You send a built-in read-only research agent to survey a codebase, and its findings quietly ignore a naming convention your project records in its instructions file. You would like to point that out and have it carry on from where it got to. What is actually going on, and what can you do next?

The principle — a worker optimised for speed is a worker that was told to skip things, and what it skipped is the first place to look when its output is subtly wrong.

Why the key is correct — two documented facts have to be joined, and each rules out one instinct. Anthropic states the skip and its reason together: Explore and Plan skip your CLAUDE.md files and the parent session's git status to keep research fast and inexpensive. That is a deliberate trade for a fast, read-only agent optimized for searching and analyzing codebases, not a fault. Then the recovery: the built-in Explore and Plan agents are one-shot and return no agent ID, so they can't be resumed. Nothing can be continued, so the convention has to go into the brief of a fresh run — which works, because the brief is the one channel that reaches a worker that reads no project files.

Why the others are wrong — the first assumes the file was read and its guidance out-argued, when it was never loaded. The second reads a documented trade as a broken setup, which sends you looking for a misconfiguration that does not exist. The last gets the diagnosis exactly right and then reaches for a resume, which is the half of this that catches careful readers: a correct understanding of why the output is wrong does not tell you whether the run can be salvaged, and here it cannot.

Remember this — the fast built-in researchers know only what you put in the brief, and you get one shot at that brief.

Sources — Anthropic's Claude Code subagents documentation.