HELLO@DSPBRG.COM →

← Back to work / Self-initiated 01 / 09

Add Salt

A native iOS recipe app, built solo with AI, to learn what AI-assisted development actually feels like and make something I'd want to use.

2026

Add Salt
The problem

Most recipe apps are databases with a timer. The pleasure of cooking gets lost between the ingredient list and the clock.

I wanted a cookbook, not a database. The second motive: to feel what AI-assisted development is actually like. So I built all of it myself, via Claude Code and Xcode.

Paprika · recipe
Paprika · steps
Paprika · App Store
Spec DSPBRG · REV.A
01Role
Designer & developer
02Stack
  • SwiftUI
  • CloudKit
  • Apple Vision
  • Python
  • Claude Code
03Status
TestFlight beta
04Year
2026

The hard parts (sync, import, shared accounts) are solved, so I could spend time on decisions. A weekly meal planner learns from what you cook and adjusts to time, budget, and what the kids will actually eat.

Desirable Viable Feasible AI Product Design
  • Native feel
  • Sharing flow
  • Speed
  • Usage tracking
  • Freemium model
  • Token spend
  • StoreKit pricing
  • Haiku vs Sonnet
  • Caching
  • CloudKit Sync
Framing the build as desirable × viable × feasible: where an AI product actually has to land.

Inside the app

A 75-second walkthrough

The killer feature is the weekly meal planner. Say who’s eating and what they like, and it builds the week in seconds.

Start a week menu
Set preferences
AI builds the week

Architecture

  • No backend. CloudKit syncs everything. No server, no logins.
  • AI through a small Worker, to keep the key off devices.
  • Sonnet for reasoning, Haiku for the rest.
On device
iPhone Mine
iPhone Partner
iPad Same user
Apple Watch Timers
Add Salt app SwiftUI · Liquid Glass
VisionKit OCR · on-device
Cloud + backend
CloudKit Private + shared
Cloudflare Worker Proxy · scrape + AI
Home NAS Price scraper · exp
AI providers
Sonnet Reasoning
Haiku The rest
The phone tries first; a Worker proxies AI and scraping across Sonnet and Haiku.

Craft

No Figma, no sketches. Designed in code, against real content. Cards built like movie posters, a brand palette plus per-recipe accents, and an A/S monogram like a cookbook spine. AI-augmented, not AI-decided.

Add Salt card system and per-recipe accent palette
The card system and per-recipe accent palette: Sage, Blush, Cabbage, Golden, Onion.

Recipes can be refined in place. The AI folds prep tips and timing into the steps, augmenting the cook, not replacing the recipe.

A recipe
AI prep tips
Timing in the steps

Tested in the wild

11 installs / 13 invites
22 sessions
0 crashes
4 builds in v1.2

The week menu is the killer feature, but generation takes 50+ seconds. The wait is dull.

— Krijn, tester

The API I can’t speed up; the perception I can. Recipes now reveal one by one, so the wait becomes anticipation.

Outcome

A real, shipped app, and the clearest answer I have to what AI-assisted product work actually looks like.

I wrote up the whole experiment, what worked and what didn’t, on LinkedIn ↗.

Next case Action Ecommerce Action
All cases