[FEATURE] Skill management is fragmented and confusing

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

Managing skills in Claude is currently a complete mess.

Two overlapping systems for the same job

It's confusing when to use Cowork scheduled tasks vs Claude Code routines:

• For an end user they're almost the same: a recurring task with a trigger.
• Scheduled tasks aren't nice to manage. Routines offer improvements (e.g. renaming).
• Now with routines, there are two separate places to create recurring tasks.
• Routines can't properly access Cowork skills.

Skill storage is fragmented across several locations

Skills can live in ~/.claude/skills/, ~/.claude/commands/, Cowork installed plugins, or inlined in routine prompts. None is reachable from all execution contexts (interactive chat, Cowork chat, local routine, remote routine).

The result: a routine asked to use a skill installed via claude plugin install refuses with "skill not in the available list". This forces full skill bodies to be inlined into every routine, killing the single-source-of-truth benefit of skills.

End-user impact

It's gotten so complex that it's not properly usable for end users in non-technical roles. I have friends in customer success and sales who've given up - the maintenance overhead of skill management has become too painful for the value it delivers to them.

Proposed Solution

1. One source of truth for all skills, stored on Anthropic's servers and mirrored locally on the user's devices (like iCloud), reusable across every Claude session type - Chat, Cowork, local routine, remote routine
2. One place to create recurring tasks, unifying Cowork scheduled tasks and Claude Code routines into a single management surface.

Alternative Solutions

Alternative Solutions / Workarounds tried

None of the workarounds I've tried fully solve this - each covers only one execution context:

1. Local marketplace (~/.claude/super-skills/ registered via claude plugin marketplace add): works for local Claude Code chat, invisible to Cowork and routines.
2. Inlining the full skill body into each routine's Instructions field: works, but means maintaining the same skill in N+1 places. Drift is inevitable.
3. /create-cowork-plugin: closest to a "central" solution, but requires opening a Cowork session, manually packaging, and clicking install through the UI. There's no programmatic install API, and even after install it's unclear whether routines pick it up.
4. Custom bash scripts (~/.claude/bin/sync-cowork-mirror.sh, sync-my-skills.sh, diff-cowork-mirror.sh) to mirror Cowork plugins and session-output drafts into a browsable local folder and detect drift between local and Cowork copies. Read-only — doesn't help routines.
5. Telling a local routine to read the SKILL.md from disk: the routine refuses, citing "I cannot load skills from arbitrary file paths."

The fact that I needed to build three custom shell scripts and still don't have a working single-source-of-truth setup is itself a sign the current model isn't working for end users.

Priority

High - Significant impact on productivity

Feature Category

File operations

Use Case Example

I built a skill that syncs my newly accepted LinkedIn connections into my Attio CRM and emails me a summary. I want it to:

1. Run on a schedule (Mon–Fri, 11am and 6pm) — even when my laptop is off.
2. Be runnable on demand in any Cowork chat ("hey Claude, run my LinkedIn sync now").
3. Be editable from one place so I don't have to update three copies when I tweak the workflow.

What it took today:

• Authored the skill at ~/.claude/super-skills/plugins/gtm/skills/check-linkedin-accepted-requests-update-in-crm/SKILL.md.
• Couldn't reference it from the routine — had to copy the entire skill body into the routine Instructions field.
• To make it work in Cowork chats and remote routines, I'd also need to package and install it as a Cowork plugin via /create-cowork-plugin — a third copy.
• Result: 3 copies of the same workflow. Edit one, the other two silently drift.

What it should take: write the skill once, tag it as "recurring on this schedule," done.

Additional Context

_No response_