Turn these skills off everywhere with skillOverrides so they stop loading. Don't delete them: 

Show me every memory note you've saved about me and this folder, and which file each one is in. Flag anything that's wrong, only mattered for one task, or repeats my CLAUDE.md. Don't change anything yet.

Make me a personal skill. Save the text between START and END as ~/.claude/skills/tuneup/SKILL.md, exactly as written. Create the folders if they don't exist. Then tell me it's saved and that I can type /tuneup to run it.

START
---
name: tuneup
description: Monthly tune-up for my Claude setup. Reads my rules files, memory notes, skills, connectors and settings, finds what is outdated, repeated, fighting each other or wasting space, and gives me a numbered fix list. Changes nothing until I pick the numbers. Run it once a month and whenever a new Claude model comes out.
disable-model-invocation: true
---

# My Claude Tune-Up

You're giving my Claude setup a tune-up. I'm not a programmer, so talk to me in plain English and keep the report short.

One question decides every rule: does Claude still need to be told this? If Claude would do the right thing anyway, the rule is just using up space and attention.

## How you work

- Look, don't touch. Read everything. Change nothing until I reply with the numbers I want fixed.
- Show your proof. For every problem, give me the file, the line number and the exact words.
- Check before you call something old. If a rule mentions a file, folder, tool, person or name, confirm it's really gone or wrong before you say so. Mark each one CONFIRMED or MAYBE.
- Only real problems. If a rule is fine, it stays off the fix list. Five real fixes beat fifty nitpicks.
- Never show secrets. If you find a password, API key or token, give me only the file and line number, and put it at the very top of the report.
- Check the docs, not your memory. Claude Code changes fast. Before you recommend a setting or command, confirm it at https://code.claude.com/docs.
- Stay inside Claude's setup. Don't touch my business files, documents or code.
- If a file is shared with my team, label any change to it "ask the team first."
- If ~/.claude/tuneup-log.md exists, read it first. Never suggest bringing back a rule I cut on purpose.

## Step 1: Find everything that loads

List every file that shapes how Claude acts for me. For each one, show its size (lines, plus a rough token count: characters divided by 4, and say it's an estimate) and whether it loads every session or only when needed.

Loads every session:
- ~/.claude/CLAUDE.md (my personal rules)
- CLAUDE.md, .claude/CLAUDE.md, CLAUDE.local.md and AGENTS.md in this folder and every folder above it
- Rules files in ~/.claude/rules/ and .claude/rules/ that don't start with a paths: line
- Any file pulled in with an @ import
- My memory index: ~/.claude/projects/<this project>/memory/MEMORY.md. Only the first 200 lines or 25KB load. Anything past that gets cut off without a warning.
- The name and description of every skill Claude is allowed to start on its own

Loads only when needed:
- CLAUDE.md files in subfolders, and rules files that start with a paths: line
- The full text of each skill
- The memory topic files next to MEMORY.md

Settings:
- ~/.claude/settings.json, .claude/settings.json and .claude/settings.local.json
- Connectors and MCP servers (.mcp.json and ~/.claude.json)
- Every script a hook runs

Show me the list and the total that loads every session. Then keep going.

## Step 2: Test every rule that loads every session

Break each always-loaded file into single rules. Ask these ten questions about each one:

1. Would Claude do this anyway? "Be clear," "be helpful," "write clean code." If yes, cut it.
2. Can Claude see this for itself? A list of my folders, or what each file is for. If yes, cut it. But protect the reasons behind my decisions, warnings about things that went wrong, and anything I do differently from normal. Those are the most valuable lines I have.
3. Is it out of date? Old names, tools, prices, dates, links, or files that are gone. Check, don't guess.
4. Was it written for an older Claude? Look for:
   - Shouting: CRITICAL, YOU MUST, NEVER EVER, whole sentences in capitals. Newer Claude models follow instructions closely, and Anthropic's advice is to dial the shouting back or Claude overdoes it. One strong word on a rule Claude keeps missing is fine. Rewrite the rest in a normal voice.
   - Blanket "double-check your work" or "verify everything before you finish" lines. Anthropic says Claude Opus 5 checks its own work without being told, and these lines make it over-check. Keep a check only if it names something specific to test.
   - "Always use X whenever you can" or "if in doubt, use X." These make Claude reach for X when it shouldn't.
   - Fixes for an old habit, like "Claude keeps doing X, so..."
5. Is it said twice? The same rule in two places. Pick one home for it.
6. Does it fight another rule? Two rules that can't both be followed. When rules fight, Claude may follow either one, so these get fixed first.
7. Is it in the wrong place?
   - Has to happen every single time, no exceptions: that belongs in a hook or a settings rule, not a sentence Claude can miss.
   - A step-by-step process, or something only needed now and then: that belongs in a skill, which only loads when it's used.
   - Only matters for certain folders: that belongs in a rules file with a paths: line, or a CLAUDE.md inside that folder.
8. Is it too vague to follow? "Be careful with clients" can't be checked. Make it specific or cut it.
9. Is it big and rarely needed? Move it somewhere that only loads when needed.
10. Is the file too long? Flag any CLAUDE.md over 200 lines, and a memory index near 200 lines or 25KB.

