Create local skills as shortcuts — makes real /commands in .claude/skills/. Use when user says "create shortcut", "create skill", "make a command for", "add shortcut", or wants a quick custom /slash-command. Also lists and deletes local skills. ALSO triggers on "Unknown skill", "skill not found", or any unrecognized /slash-command — auto-creates it on the fly.
SKILL.md
/create-shortcut - Local Skill Factory
Create real local skills (.claude/skills/<name>/SKILL.md) that show up as /commands in autocomplete.
Usage
/create-shortcut # list local skills
/create-shortcut list # same, with numbers
/create-shortcut list --mine # only stamped skills (installer: create-shortcut)
/create-shortcut list --untagged # only skills with no installer: field
/create-shortcut create deploy "Run tests then deploy"
/create-shortcut delete deploy # delete by name
/create-shortcut delete 3 # delete by number
/create-shortcut --cleanup # interactive bulk-archive of --mine skills
/create-shortcut --cleanup --auto # auto-archive anything stamped + unused 30d+
How It Works
Creates a SKILL.md in .claude/skills/<name>/ (project-local) or ~/.claude/skills/<name>/ (global with --global).
The skill immediately appears in / autocomplete after creation.
⚡ Local Skills (.claude/skills/)
1. deploy Run tests then deploy to prod
2. lint-fix Fix all linting errors
3. db-migrate Run database migrations
⚡ Global Skills (~/.claude/skills/)
4. trace (v3.4.8) [core] Find projects, code...
5. recap (v3.4.8) [core] Session orientation...
...
Delete local: /create-shortcut delete <name or number>
Mark core (arra-oracle-skills-cli installed) skills with .
Mark skills (with in frontmatter) with plus a relative-age hint (e.g. ).
Local skills with no field have no tag (origin unknown).
Installs
0
[core]
stamped
installer: create-shortcut
[mine]
2d ago
installer:
Filter flags
/create-shortcut list # default — all local + global, marks [core] / [mine]
/create-shortcut list --mine # only skills with `installer: create-shortcut`
/create-shortcut list --untagged # only skills with NO `installer:` field (origin unknown)
For --mine and --untagged, the listing additionally shows:
created_at age (e.g. "2 days ago", "3 weeks ago") — pulled from the frontmatter created_at: field for --mine; absent for --untagged.
Approximate usage count — derived by greping the command name across ~/.claude/projects/*/*.jsonl session files (read-only, no stored counter — counts stay computed, not persisted).
Example --mine output:
⚡ Your skills (installer: create-shortcut)
1. deploy [mine] 2d ago · used 7× Run tests then deploy
2. read-and-deep-... [mine] 1d ago · used 1× Read a target and deep-analyse...
3. resonance [mine] 3w ago · used 24× Capture a resonance moment...
Bulk-archive: /create-shortcut --cleanup
Implementation hint: parse each SKILL.md frontmatter (top YAML block between --- lines) and check for the installer: key. --mine keeps only installer: create-shortcut; --untagged keeps only entries with no installer: key at all. Core skills (installer: arra-oracle-skills-cli ...) are excluded from both filtered views.
Write SKILL.md with a provenance stamp in the frontmatter (so the skill can later be filtered via --mine and bulk-cleaned via --cleanup):
---
name: <name>
description: <description>
installer: create-shortcut
created_at: <ISO 8601 timestamp with timezone, e.g. 2026-05-14T04:15:00+07:00>
created_session: <first 8 chars of $CLAUDE_SESSION_ID, optional but nice>
---
# /<name>
<description>
## Step 0: Init
Chain date with the first real command — never call date alone (saves 1 tool call):
\```bash
date "+🕐 %H:%M %Z (%A %d %B %Y)" && <first-real-command-here>
\```
## Instructions
<Ask user what the skill should do, or generate from description>
---
ARGUMENTS: $ARGUMENTS
Generating the stamp values:
# created_at — ISO 8601 with local timezone offset
CREATED_AT="$(date -Iseconds 2>/dev/null || date +%Y-%m-%dT%H:%M:%S%z)"
# created_session — first 8 chars of session id (skip silently if unset)
CREATED_SESSION="${CLAUDE_SESSION_ID:0:8}"
These three fields (installer, created_at, created_session) are always written for new skills created via this mode. They are the audit trail — without them, a skill created here is indistinguishable from one created by hand.
After creating, confirm:
✅ Created /<name>
📁 .claude/skills/<name>/SKILL.md
📝 <description>
Try it: /<name>
With --global flag
/create-shortcut create deploy "Deploy to prod" --global
Creates in ~/.claude/skills/ instead of .claude/skills/.
Mode 3: Delete
/create-shortcut delete <name or number>
Find the skill (by name or list number)
Show its content
Ask confirmation: "Delete /? (yes/no)"
If yes:
Move to trash (Nothing is Deleted): mv .claude/skills/<name> .claude/skills/.trash/<name>_$(date +%Y%m%d_%H%M%S)
Create .trash/ if needed: mkdir -p .claude/skills/.trash
Confirm: "Archived: / → .claude/skills/.trash/"
If no: "Kept: /"
Only delete local skills. Never delete global/core skills — warn instead:
⚠️ <name> is a core skill (installed by arra-oracle-skills-cli).
