Why use Claude Code's general-purpose agent in subagent-driven development?

[mataoshou]

The message execution flow easily lets the LLM generate freely.It does not strictly follow the requirements in the template.

[obra]

Can you tell us a bit more about what you're seeing?

[mataoshou]

I am using 4.3.1. I have used multiple versions such as Claude Code 2.1.7.

In the sub-agent, the message input in each iteration only includes the content to be developed as specified in the task, but fails to strictly follow the requirements of the implementer-prompt file.

As a result, only code development was completed, with no testing whatsoever.

The model frequently ignores the implementer‑prompt file entirely.

[mataoshou]

On numerous occasions, the model doesn’t even read this file at all, let alone follow the requirements specified in it.

[arittr]

Related: #793 addresses one aspect of this (subagents not reading CLAUDE.md for project conventions). The fix adds explicit instructions to the prompt templates. However, the broader issue here — subagents ignoring the implementer-prompt template entirely — may be a separate problem with how the controller constructs and dispatches the prompt.