[DOCS] Direct contradiction in Explore agent tool scoping between internal system prompt and developer documentation.

Documentation Type

Unclear/confusing documentation

Documentation Location

Discovered by comparing the Claude Code v2.1.4 internal system prompt definition of the Task tool against the "Create custom subagents" documentation page.

Section/Topic

The Explore and general-purpose agent definitions in the Task tool toolset.

Current Documentation

• System Prompt says: "Explore: ... (Tools: All tools)" and "general-purpose: ... (Tools: *)"
• Developer Documentation says: "Explore: ... Read-only tools (denied access to Write and Edit tools)" and "General-purpose: ... All tools"

What's Wrong or Missing?

The system prompt (the instructions given to the model) claims that the Explore subagent has access to "All tools." However, the public-facing documentation explicitly states that Explore is a read-only agent denied access to Write and Edit. This creates a fundamental misunderstanding for the AI:

• The AI may attempt to delegate write/edit tasks to Explore believing it has the tools to perform them.
• Users relying on the documentation will expect Explore to be safe/read-only, whereas the AI believes it is a full-privileged agent.

Suggested Improvement

Standardize the permission definitions. If Explore is intended to be read-only (as the documentation states), the system prompt must be updated to reflect this restriction so the parent agent correctly understands the subagent's capabilities. If Explore is intended to have all tools (as the prompt suggests), the documentation must be corrected.

Impact

High - Prevents users from using a feature

Additional Context

This inconsistency likely leads to "Tool not found" errors or failed delegations when the parent agent tries to use Explore for modifications based on its internal prompt instructions, only to be blocked by the actual system-level permissions described in the documentation.