8:47 AM - Inbox triage
I open the admin dashboard before coffee. The Slack notification came through at 6:15 AM: three new signups overnight. I don't open Slack yet. I go straight to the dashboard and see the numbers: $742 in MRR added since yesterday, a net addition of two customers (one churn from a pizza place in Austin, which stung), and fourteen menu items queued for processing.
The inbox shows me the raw submissions. I scroll through them:
Marcus Webb at Webb Street Bistro submitted a duck breast dish with a handwritten note asking if I could "make it sound fancy but not pretentious." That's the kind of direction that makes the agent shine. I've trained the system to flag vague requests, so there's a note attached: "Tone mismatch detected. Customer values approval over speed."
Priya Sharma at Sharma's Indian Cuisine submitted six items at once. Her onboarding profile says she runs a family restaurant in Portland, seats thirty, prefers concise descriptions under 35 words. The agent flagged two of her submissions because they contain ingredients it doesn't recognize: "kasuri methi" and "langcha." I make a note to email her back by end of day. The system can't guess what it doesn't know.
The third is from someone new: Carol Reyes at Reyes Family Practice. Except it's not a practice, it's a restaurant she opened with her husband two months ago. Their tagline is "Home cooking you didn't have to make." The submission is clean, simple. Salmon, lemon, dill, pan-seared. The agent processed it fine. No red flags.
I mark Marcus's as "review priority" and move on.
10:23 AM - The flagged conflict
I've pulled up Marcus's three agent-generated descriptions in the dashboard. The system gives me three options per dish, ranked by likelihood of matching the house style.
Option 1 (Confidence: 94%): Pan-seared duck breast with cherry gastrique, charred broccolini, and smoked potato puree. Finishes with sea salt and microgreens.
Option 2 (Confidence: 87%): Duck breast, roasted, with cherry and smoke notes.
Option 3 (Confidence: 79%): Perfectly cooked duck. Cherry gastrique. Potato puree that tastes like something your grandmother would make if she had a smoker.
I'm looking at Option 1, which is technically perfect - probably the one the agent would choose. But there's a flag in red underneath it: "Descriptor creep detected. Four adjectives in first clause. Customer profile suggests moderation."
I remember Marcus's note. "Fancy but not pretentious."
I mark Option 3 as approved and add a note: "This one hits your brief. Ships tomorrow." I send an email:
"Marcus, I had the agent run three takes on your duck. Attached is the one I'm sending to print. I chose it because it leans on specificity (cherry, smoke, the potato angle) instead of stacking adjectives. The other options were technically gorgeous but felt like they were trying too hard for a place that's clearly built on real cooking. Let me know if you want to revise."
I hit send. It's 10:34 AM. I open my Stripe dashboard in a new tab and pull up Webb Street Bistro's subscription. He's on the $35 tier. I've been operating for eight weeks. This is the first time I'm selling someone, not explaining what Saltline might do someday.
12:18 PM - Lunch and the churn pain
I make a sandwich. The Austin pizza place - Rosa's Woodfired - churned yesterday. I'd checked the notes this morning. They'd been on the system for three weeks. Three weeks.
Rosa was a solid customer. She ran a sixteen-table place in East Austin, high foot traffic, posted constantly. She used two batches of descriptions before going quiet. I didn't notice for ten days. When I finally sent a check-in email asking if everything was working, she'd already cancelled. No response to my follow-up.
I'm trying to figure out what happened. Was the output bad. Did the pricing feel wrong for her volume. Did she think it was too hard to use. I'll probably never know. In a SaaS, you live with that uncertainty.
I pull up my metrics sheet. Embedded in my admin dashboard, I can see week-to-date numbers:
- New signups: 12
- Trials converted: 4
- Churn: 2
- MRR: $1,847 (up $156 from last Tuesday)
- Pipeline (warm leads): 23
The pipeline number is doing work. It means there are twenty-three restaurant owners who have signed up for a trial and not yet cancelled. Some will convert. Some won't. The churn will keep happening. That's the part they don't really tell you about when you're building software.
2:02 PM - Customer escalation
Slack lights up. A customer named David Kim just messaged: "Invoice question."
David runs a ramen house in Seattle, been a paying customer for five weeks, tier-one plan at $25/month. He's asking why he was billed $34.99 last cycle instead of $25.
I cross-reference his account. He upgraded his plan mid-cycle to the $35 tier, then downgraded back to the $25 tier. Stripe prorated him, but something about the way it calculated didn't track to what he expected.
I walk through the math in his Slack message. I explain the proration clearly, offer him a $10 credit for the confusion, and ask him to let me know if it still doesn't feel right. He comes back within ten minutes: "Got it, that makes sense. Appreciate you explaining it. The tool is saving us hours though, seriously."
That message sits on my screen for a moment. David is doing real work with the software. He's cutting his menu-writing time from hours to minutes.
3:47 PM - The kasuri methi question
I come back to Priya Sharma's submission. The two ingredients the system flagged are real. They're just not in the training dataset. What I could do is ask Priya to provide translations or explanations. What I actually do is spend twenty minutes researching both ingredients myself, add brief context to the agent prompt, and regenerate descriptions for those two dishes.
Kasuri methi is dried fenugreek leaves. The agent needs to understand that it's a flavor note - bitter, slightly maple-sweet, used in Indian cooking more often in the finishing than the base. I feed that into the system, and when it regenerates, the descriptions are suddenly competent.
It's not scalable to do this for every unfamiliar ingredient. But right now, with fourteen restaurants using the system and a hundred dishes a week flowing through, I can spot-fix the gaps before they become customer complaints. By month four, I'll need to automate this differently. Today, I'm the plugin.
The descriptions that come out are clean:
"Saag paneer with kasuri methi - the dried fenugreek gives it a subtle sweetness that most versions miss."
I send them to Priya with a note: "These two dishes needed some context added to the system. Just shipped the updated descriptions. Curious if they land right for your customers."
5:15 PM - Pipeline and the thank-you
I open Linear, where I track bug reports and feature requests. Three items came in this week:
- Two customers asking for "automatic seasonal refresh" flagging (what the spec called the proactive prompt for spring menu updates).
- One asking for social post scheduling directly to Instagram, not just Meta Business Suite.
- One asking for a "dark mode" in the dashboard.
Dark mode is a rabbit hole I won't chase right now. The other two are patterns. If three people ask for the same thing in a week, it matters.
I also have an email that came in at 3:42 PM from Carol Reyes. Subject: "This is amazing."
She wrote: "I was staring at our menu yesterday feeling like every description sounded generic. Threw the salmon into your system as a test. Read the description you generated and immediately showed it to my husband. He laughed and said, 'That's exactly what we want people to taste when they order it.' We're keeping the salmon. Also just upgraded to the $35 tier and I'm going to submit our full menu next week. Thank you."
I stare at that for a minute. Carol spent two months building a restaurant. She wants the words to match the food. Saltline helped her see that they do.
6:22 PM - Closing the laptop
I check Slack one more time. Marcus replied: "Perfect. This is exactly the voice we want. Running tomorrow's special description through right now."
I close the admin dashboard. The day added $742 in new MRR and lost a customer I should have followed up with sooner. I made one customer's expansion feel easy. I spent forty minutes learning about Indian spices so three descriptions would be better. I answered a billing question clearly and got thanked for it.
The software is real. The work is real. Some of it is me, some of it is the agent, all of it is running a business that lives between the two.
I'm not scaling restaurants' entire operations. I'm solving one specific problem that turns out to matter: the words they put on their menus and their social accounts shape whether customers understand what they're trying to do. When those words are good, the customer sees what the owner sees.
That's worth building.