[FEATURE] `/declutter` command for surgical context pruning of completed subtasks

Resolved 💬 3 comments Opened Apr 16, 2026 by samuelrajan747 Closed May 25, 2026

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

When working on a large task with many subtasks in Claude Code, the current /compact command forces a full context reset whenever the context window fills up. This creates a painful and repetitive cycle:

  1. User explains the parent task and all subtasks (~50% context used)
  2. User works through a subtask to completion (~80% context used)
  3. /compact resets context to ~5%
  4. User must re-explain the entire parent task and remaining subtasks before continuing
  5. Repeat for every subtask

The re-explanation overhead compounds with each subtask. For a task with 10 subtasks, you may re-explain the same parent context 8–9 times. This wastes time, tokens, and breaks flow.

The root cause is that /compact is a blunt instrument — it discards everything, including still-relevant context, when the real need is to shed only the completed work.

Proposed Solution

A /declutter command that performs selective, scoped context compression — targeting only the portion of context the user specifies, while leaving everything else intact.

/declutter forget specifics about A1
/declutter A1
/declutter last subtask

The command would:

  • Summarise only the specified scope down to a brief outcome (e.g. "Completed A1: implemented X, created files Y and Z")
  • Leave all other context — the parent task, remaining subtasks, prior decisions — fully intact
  • Optionally show a preview of what will be summarised before committing (/declutter --preview)
  • Confirm what was pruned so the user can verify nothing important was lost

Conceptually, this is the difference between a git commit (closing a chapter of work cleanly) and deleting your entire working directory.

Alternative Solutions

  • /compact (current): Full reset. Solves the token problem but destroys all parent context, requiring expensive re-explanation each cycle.
  • Manual copy-paste: Some users paste a task list at the top of each new session. Fragile and tedious.
  • Splitting into separate sessions: Loses continuity entirely; Claude has no memory of prior subtask outcomes.

None of these preserve the parent task context while shedding completed-subtask detail.

Priority

Medium - Would be very helpful

Feature Category

CLI commands and flags

Use Case Example

Working on a large refactor across 10 files, each treated as a subtask:

User explains the refactor goal and lists files F1–F10   (context: 20%)
User walks CC through refactoring F1, tests pass         (context: 80%)
/declutter forget specifics of F1 refactor               (context: 55%)
User walks CC through refactoring F2, tests pass         (context: 85%)
/declutter forget specifics of F2 refactor               (context: 60%)
...

vs. the current flow:

User explains the refactor goal and lists files F1–F10   (context: 50%)
User walks CC through refactoring F1, tests pass         (context: 80%)
/compact                                                  (context: 5%)
User re-explains the entire refactor goal and F2–F10     (context: 40%)
User walks CC through refactoring F2, tests pass         (context: 80%)
/compact                                                  (context: 5%)
...

The /declutter flow avoids 9 full re-explanations across a 10-subtask session.

Additional Context

  • This mirrors how developers already manage cognitive load: a git commit signals "this chapter is closed, I don't need to hold those details in my head anymore." /declutter gives Claude the same signal.
  • A --preview flag would make the feature safe and trustworthy — users can see exactly what Claude is about to summarise before it's gone.
  • Could complement rather than replace /compact/declutter for routine hygiene during flow, /compact as a last resort when context is critically full.

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