How to build an invoice and quote builder with OpenAI Codex

Rule of thumb: if you send a handful of invoices a month and want to be paid by card tonight, rent one, and Wave or Zoho Invoice will do it for nothing but the card fee or their logo on your PDF. If invoicing is part of a product you sell, if your quotes carry revisions and your PDFs carry your brand, or if you bill in several currencies and do not want them converted for you, the software is worth owning, because the caps above are exactly where a growing business pays and your clients’ names and amounts are yours to keep in a database you control. The honest middle case is a business that needs to be paid online. Build this and connect a payment rail, and read the block on choosing how you get paid, because that fee exists whether you rent the software or own it.

* On this template the free-tier line to watch is the pause, not storage. The only file is a logo and the rows are small, but a free project is paused after a week without activity, and a business that invoices at month end leaves it idle for three. Pro, from $25/mo, ends the pause and keeps a daily backup for 7 days, which matters more here than anywhere else in the catalogue, because an invoice is a record you may be required to keep.

Prices and rates from supabase.com, developex.com and learn.chatgpt.com, checked September 2026.

Nothing here is worth memorizing. These four just need to exist on your machine. From step 01 on, you say what you want and Codex runs the commands.

1. 01

   Write AGENTS.md before anything is scaffolded

   Codex loads AGENTS.md at the start of every task, so the three rules that matter most on this build (every row has an owner, money is never converted, the AI key stays on the server) are loaded before any code exists to break them.

   PromptSet up the projectCopy

   Start by writing an AGENTS.md at the project root: the stack is React 18, Vite, TypeScript, Tailwind and the Supabase JS client, and this project is an invoicing and quoting app. Fix the vocabulary (clients, quotes, invoices, line items, templates, tax rates) and record three standing constraints: every row belongs to the signed-in account and the database enforces that rather than the screen, money is stored as numbers in each document’s own currency and never converted, and any AI key lives in a server-side function and never in the browser. You load that file automatically at the start of every later task. Then scaffold the project: a typed Supabase client under src/lib reading VITE_SUPABASE_URL and the publishable key (the sb_publishable_… key that replaced the older anon key) from .env, with .env confirmed present in .gitignore.

   The three constraints are the ones worth writing down. Filtering by user in the component, converting currencies with a made-up rate, and calling the provider from the browser are all reasonable-sounding shortcuts, and each one costs somebody money.

2. 02

   Narrow permissions, then have Codex add sign-in, the schema and the ownership rules

   This task touches authentication and every table, which is exactly the kind of change worth scoping Codex down for and reading the SQL of before it runs.

   PromptAdd auth, tables and ownership rulesCopy

   This task touches authentication, so narrow your permissions to just this one. Wire in Supabase Auth with email and password: sign-up with a full name and a company name, login, logout, a persisted session, a useUser hook and a profiles row per account created by a trigger. Then write the migrations and show me the SQL before running it. Clients with name, email, phone, billing and shipping address, tax ID, account number, contact person and notes. Quotes with a number, a Draft, Finalized or Archived status enforced by a database check and defaulting to Draft, the client’s details, a quote date and a valid-until date, a tax rate, a discount that is a fixed amount or a percentage, shipping, a currency code, notes, terms and a link to the invoice it became. Invoices with the same shape plus a due date, payment terms in days, payment instructions and bank details. A line-items table for each with description, quantity, unit price and a per-line discount. Tax rates, templates (name and settings JSON), settings as key and value JSON, and custom fields scoped to quotes, invoices or both, with a values table for each. Every table carries the owning account, has row-level security on and lets a user reach only their own rows with explicit WITH CHECK clauses, line items included through a function that checks the parent document’s owner. Show me how to verify a second account gets nothing back.

   Reading the SQL here costs ten minutes. Finding another company’s invoice in your list costs a customer.

3. 03

   Give the clients screen its own task, and the quote editor the next one

   The quote editor is the hard centre and the clients screen feeds it. Keep them as two tasks so each change is small enough to read.

