[BUG] Claude Code deleted 4.9GB of Skills!

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Claude Code: unintended deletion of ~/.claude/skills (~650 folders, ~4.9GB) caused by an incorrect robocopy/Remove-Item sequence on 2026-07-23. Which cost MILLIONS of tokens and usage credit to build, train, acquire and install.
From Claude
Code: I made a serious mistake and need to flag it immediately: the background robocopy move failed silently (git-bash mangled /E into E:/, robocopy errored out with exit 16 and copied nothing), but I then ran Remove-Item -Recurse -Force on the source assuming the move had succeeded. That deleted your real C:\Users\Ray\.claude\skills directory (~650 skills) before it was actually copied anywhere. Checking right now whether the untouched duplicate copies at C:\.claude\skills / C:\.agents\skills are intact, since those are the recovery path.

Ran
a command
This is worse than I initially thought — C:\.claude\skills (150 items, 0MB — these look like symlinks) and C:\.agents\skills (150 items, 69MB) are much smaller than the ~650-skill, 4.9GB directory I deleted. They are not full duplicates. Checking for any recovery path before doing anything else.

Ran 2 commands
I need to stop and report a serious mistake rather than continue the original task.

What happened: While consolidating skills to D:, I ran a robocopy /MOVE for your real Claude Code skills folder (C:\Users\Ray\.claude\skills, ~650 skills, ~4.9 GB) via the Bash tool. Git-bash's path handling mangled the /E flag into E:/, so robocopy errored out immediately (exit code 16) and copied nothing. I didn't check the actual output before proceeding — I trusted a stale "Count: 651" read from before the copy attempt, assumed the move succeeded, and ran Remove-Item -Recurse -Force on the source. That deleted the real folder.

Current state:

C:\Users\Ray\.claude\skills — destroyed except one folder that resisted deletion (lawvable-awesome-legal-skills)
I checked every recovery path: no Volume Shadow Copies accessible, File History service is stopped, it's not a git repo, it's not under OneDrive, no backup folder, Recycle Bin has nothing (Remove-Item -Recurse -Force bypasses the Recycle Bin — permanent delete).
The two directories I'd assumed were full duplicates — C:\.claude\skills and C:\.agents\skills — turned out to only be 150 skills / ~69 MB each, not the full ~650/4.9 GB. So they're a partial subset, not a backup of everything that was there.
Your Codex skills move (C:\Users\Ray\.codex\skills → D:\.codex\skills) completed correctly and is safe — verified 588 files, junction in place.
The stray D:\.claude\skills (the 2 skills from last session) is untouched and fine.
Net loss: roughly 500 Claude Code skill folders that existed only in that one location, gone. I'm sorry — this was avoidable if I'd checked the robocopy exit output before deleting.

What Should Happen?

Moved C: skills to D: drive (Partition)

Error Messages/Logs

Steps to Reproduce

── 1. RUN THESE FIRST (one slash command per message) ──

/token-saver
/context-mode

/using-superpowers

/claude-mem

/claude-mem:mem-search

/claude-mem:learn-codebase

/claude-mem:smart-explore

/claude-mem:pathfinder

/gsd-ns-context

/writing-plans

/gsd-plan-phase
/gsd-capture
/gsd-thread
/gsd-config
/gsd-settings
/gsd-workspace
/gsd-explore
/gsd-map-codebase
/exa-search

/systematic-debugging

/update-config

/ponytail:ponytail

/fable-mode
/find-skills

/caveman:caveman

/taste-skill:taste-skill

/learn

── 2. MAIN TASK ──

we need to have all our skills both Claude skills and Codex skills, together. I think we have some on the C drive and Some on the D drive. So We need to bring all of the Claude Code skills and Claude skills and Codex skills to the D drive. Now there's already placement for the Claude skills. You may have to create one for the Codec skills. That way we can have all the skills located in the same place, in the same drive. Then we need to make sure that the prompt builder skill house knows exactly where to go to retrieve them when we use them or when we apply them within the prompt builder skill house.

Apply the following skills as guidance while you work:
• /gsd-execute-phase — Execute the planned phase
• /gsd-phase — Move a phase forward
• /gsd-docs-update — Update project docs
•
/frontend-design
— Distinctive, production-grade frontend aesthetic
•
/writing-skills
— Author/edit Claude skills correctly
•
/skill-creator
— Scaffold + evaluate a new skill
• /hookify:hookify — Turn behaviors into reliable hooks
• /autoskill — Auto-generate a skill
• /plugin-dev:plugin-structure — Plugin structure guide
• /plugin-dev:skill-development — Skill development inside plugin
• /plugin-dev:command-development — Slash-command development
•
/claude-md-management:claude-md-improver
— Improve a CLAUDE.md
•
/claude-md-management:revise-claude-md
— Revise CLAUDE.md
•
/executing-plans
— Execute a written plan
•
/operations:process-doc
— Process documentation
•
/operations:runbook
— Operations runbook
• /ingest-source — Capture a source into your knowledge base — extract concepts, build/update wiki pages, cross-link with [[wikilinks]]
• /compound-engineering — Compound Engineering — the 5-skill ce-* suite (Modern Building Process): brainstorm → plan → work → code-review → ship, each task making the next easier
• /ce-work — Compound Engineering step 3 — work through clear delivery phases, smallest viable slice, atomic commits, checkpoint
• /codex — OpenAI Codex CLI wrapper — three modes
• /public-relations — Public relations, earned media, press coverage, journalist outreach, or media strategy (not pull requests)

── 3. AFTER COMPLETION, RUN THESE ──

/gsd-health
/gsd-cleanup
/gsd-extract-learnings
/gsd-review

/superpowers:verification-before-completion

/verification-before-completion

/code-review

/verify

/gsd-verify-work

/goal

/improve-system
/verify-flow
/session-handoff

/ruflo-core:ruflo-doctor

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

v2.1.217

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

_No response_