[Feature] Claude Code Discuss Mode

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/1946
2. https://github.com/anthropics/claude-code/issues/2058
3. https://github.com/anthropics/claude-code/issues/4764

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

I just ran into exactly this -- I want a gentler way to interrupt than hitting escape, like "finish the edit you're making and then let's chat." Plan mode doesn't do this because there's no way to say "yes that plan looks good but I have a question about a line you just wrote" and hitting escape cancels an edit that might itself look fine (which it would then have to re-express after discussion).

Plan mode doesn’t exist to ask questions or have a discussion, it exists to
create an action plan and usually doesn’t involve any sort of
communication, it always has an action at the end.

This doesn’t fit the use case we want, we don’t need any specific type of
code or action.

I agree with the comment.

On Thu, 14 Aug 2025 at 22:54, Jacob Rothstein @.***>
wrote:

jbr left a comment (anthropics/claude-code#5787) <https://github.com/anthropics/claude-code/issues/5787#issuecomment-3189846700> I just ran into exactly this -- I want a gentler way to interrupt than hitting escape, like "finish the edit you're making and then let's chat." Plan mode doesn't do this because there's no way to say "yes that plan looks good but I have a question about a line you just wrote" and hitting escape cancels an edit that might itself look fine (which it would then have to re-express after discussion). — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/5787#issuecomment-3189846700>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AX3T4JJWWLKEBHUUPLBJOD33NTZR7AVCNFSM6AAAAACD52FNY6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTCOBZHA2DMNZQGA> . You are receiving this because you authored the thread.Message ID: @.***>

Title: Feature Request: Add a 'Chat' or 'Q&A' Mode for Non-Agentic Conversation

Is your feature request related to a problem? Please describe.
Yes. Claude Code is heavily optimized for being an agentic coding assistant. The internal system prompt and available tools strongly guide it towards taking action: planning tasks, searching the codebase, implementing solutions, and verifying them.

The System Prompt explicitly states:

You are an interactive CLI tool that helps users with software engineering tasks. Use the instructions below and the tools available to you to assist the user. ... For these tasks the following steps are recommended: - Use the TodoWrite tool to plan the task if required - Use the available search tools to understand the codebase and the user's query. - Implement the solution using all tools available to you - Verify the solution if possible with tests.

This action-oriented design is powerful but can be counterproductive when the user's goal is simply discussion, brainstorming, or understanding a concept without any intent to modify the codebase. In these scenarios, Claude Code may still try to formulate a plan or use tools unnecessarily, which consumes tokens, adds latency, and doesn't align with the user's conversational intent.

Describe the solution you'd like
I propose the introduction of a new "Chat Mode" (or "Q&A Mode" / "Discussion Mode"). This mode would be explicitly non-agentic and conversational.

Behavior of Chat Mode:

1. Restricted Toolset: In this mode, Claude Code would have its tool access strictly limited to read-only operations.

• Allowed Tools: Read, Grep, LS, Glob, WebSearch, WebFetch.
• Disabled Tools: Bash, Edit, MultiEdit, Write, NotebookEdit, Task, TodoWrite, KillBash.

1. Shift in Intent: The underlying system prompt or a prompt amendment for this mode should instruct Claude to act as a knowledgeable conversational partner for Q&A and brainstorming, rather than an agent that plans and executes tasks. It should understand that the goal is explanation, not implementation.
2. User Experience: When in Chat Mode, Claude would answer questions, explain code snippets (using its read-only tools), and discuss concepts without ever attempting to create a to-do list, formulate a multi-step implementation plan, or ask for permission to modify files or run commands.

This could be implemented as a new permissionMode in settings.json and exposed via a slash command like /chat or /ask for easy toggling within a session.

Describe alternatives you've considered

1. Using plan mode: The current plan mode is the closest alternative. As per the iam.md documentation, it "disables file modification and command execution." However, its primary purpose is still to create an action plan that the user will approve. The user is still prompted to exit plan mode to begin coding. This is different from a purely conversational mode where no action plan is expected.