   PromptBuild the clients screenCopy

   Build the clients screen: a searchable list and add and edit dialogs covering every client field, with a status of active or inactive. Walk me through the component structure before you build it.

   PromptBuild the quote editorCopy

   Build the quote editor: pick a client from file to fill their details or type them in, add line items with description, quantity, unit price and an optional per-line discount, and show the subtotal, a document discount as a fixed amount or a percentage, a tax rate chosen from my tax rates, an optional shipping line and the total, recalculating as I type. Let me choose the document’s currency from a list of about thirty common ones and format every amount with the browser’s own number formatting for that currency, converting nothing. Add a quote date, a valid-until date, notes and terms, then a quotes list with search, a status filter and a sort, with the quote number as the address of its detail page.

   Test with a percentage discount, a tax rate and shipping together, in a currency with no decimals. Rounding that is right in dollars and wrong in yen is the bug this task exists to find.

4. 04

   Turn a quote into an invoice, and give both a lifecycle

   Converting without retyping, revisions of a quote, sequential numbers and the three statuses. No sent and no paid, because nothing here sends or takes money.

   PromptAdd conversion, revisions, numbering and statusesCopy

   Add a Convert to invoice action on a finalized quote that creates an invoice with the same client, lines, discount, tax, shipping and currency, numbers it INV-1, INV-2 and so on, links the two documents both ways and opens the invoice editor with a due date and payment terms in days added. Add a Duplicate action on a quote that creates a revision numbered from the original (R becomes R/1, then R/2) and make the routes handle a number with a slash in it. Then the statuses Draft, Finalized and Archived on both documents, changed from the detail page, with the database refusing any other value. Build the invoices list and editor to match the quote screens.

   Keep the status list to those three. Adding sent or paid without the machinery behind them is how a status lies to the person reading it.

5. 05

   Build the PDF and branding, then hand the AI analyzer to its own task

   A handful of PDF designs generated in the browser with your branding on all of them, and then the one AI feature, which is self-contained enough to be its own task with its own log and ceiling.

   PromptBuild the PDF designs and brandingCopy

   Generate the PDF in the browser with a PDF library. Build a template gallery with a handful of designs (professional, classic, a ledger style and a couple more), each rendering the document’s header, addresses, line items, totals, notes and terms in its currency. Add branding settings: a logo uploaded to a storage bucket limited to 5 MB and image types, primary and accent colours, a font choice and header and footer text, applied to every design with a live preview. Add a print-ready preview page per document with a Download PDF button.

   PromptAdd the AI template analyzer, with a log and a ceilingCopy

   Create a Supabase edge function called analyze-template that requires a signed-in user, accepts an uploaded image or PDF of an existing invoice as a data URL, sends it to my AI provider with the key held in function secrets, and asks the model, through a forced structured tool call, for a layout specification only: colours, typography, header layout, address sections, line-item columns, how totals are shown, footer and decorative elements. It must not read amounts or line items. Before contacting the provider, write a row to an ai_calls table and check the caller’s calls today against a limit read from settings. Save the specification as a custom template rendered by the same preview and PDF code as the built-in designs. Show me the system prompt you wrote before you run anything.

   Asking to see the system prompt first is worth the round trip. Whether the analyzer stays a designer or drifts into reading amounts is decided in that text.

6. 06
   Destination

   Hand the dashboard, reports and settings to a cloud run, then test it yourself

   The screens around the documents are self-contained and well specified, a reasonable set to hand to a cloud run while you do something else, provided you read the change before it merges.

   PromptAdd dashboard, reports, settings, demo, and testCopy

   Four remaining pieces. A dashboard with counts and totals by status and the most recent documents. A reports page with cards for Draft, Finalized and Archived, a sortable table of every document with client, date, number and total, and CSV and PDF export. An activity page listing recent changes to quotes, invoices and clients from their timestamps. A settings page for the company profile, branding, tax rates, custom fields, numbering and the default currency, plus a demo account seeded on sign-in and wiped on sign-out behind a Try the demo button I can remove before launch. When they are in, run the app and take it end to end yourself: sign up a second account and confirm it sees no documents, create a quote in a currency with no decimals, convert it, download the PDF and check the totals match the screen, hit the analyzer limit on purpose and read what the app says, and fix whatever does not hold up.

