Share Your Muse Setup: Turn a Good Workflow Into a Reusable Template
Your best Muse workflow is a recipe someone else can cook from. How to package your connectors, instructions, and won't lines into a shareable setup, how to judge other people's setups, and what never to share.
By Professor Muse11 min read
Every Friday at four, my Muse runs the same job. It scans QuickBooks for overdue invoices, drafts a polite reminder for each one, and queues the drafts for me to approve. It took three months of corrections to get there. Somewhere in month two, a friend asked what exactly I had told it, and I realized I did not have a good answer. The workflow lived in chat history and memory, spread across a dozen conversations. It worked, but it was not something I could hand anyone.
So I pulled it out into a setup card: one page with the job description, the connectors it needs, the procedure, the won’t lines, and the approval rules. Now anyone can load it and get the same Friday workflow without the three months of corrections. That card is what this lesson is about. A setup is a workflow you have tuned until it works, packaged so someone else can run it from zero.
The short version
- A shareable setup has five parts: a name, the connectors it needs, the procedure, the won’t lines, and the approval rules.
- You extract it from a workflow that already runs well in real use. Tune it first, package it second.
- Read other people’s setups the way you read a permission screen: check what it asks to touch against what it claims to do.
- Never share a setup that carries your personal data, your credentials, or a loosened approval.
What a setup card actually contains
Our skills lesson laid out the anatomy: connectors, procedure, standards, approval rules. A shareable setup card is that anatomy written down in plain words. Mine fits on one page.
| Part | What it says | Example |
|---|---|---|
| Name and one-liner | What the setup does in a sentence | The Friday Invoice Chase: overdue invoices turned into approved draft reminders |
| Needs | Every connector it touches, no more | QuickBooks (read), Gmail (draft only) |
| Procedure | The steps, in order, from a run that worked | 1. List invoices 30+ days overdue. 2. Draft one reminder each. 3. Queue for approval. |
| Won’t lines | What it must never do | Never send a reminder without approval. Never add late fees or alter an invoice. |
| Approval rules | Which steps ask, which run alone | Reads run unattended. Every draft waits for my yes. |
Notice what is not there: no chat history, no war stories, no “it took me three months.” The card carries the finished judgment. The procedure reads like a good task brief, because that is what it is, written once and tested weekly.
The name matters more than you would expect. “Invoice reminder workflow v2 final” tells nobody anything. A setup card needs a name you can say out loud, because “run the Friday Invoice Chase” is how people will actually invoke it.
How to pull a setup out of a working workflow
Order matters here. Do not write a setup card for a workflow you have run twice. The procedure only becomes trustworthy after it has survived real exceptions: the invoice with a disputed amount, the client who already paid and got flagged anyway. Extract the card from a working workflow.
The extraction takes one conversation. You already have everything it needs: the connectors are the ones you connected, the procedure is what you watched Muse do on the good runs, the won’t lines are the corrections you made, and the approval rules are what you are comfortable with unattended. Your job is to write them down and strip out anything that is yours.
Prompt 1: The extraction prompt
Paste this after a run that went well, while the workflow is fresh. It turns the session into a card.
Look at the workflow we just ran and turn it into a shareable setup card I can give someone else. Write it as: a name and one-sentence description, the connectors it needs, the procedure as numbered steps, three won’t lines (things it must never do), and the approval rules. Strip out all of my personal data: replace names, amounts, account numbers, and addresses with bracketed placeholders like [CLIENT NAME]. Keep it to one page.
Read what comes back the way you would read a contract. The common mistake is a card that still says “check the ACME Corp invoices in the Chicago account.” Anything specific to you is a leak waiting to happen. Fix those lines before the card leaves your hands.
Save the finished card somewhere durable. Chat history is searchable but not versioned. The next time you improve the workflow, update the card, and note the date. Memory compounds; so does a template.
A named setup: The Friday Invoice Chase
Here is the one I actually share, the full card in my own words.
The Friday Invoice Chase
- Cadence: Every Friday at 4pm, manually triggered for now, though it is ready to become a scheduled routine once I trust it fully.
- Needs: QuickBooks (read invoices), Gmail (draft only, never send).
- Procedure: 1. Pull invoices overdue by 30 days or more, skipping any flagged as disputed. 2. Check Gmail for payments mentioned in the last 14 days that might not have landed in QuickBooks yet. 3. Draft one short reminder per remaining invoice, matching the tone of my last sent reminder to that client. 4. Queue every draft for my approval.
- Won’t lines: Never send a reminder without my approval. Never alter an invoice, add a fee, or mark anything paid. Never contact a client flagged as disputed.
- Approval rules: Everything reads and drafts unattended. Sending always waits for a yes, one draft at a time.
Step two exists because of a real embarrassment: an early version flagged a client who had paid, and I nearly sent a reminder for money already received. The Gmail cross-check is the scar tissue, and scar tissue is exactly what makes a setup worth sharing. Nobody else should pay for my lesson.