1. Prompting Claude to not take action: A user can try to start their prompt with "Don't use any tools, just answer my question..." but this is unreliable. The core system prompt is so heavily weighted towards action that it often overrides these user instructions, leading to a frustrating experience. A dedicated mode would provide a much more consistent and efficient conversational experience.

Additional context
The need for this mode is directly evident from the current system prompt's design, which is focused on a "Doing tasks" loop. A dedicated chat mode would provide a much-needed escape from this agentic loop for purely exploratory and educational interactions, making Claude Code a more versatile tool for all stages of the development process, not just the implementation phase.

I agree with all of that, but can't imagine how folks implement without constantly interrupting to say "hey can you explain that line? I'm pretty sure it's a bug" (or dead code, or unnecessarily defensive programming, etc). This isn't just an educational/conversational thing, it's a pair programming implementation concern. If you wait until after all the edits are done, mistakes can cascade and you spend far more time iteratively chasing fixes. Constantly hitting escape and canceling edits is also wasteful and is an inadequate solution

The idea is not to implement anything, rather just focus on understanding and making the hard part, which is reading code, easier.

This is also an issue, I also have that part Minimal Defensive Programming in my CLAUDE.md but yes it's very hard to put guardrails on the LLMs because at the end of the day it only evaluates text and these definitions are very hard to get right, what's minimal, what's the proper, correct code. Most of the clean code and best practices are subjective, you have things like DRY don't repeat yourself but does that mean we will make everything and env variable or bundle two print statements if they are the same.. very difficult topics..

By closing #2020 because it's similar you might loose the important fact that this exact feature worked very well from start. When plan-mode arrived in CC it was great. We (me and Claude) would iterate many rounds until a perfect todo list was created. Then I would switch out of plan mode with cmd-tab. The current plan-mode has two major flaws imo:

• Claude constantly nags to get out of plan mode
• When I tell him that we need to discuss further he keeps proposing the same implementation without listening to me.

Why closing an issue that was created long before this one? I could not help but feel like this is a lazy way to avoid feedback

The original release of plan mode was superior to what we have now.

'plan mode' should just be 'read only' mode. When I want to distill the conversation into a plan after multiple turns of refining context, I would like to prompt it to do so. I've stopped using plan mode all together because of the current UX.

It includes both changes and restrictions to the prompt and the Claude
Permissions Tools.

On Tue, 2 Sep 2025 at 03:02, kylesnowschwartz @.***>
wrote:

kylesnowschwartz left a comment (anthropics/claude-code#5787) <https://github.com/anthropics/claude-code/issues/5787#issuecomment-3243487062> The original release of plan mode was superior to what we have now. 'plan mode' should just be 'read only' mode. When I want to distill the conversation into a plan after multiple turns of refining context, I would like to prompt it to do so. I've stopped using plan mode all together because of the current UX. — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/5787#issuecomment-3243487062>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AX3T4JIMLCLR4OHEASHBA633QTUCDAVCNFSM6AAAAACD52FNY6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTENBTGQ4DOMBWGI> . You are receiving this because you authored the thread.Message ID: @.***>

I would love this too. Claude is far too eager to rush ahead, present a plan, and get on with implementing it.

I want to do analysis and have a discussion first, without the fear that Claude might start going of and starting to work on things, and without constantly being presented with plans to reject.

A discussion mode, or pre-plan mode would be great. Or just make plan mode less keen to present a plan until asked (especially if explicitly asked not to provide a plan until asked to).

I want to have and use this daily. Especially on codebases that are not mine, to be able to reason about the possible changes, discuss the implementation... would be a valuable addition.

This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.

still an issue

This issue has been automatically closed due to 60 days of inactivity. If you're still experiencing this issue, please open a new issue with updated information.

still an issue

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.