Review the access rules (row-level security policies) on every table. For each one, tell me in simple terms who can view, add, edit, and delete records, confirm that people can only reach their own data while the right roles can reach more, and flag anything left open that shouldn’t be.

One rule outranks everything above it: the database service key and the AI provider key stay on the server, and neither one ever goes into the app a browser downloads or into a repository. Should either leak, assume it has been used, look at what your provider says was spent with it, and issue a new one the same day.

Two questions decide this, and neither is about speed. Does the plan permit commercial use, because an app that bills clients is a business and Vercel’s Hobby tier is not for one. And does a saved link to one invoice open that invoice after sign-in, which takes one routing rule on each of these hosts, with Vercel’s already in the template. Nothing on this app is public, so a search engine has nothing to index and that is fine. The host serves one small static folder, and every amount comes from your database.

A percentage has no size until you multiply it by your own month, so do that first. Twenty invoices of $500 paid by card in a month, $10,000 in all, costs roughly $336 through Stripe (the card fee, the 30 cents on each, and the 0.4% for the invoice), about the same through Square on the free plan, and roughly $309 through PayPal at its card rate. Bank details printed on the PDF cost nothing and arrive when the client gets round to it. Whichever you pick, the paid status and the reconciliation are yours to build, because the template has neither.

Open the analyze-template function. Make the model name a setting I can change without a deploy, and set it to a current model on my provider’s price list that accepts images and structured tool calls. Add an ai_calls table that records every model call (which account, which task, which model, tokens in and out, a cost estimate, success or failure) and write to it from the function, successful or not. Then add a ceiling: a limit per account per day and a total across the whole app, both settings I can change without a deploy, checked before the function contacts the provider and returning a friendly “you have reached today’s limit” rather than an error. Do not let the browser send the limit or the cost. Then add a small usage panel to settings showing calls and estimated spend this month.

Add a “Draft from description” box at the top of the quote editor. It sends my description and my last twenty finalized quotes’ line items to the ai function and gets back suggested lines with a description, quantity and unit price each, drawn from prices I have used before where they match. Fill the editor with the suggestions marked as drafts for me to edit or delete, never save anything until I do, and keep the totals recalculating as I change them.

Add an “Overdue” filter to the invoices list for finalized invoices past their due date, and a “Draft reminder” action on each that sends the client’s name, the invoice number, the amount, the currency, the due date and how many days late it is to the ai function and gets back a short, polite reminder in my company’s voice, with a firmer version for anything more than thirty days late. Show it in a box I can edit and copy, and record in the activity feed that a reminder was drafted, since this app does not send email.

Add a second function called read-invoice that requires a signed-in account, accepts an uploaded image, PDF or CSV of an invoice or a price list, sends it to the ai function and asks for the line items (description, quantity, unit price), the currency, the tax rate and the client details it can find, each with a confidence. Show the result as a prefilled draft quote or invoice with every figure the model was unsure about highlighted for me to confirm, and never save until I do. Log it and count it against the same ceiling as the analyzer.

Add a “Summarize this month” button on the reports page that gathers the documents finalized in the selected period, grouped by currency with no conversion between them, and asks the ai function for a short summary: what was invoiced, what was quoted and not yet invoiced, the largest clients by amount, and anything unusual compared with the previous period. Show it in a box I can edit and copy, and make clear which currency each figure is in.

Each prompt above should pick whichever model tier fits the job: a fast, lower-cost model for high-volume work, and a stronger reasoning model for anything that weighs tradeoffs against each other. Codex’s model names change faster than this page does, so check OpenAI’s current model line-up (linked in the references below) before you build, rather than copying a name you saw once. Keep all of it behind that same ai function, so one key and one rule set covers every feature you add.