[FEATURE] Client-side prompt comment (,,) that are stripped before API transmission
Preflight Checklist
- [x] I have searched existing requests and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)
Problem Statement
Currently, there is no way to include personal notes, reminders, or scratchpad text in a prompt without sending it to the model. Everything entered in the prompt is processed, which consumes tokens and can unintentionally influence the model's output or reasoning.
Proposed Solution
Introduce a client-side comment syntax (e.g., lines starting with ",," or something else extremely unlikely to be in a prompt you want processed) in the interactive REPL and single-prompt (-p) flags. The Claude Code CLI should parse and strip these lines locally before sending the payload to the Anthropic API.
If I type:
,, This is just a note to myself about the bug
Fix the off-by-one error in the loop
The CLI should locally strip the comment and only transmit "Fix the off-by-one error in the loop" to the API.
Also useful for ephermal notes-to-self like
,, On hold until 2:32pm due to token depletion
Alternative Solutions
I've been leaving comments on the line without pressing enter, however this interferes with my auto-compact code where claude uses herdr to input "/compact yadayadayada" lines as it sees there is something in the prompt box so does not do the /compact.
Priority
Low - Nice to have
Feature Category
Interactive mode (TUI)
Use Case Example
- Context tracking: Leaving a note for myself about why I am asking for a specific change without paying to process it.
- Token savings: Keeping a local history of thoughts in the prompt box without sending them to the model.
- Preventing model confusion: Ensuring the AI does not try to act on a personal note or get derailed by unstructured brainstorming.
Additional Context
_No response_