation.
• find or grep through Bash: they work, but the exam expects the built-in tool made for the job.
• Writing from memory with Write without reading first: it destroys the rest of the file.

In the exam

Identify what’s being searched for: text inside files, or a naming pattern? A change with a unique or a repeated anchor? Are there aliases or re-exports in the scenario? Those three questions give you the tool sequence.