Use 'arra-oracle-skills uninstall -s <name>' to remove it.
Mode 4: Auto-Create (catch unknown commands)
When the agent encounters an unknown /slash-command (e.g. "Unknown skill: resonance", "skill not found: push-further", or any unrecognized /command), this skill activates automatically.
Flow
Parse the command name from the error or user input (e.g. /resonance → resonance)
Infer intent from the command name + current conversation context
/resonance → "capture a resonance moment to ψ/memory/resonance/"
/push-further → "challenge the current approach, suggest improvements"
/deploy-staging → "deploy the project to staging environment"
Execute immediately — do what the user intended, don't block on skill creation
Offer to save: after executing, ask:
Save as /command for next time? (yes/no)
If yes → create a minimal SKILL.md stub using Mode 2 (create), pre-filled with:
Name from the command
Description inferred from what was just executed
Instructions based on the action taken
Provenance stamp — same as Mode 2: installer: create-shortcut, created_at: (ISO 8601 + tz), created_session: (first 8 chars of $CLAUDE_SESSION_ID). Auto-created skills must be stamped too — that's how --cleanup later finds them.
If no → done, one-shot execution only
Intent Inference Rules
Treat the command name as a natural language hint: split on -, read as phrase
Use conversation context (last few messages) to disambiguate
If intent is truly ambiguous, ask ONE clarifying question before executing
Never refuse — always attempt something reasonable
Example
User: /resonance
Agent: [sees "Unknown skill: resonance"]
→ Infers: "capture a resonance moment"
→ Creates ψ/memory/resonance/<timestamp>.md with context
→ "Save as /resonance for next time? (yes/no)"
→ User: "yes"
→ Creates .claude/skills/resonance/SKILL.md
→ "/resonance is now a real command"
Key Principle
"You think it, you slash it, it exists."
Skills create themselves from usage. The user never hits a dead end.
Mode 5: Cleanup (interactive bulk archive)
/create-shortcut --cleanup
Bulk-archive skills you created via /create-shortcut (anything stamped with installer: create-shortcut). Six months of accumulated one-shot skills, gone in one gesture — without touching [core] skills or hand-made externals.
Flow
Scan both .claude/skills/ (local) and ~/.claude/skills/ (global). Read each SKILL.md frontmatter and keep only entries with installer: create-shortcut.
Enrich each candidate with:
Age — relative form of created_at (e.g. "2 days ago", "3 weeks ago"). If created_at is missing, show "age unknown".
Last invoked — grep the command name (e.g. /deploy) across ~/.claude/projects/*/*.jsonl and report days since the most recent match (e.g. "used 4d ago", "never used").
Description — first non-empty line of the description: field, truncated to ~60 chars.
Multi-select picker using @clack/promptsmultiselect (same library + style as the existing skill picker elsewhere in the CLI). Each row:
◯ deploy [local] 2d ago used 4d ago Build, test, deploy to prod
◯ resonance-lite [global] 3w ago never used Capture small resonance moments
◯ pr-review [local] 6mo ago used 5mo ago Review the current PR with checklist
Pre-checked rows: anything never used OR last-used >30 days ago. User can toggle freely before confirming.
Archive selected skills to .trash/<name>_<unix-timestamp> in their respective skills directory (Nothing-is-Deleted, same convention as Mode 3):
✅ Archived 7 skills.
Recover any of them via:
mv .claude/skills/.trash/<name>_<ts> .claude/skills/<name>
If nothing matches the scan (no stamped skills found), print:
No skills with `installer: create-shortcut` found.
Tip: only skills created after the provenance-stamp release show up here.
Older user skills can be retroactively migrated with `--migrate` (future).
Flag: --cleanup --auto
/create-shortcut --cleanup --auto
Skips the interactive picker and archives everything matching the auto-prune rule:
installer: create-shortcut (stamped by us), AND
created_at is older than 30 days, AND
last invocation is older than 30 days (or never invoked).
Default behavior is OFF — you must opt in with --auto. Designed for power users who trust the heuristic and want a one-keystroke sweep. Still archives to .trash/ (recoverable, never destroyed).
Boundaries
Never touches skills with installer: arra-oracle-skills-cli ... (those are [core]; use arra-oracle-skills uninstall instead).
Never touches skills with no installer: field (those are [untagged] — origin unknown, not safe to assume the user made them via this skill).
Always moves to .trash/, never deletes outright. Recovery is a single mv away. Nothing is Deleted.
Examples
/create-shortcut create deploy "Build, test, and deploy to Cloudflare Workers"
/create-shortcut create db-seed "Reset and seed the development database"
/create-shortcut create pr-review "Review the current PR with checklist"
/create-shortcut create morning "Run standup + check inbox + show schedule"