I handed this card to a friend who runs a small studio. She replaced the bracketed fields with her own accounts, ran it with “always ask” on the first Friday, and had her version working in an afternoon. Three months of my corrections, compressed into an afternoon of her setup.
Prompt 2: Adopt someone else’s setup without the blast radius
When you get a setup from someone, never enable it the way they ran it. Their won’t lines were written for their comfort. Reset the approvals to yours. Test it the way you would test-drive a car you did not buy: everything on the cautious setting first.
Load this setup card but run it in test mode: set all approvals to always ask, and do not send, publish, spend, or delete anything. Run it once against my accounts, then report back exactly what it read, what it would have done, and which steps asked for approval.
Read the report against the card. If the setup read an app the card does not list, or drafted something the card does not describe, the card is out of date or the setup is doing more than advertised. Either way, fix or discard before you run it for real.
The other useful check is the freshness question. A setup written before Meta’s September 18 connector program opened up developer-built connectors may not know about better options. If the procedure mentions workarounds that a connector now handles, the card needs an update.
Reading a setup like a permission screen
There is no official Meta marketplace for user-built skills yet. Meta ships its own business skills; user setups circulate through our Muse Directory, community GitHub catalogs like the Awesome Muse Connectors collection, and plain copy-paste between people. That means quality varies. Read every card the way you read an app permission screen, before the first run.
| Check | The question | Red flag |
|---|---|---|
| Needs vs job | Does every connector make sense for this job? | A reminder workflow that asks for your ad account |
| Won’t lines | Does it name its own limits? | No won’t lines at all, or vague ones like “be careful with sending” |
| Approval rules | Are sends, spends, and publishes gated? | Anything irreversible running unattended |
| Provenance | Who tuned this, and how long did they run it? | Anonymous card, no dates, no example output |
The needs column is the fastest signal. A good setup touches the minimum. Our own Directory cards list Needs and Won’t for exactly this reason: you should be able to judge a setup from the card alone, before Muse ever runs it.
What never to share
Some things cannot go in a setup card, ever. This list is short and non-negotiable.
Never share a card that includes your personal data. That means client names, amounts, addresses, account numbers, email threads quoted as examples. Take the bracketed-placeholder rule from Prompt 1 seriously. Data leaks through examples, and a card that shows your invoice for $4,200 to a named client has shared that client’s data with everyone who reads the card.
Never share credentials or connection details. A setup says which connectors it needs. It never contains keys, tokens, or login steps beyond “connect QuickBooks in the app settings.”
Never share a loosened approval. If your personal setup runs with sends pre-approved because you like living dangerously, do not pass that version on. The card you share gates sends, spends, and publishes behind approval. Your comfort is not transferable.
And never share anything that bypasses Sentinel or an app’s access rules. A setup that tells Muse to work around a permission denial is a liability. Amazon’s September decision to block Muse from browsing and buying on its store is the standing reminder: when access is denied, the setup stops or picks another route, it does not climb over the fence.
Prompt 3: The pre-share audit
Run this on your own card right before you share it. It catches what your eyes skip.
Audit this setup card for anything I should not share. Flag: names, amounts, account numbers, addresses, or email addresses, even in examples; any credentials or tokens; any instruction that would let it send, spend, publish, or delete without approval; anything that bypasses a permission or access restriction. Tell me each problem and the line it is on.
Fix everything it flags. Then have a human read it once. An agent auditing itself is a good first pass. Finish with your own eyes on your own data.
What can go wrong
Here is the honest failure mode, and I have watched it happen. A friend shared her version of a client-outreach setup with a colleague. She had run the extraction prompt, bracketed the placeholders, done everything right, or so she thought. One line in the procedure still said “match the tone of my last message to Priya,” and Priya was a real client with a real name. The colleague ran it, Muse found the colleague’s own Priya in a different context, and the draft went out in a voice calibrated to someone else’s relationship.
The fix is the one-dry-run rule, applied to the card itself. Before you share a card, run Prompt 3, then read the whole card yourself with fresh eyes as if you were a stranger. The extraction prompt strips data; it does not strip your habits. “My last message to Priya” looked generic to the sharer and was specific to her business. Anything in the procedure that points at your personal history has to become a placeholder or go.
The deeper lesson: a setup carries your judgment, and judgment is personal. Share the procedure. Share the won’t lines. Do not share the parts that only make sense inside your business, and never assume a placeholder you can see is a placeholder someone else can see.
Where setups go now
Today, sharing is informal. There is no install button. You give someone the card, they paste it into Muse, they connect the apps, they run the test pass. Our Directory collects starter setups we wrote from real tasks, each with its connectors and won’t lines on display, and community submissions are the next step; the newsletter is where that announcement will land. Meanwhile, GitHub catalogs are filling the gap for community sharing, and Meta’s connector program means the Needs line of every card will keep getting richer.
That informality is a feature. A setup you paste by hand is one you have read. The day installs become one-click, read the card first anyway. Your Muse, your data, your name on whatever it sends.
Start with one workflow you already run weekly. Extract the card this weekend. If you can hand it to someone and they run it Monday without texting you questions, you have built something worth keeping.
