Two AI workflows,
one overdue loan.

I built the Jupiter EMI assistant and the payment result screens in two days. AI helped in two big ways. This page shows both, with real before and after screens.

2days
39commits across 2 repos
16fintech apps studied
4directions per screen
4payment states shipped
Workflow 01

The copy engine

How it works

  1. Look. An AI agent read money screens from 16 real fintech apps on Mobbin.
  2. Write the rules. What it found became a 10-rule copy guide. Every rule links to a real screen that proves it.
  3. Check every screen. No screen ships until its words pass the guide. The git log shows every pass.
Beforesoft words, fuzzy limits
Options screen before the copy pass
Afterclear stakes, exact limits
Options screen after the copy pass

What changed, word for word

Missed payments happen. Let's sort this one out. Avoid late fees and a credit score drop. says what is at risk
Move your due date by up to 5 days Up to 5 days instantly, or more with approval tells the real limit
Pause payments for 1 to 3 months Up to 3 months off, credit score protected leads with the result
Warning: Penalty charges up to ₹300 may be applicable! Pay by 10 Aug to avoid late fees. real date, no siren
Don't ignore this! Missed payments can severely damage your credit score. Protect your CIBIL score by paying your EMI by 10 Aug. the way out, not the threat
Beforethe words the guide bans
Overdue screen wearing the copy the guide bans
Afteroverdue screen, shipped
Overdue screen after

Left screen: the same build wearing the copy the guide exists to kill. "Oops", "URGENT", "Warning", a fee that "may be applicable", and shouting buttons. Right screen: what shipped. Same layout, different words, different feeling.

Same rules, next project

The payment result screens follow the same guide from day one. Success, partial, failed and prepay each tell the truth: exact amounts, real dates, calm words even when the payment fails.

Payment successful state
Success
Partial payment state
Partial
Payment failed state
Failed
Prepayment state
Prepay
Workflow 02

Four directions before one pixel

How it works

crazy-four is a skill I built. It makes the AI pitch four different visual takes on every screen, then helps me compare them: what each one does well, where each one breaks.

  1. Make four. No screen gets designed until four different takes of it exist.
  2. Build them for real, then weigh them. The four docks were working builds on one page, all animating. Each carries its own pros and cons, written on the pitch itself.
  3. Pick one. One wins. Three die. The winner gets all the polish.
The four pitchesA · Live Agent Console  B · Command Pill  C · Loan Island  D · Split Console
Four dock directions compared side by side

All four run the real colors, type and motion. Static mocks only show layout. Running builds show behavior, so the pick is honest.

The winner, shipped

The pick: one dark dock at the bottom that belongs to the assistant. It holds the status line and the actions. When a human joins, the same surface goes live: the reviewer sees your case, Meera joins, and reply chips carry the talk.

Winnerthe dock, holding the actions
The shipped winning direction with the dock
Winner, livea human joins on the same surface
The shipped build live with the relationship manager joined

References

Everything used to build both projects, 9 to 10 Aug 2026.

Repositories

  1. [1]TequilaTejas/jupiter-emi — the overdue EMI assistant. 38 commits in two days. Live at tequilatejas.github.io/jupiter-emi.
  2. [2]TequilaTejas/emi-success — the four payment result screens. Live at tequilatejas.github.io/emi-success.

Skills, authored

  1. [3]fintech-voice — the 10-rule money copy guide from the Mobbin research. Start with a verb, use exact numbers, no drama words.
  2. [4]crazy-four — my ideation skill. Four takes per screen, compared on pros and cons before anything gets built out.
  3. [5]stop-slop — removes AI-sounding writing from every line.
  4. [6]i-have-adhd — keeps output short, numbered and action-first.

Skills, imported

  1. [7]emilkowalski/skills — 9 design and animation skills: animate, animation-vocabulary, apple-design, emil-design-eng, find-animation-opportunities, improve-animations, pick-ui-library, prototype, review-animations.

Tooling

  1. [8]Claude Code — Fable 5 plans and reviews. Sonnet and Opus agents write the code.
  2. [9]Mobbin MCP — searches real app screens and flows. The research behind the copy guide.
  3. [10]Playwright MCP — clicks through the builds and screenshots every state. Every image on this page came from it.
  4. [11]Agentation v3.0.2 — lets me leave notes right on the prototype. The notes go back to the agent as edit requests.
  5. [12]tmux + git worktrees — runs many agents at once, one branch each.

Design references

  1. [13]Wise design language — the visual base: heavy black type, lime on ink, sage neutrals. Tokens live in DESIGN.md.
  2. [14]Revolut — a full second look built in Revolut's style, then set aside.
  3. [15]VOICE-TONE-GUIDE.md — the copy guide inside the repo. Each rule links to a real screen. Bad examples included.

Apps studied via Mobbin

CRED · Wise · Monzo · Revolut · Chime · KOHO · Afterpay · Zip · Cleo · PayPal · Toss · KakaoBank · Lloyds · GoPay · Turo · Credit Karma