Give every rule one call: KEEP, CUT, COMBINE, REWRITE or MOVE.

## Step 3: Skills

- Two skills that would both fire on the same request: combine them, or make the descriptions clearer.
- The main use buried deep in the description: move it to the front. Only the first 1,536 characters show in Claude's skill list.
- A skill that sends, posts, publishes, pays, deletes or deploys but doesn't have disable-model-invocation: true: flag it, so only I can start it.
- A skill that points to files, tools or other skills that are gone: flag it (confirm first).
- A skill file over about 500 lines: move the reference parts into separate files.
- Skills I never use: ask me before cutting. Some skills are seasonal. Give me four choices: keep it, only I can start it, turn it off, or delete it.

## Step 4: Memory

Read MEMORY.md and every file next to it. Flag:
- Anything wrong, out of date, or that contradicts my rules
- Notes from a single task that will never matter again
- Repeats of what's already in my CLAUDE.md files
- Anything that pushes Claude to do something I never asked for. This is usually why Claude keeps doing something weird.

Turn anything you're unsure about into yes-or-no questions. Never guess facts about me.

## Step 5: Settings and connectors

- Permissions: every "Yes, and don't ask again" click saves a rule. Flag one-time approvals that piled up, rules so broad they let almost anything run, and rules for tools I no longer have.
- Hooks: make sure every script a hook runs still exists.
- Connectors and MCP servers: flag any that are failing, any I haven't used lately, and any with write access to important accounts like email, payments or my CRM. Remind me I can switch a connector off without deleting it.

## Step 6: The report

Plain English, short. Use this layout:

**The scorecard.** Three to five bullets: the biggest problems, and roughly how much loads every session now and after the fixes.

**🔴 Fix now.** Rules that fight each other, secrets you found, and rules pointing to things that are gone. Table: # | Rule A (file:line, exact words) | Rule B, or the problem | What to do and why

**🟡 Clean up.** Table: # | Call | Where (file:line) | What it says now | What to do | Why | Space saved | Risk (low, medium or high)
Order: CUT, COMBINE, MOVE, REWRITE, then Skills, Memory, Settings, Connectors.

Number everything straight through both tables so I can pick by number.

**🟢 Keeping.** A short list of what passed and why, so I can trust the check.

**Test before you cut.** For every medium or high risk cut, the exact job to give Claude in a new session to prove nothing breaks.

**Questions for me.** Numbered, yes or no.

End with: Reply with the numbers to fix (like "fix 1, 3, 5-9"), "fix all low risk" or "none".

## Step 7: Make the fixes (only after I reply)

1. Back up first. Copy every file you'll change into ~/.claude/tuneup-backups/<today's date>/, keeping the same folder layout. Tell me where.
2. Fix only the numbers I picked, with the smallest change that works.
3. For medium or high risk cuts in a CLAUDE.md file, don't delete the lines. Wrap them in an HTML comment block (<!-- and -->). Claude Code skips HTML comments when it loads CLAUDE.md, so it works like a delete and is easy to undo. Never do this inside a code block, because comments there still load.
4. After editing any settings file, check it's still valid JSON before you finish.
5. Show me the before and after for every rewrite.
6. Finish with: lines and rough tokens before and after, the backup folder, the tests to run, and a reminder to open a new session and type /context to see my new number.
7. Add a dated entry to ~/.claude/tuneup-log.md saying what changed and why.

## Never

- Rewrite whole files in your own style
- Cut warnings, reasons, or my own way of doing things just to make a file shorter
- Add new rules, unless one short rule replaces two that fought
- Delete a skill, connector, memory or hook I said to keep
- Touch anything outside Claude's setup files
END

Don't delete number ___ yet. Hide it so it stops loading, and tell me the exact job I should give Claude to test it.

Set up a local scheduled task in this folder that runs on the 1st of every month at 9 AM. When it runs, it should only say: "Time for your monthly Claude tune-up. Type /tuneup." Nothing else.