[FEATURE] `claude-md-management` plugin should support ongoing knowledge ecosystem maintenance, not just one-off CLAUDE.md additions

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

The claude-md-management:revise-claude-md plugin is designed around a one-off, additive model:
it finds CLAUDE.md and appends lines to it. For teams using Claude Code continuously, this creates
two compounding problems:

1. Bloat — every session adds to CLAUDE.md with no pruning, consolidation, or staleness checks.

Over time CLAUDE.md becomes a dumping ground, consuming prompt tokens on every session regardless
of relevance.

1. Misrouting — for projects with a skill directory (e.g. .agents/skills/ or .claude/skills/),

the plugin has no awareness of that structure and routes everything to CLAUDE.md regardless of
whether a more specific, scoped file would be more appropriate. Domain knowledge that should only
load when relevant ends up always-loaded instead.

A side effect also is that with multiple sources offering their own solutions, there's often confusion of what should and shouldn't go into claude.md and skills, and how they should be written. This creates frustration with the ecosystem, and reduces trust.

Proposed Solution

Extend claude-md-management to support ongoing knowledge ecosystem maintenance, not just one-off
additions:

1. Routing judgement — detect existing skill/knowledge directories and route learnings to the

appropriate domain file rather than always targeting CLAUDE.md

1. Coverage check — before proposing an addition, check whether it is already covered in any

skill or instruction file

1. Pruning / review mode — a mode that reviews existing content for staleness, duplication, or

misplacement and proposes removals or moves, not just additions

1. Regular cadence support — make the plugin useful as a recurring practice (e.g. end-of-sprint

knowledge review), not just a one-time setup step

Side note: this may mean it is not really claude-md management per se - it may be a new plugin altogether.

Others do try to fill this gap, but they are not authoritative and may sometimes be doing best guess efforts as to what is needed; you guys know!

Alternative Solutions

• Users can manually audit CLAUDE.md and skill files periodically — but this relies on discipline

and domain knowledge that a well-designed plugin could provide automatically

• Third party plugins and skills, like evaluate-skills skills can help partially fill the gap, but these are not authoritative and no guarantee they will keep to up to date with the evolving landscape
• Related: #23522 (on-demand plugin loading) and #15281 (project-level skill overrides) address

adjacent symptoms of the same underlying issue

Priority

High - Significant impact on productivity

Feature Category

Configuration and settings

Use Case Example

A team runs revise-claude-md at the end of a session. The plugin identifies a code pattern or improvement that is worth capturing. The plugin proposes adding this to CLAUDE.md.

The correct home may sometimes be a scoped skill file — only loaded during work that uses that programming language for example. But the plugin has no way to know that file exists. The addition goes into CLAUDE.md instead, where it loads on every session including frontend-only and infrastructure sessions where it is irrelevant.

Multiplied across many sessions and many learnings, CLAUDE.md becomes bloated and its signal-to-
noise ratio degrades.

Additional Context

There is currently a significant amount of confusing and contradictory community advice about what
belongs in CLAUDE.md versus skill files versus memory versus agent instructions. Users piece
together guidance from blog posts, Discord threads, and trial and error.

claude-md-management has the potential to be the authoritative, Anthropic-endorsed answer to
this question — an opinionated, maintained tool for knowledge ecosystem hygiene that the whole
community can rely on. Right now it leaves that potential unrealised.

Possibly related issues I search for but aren't directly the same suggestion as what I am making: #23522, #15281, #37066