ClaudeClaude CodeLevel 4: Multi-tool systems

Should You Tell Claude Code to Explore Before It Edits?

Usually not. Claude Code already gathers context before it acts, and Anthropic's docs say current models find your layout, conventions and test framework faster than you could explain them. Add a read-only step only when a wrong guess would be expensive, or when you want to see what it learned before it touches anything.

Prompt to copy
Don't edit anything yet.

I want to [goal]. Use a subagent to find out how [feature or folder] works today: the files that own it, the pattern I should follow, and how it's tested. Report back with file paths, and separate what you confirmed in the code from what you're inferring.

Here's what you can't see in the repo:
- [a decision the team made]
- [which approach I prefer, if I care]
- [what's out of scope]

Then ask me anything you still can't answer from the code.
cost of a wrong guess →
Read-only step firstplan mode or a subagent
Plan modesee what it learned before edits
Let it reada wrong guess is cheap to undo
Let it readname the file and the symptom
how well you know this code →
A rule of thumb, not Anthropic's chart. Either way, tell Claude what it can't see in the repo.

Try it now

  1. 1Start claude in your project and ask give me an overview of this codebase, then a narrower question such as how is authentication handled?.
  2. 2Press Shift+Tab until the status bar shows ⏸ plan mode on, then describe a real task. Edits stay blocked while it reads.
  3. 3Read what it reports. It worked if the plan names the real files and the pattern you expected. If not, correct it in one sentence now, before any code exists.

Use it when

Ask for a read-only step on unfamiliar code you'll have to defend in review, on a change whose right answer depends on a pattern somewhere else, or on a bug in code nobody on your team wrote.

Skip it when

Skip it for a clear task in code you know. Anthropic's sample follow the pattern in HotDogWidget.php and build a calendar widget names the pattern directly, which is quicker than an exploring step. Delegated exploring isn't free either: each subagent sends its own requests against the same usage limits, and the docs warn that a fresh one can end up slower than working where you are.

Why it works

Claude Code works in a loop the docs describe as three phases: gather context, take action and verify results. Broad searches go to the built-in Explore subagent, which can't write or edit, and only a summary comes back to your conversation. What reading can't reveal is anything that was never written down: a decision from a meeting, which of two approaches you prefer, or how far you want it to go. Anthropic says Claude tends to read a request conservatively and build the smallest thing that satisfies it. So the prompt that helps most tells it what it can't see. Telling it to look adds little.

Example

Say you need rate limiting on a public API and haven't settled the details. Anthropic suggests opening with I want to add rate limiting to the public API but haven't thought it through yet. Interview me until you could write the spec, then write it to SPEC.md. Claude then asks multiple-choice questions on points such as per-user versus per-key limits, and what happens to requests over the limit. Start a fresh session to build from SPEC.md, so the work begins with a clean context and a written definition of done.

Going further

An @ reference skips the read step, but it has limits worth knowing. @src/utils/auth.js puts the file in the conversation when it fits the Read tool's limit, 25,000 tokens by default, and also adds CLAUDE.md files from that folder and its parents. A text file over 256KB isn't included, and @src/components gives you a file listing, not the contents. Your PreToolUse hooks don't run on an @ reference, because no tool call happens. To keep Claude away from a path, use a Read deny rule instead.

Common mistakes

  • Writing a tour of your folders into CLAUDE.md to save it exploring. Anthropic lists this as advice to drop, because Claude reads the repo itself.
  • Saying be thorough or think hard and expecting a deeper read. The docs call those ordinary prompt text. The real control is effort, set with /effort high or /effort xhigh.
  • Delegating to a subagent to get a smarter answer. Anthropic says the same model on the same task gives you parallel work or a fresh pair of eyes, not a better answer.
  • Correcting a third time. After two failed corrections, Anthropic suggests /clear and one better prompt, because the session keeps rereading both failed attempts.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course