Next.js + Supabase + Stripe Agent Rules Kit: Cursor Rules, AGENTS.md, CLAUDE.md & Pre-Deploy Audit

Stop your coding agent from shipping the classic Next.js + Supabase + Stripe mistakes. This kit gives Cursor, Claude Code, Codex and other agents a short set of rules for this exact stack, plus a read-only audit you run before you deploy.

What you get:

- AGENTS.md: the one source of truth (stack, commands, layout, security rules that are never optional, a definition of done that requires quoted evidence)

- CLAUDE.md: a thin file that imports @AGENTS.md so Claude Code and other agents can't drift apart

- 5 scoped Cursor rules (.mdc): core working agreement, Supabase RLS and security, Stripe checkout and webhooks, testing

- A 13-point pre-deploy audit prompt. It is read-only: paste it into your agent and it returns a PASS / FAIL / UNVERIFIED table with file:line evidence (secrets in the browser, RLS on every table, grants and policy shape, security definer and views, Server Actions and route handlers, webhook signature verification, idempotency, money and user ids from the server, rigged tests)

- Claude Code permissions.deny for .env files

- A read-only grep helper script that prints the mechanical hits for the audit

- Two tiny sample apps (one with planted flaws) and the ground truth, so you can try the audit on something first

- VERIFY-LOG.txt: the raw output of what we actually ran, what it caught, what it missed, and what we did not test

Based on the current official docs, cited in every file: Next.js data security guide, Supabase Row Level Security guide, Stripe webhooks docs, Claude Code permissions docs.

What we tested (details and raw output in VERIFY-LOG.txt, and the free lite page at https://agentpromptsatoz.com/guides/nextjs-supabase-stripe-rules/):

- The Stripe webhook rule's code was run for real under next dev (Next.js 16.4.3, stripe-node 23.0.1) with a stubbed database: valid signature accepted once, wrong, missing and stale signatures rejected with 400, a repeated event id ignored, and a forged unsigned event accepted by the flawed control route.

- The 13-point audit prompt was run twice on the two sample apps. It found all the planted flaws both times, but on the first run it wrongly passed a placeholder test, so we tightened the wording and re-ran.

- Every file was validated (frontmatter, JSON, shell syntax) and the SQL was syntax-checked with PostgreSQL's parser.

Honest limits:

- It makes an agent much less likely to ship the classic mistakes. It does not make your app secure. Have a human review anything that handles real money or personal data.

- The SQL policies and pgTAP tests were only syntax-checked, not executed against Postgres or Supabase.

- We did not run it inside Claude Code or Cursor themselves, and the audit was tested on a small sample app with one model, not on your codebase.

- Claude Code's permissions.deny covers its file tools and recognized file commands, not every subprocess. Use its sandbox for hard isolation.

- Rules reflect the docs as of 2026-10-23; Supabase, Stripe and Next.js change.

Format: a zip of plain text files you own (Markdown, JSON, one shell script, one Python helper). Works with any agent that reads AGENTS.md, CLAUDE.md or Cursor .mdc rules. No account, no installer.

Refund line (same wording as the site's other packs): if it doesn't help, message us through Payhip for a refund.

Update v1.1 (2026-10-23): added a security-patching rule (checks next and sharp against the patched versions from the Next.js 2026-10-14 security update, no downgrades, lockfile, npm audit before deploy, no SVG uploads, tight image sources), AGENTS.md and CLAUDE.md lines, and check 13 in the pre-deploy audit (now 13 checks). Rules reduce risk; they do not guarantee you are patched.

Update v1.2 (2026-10-23): the security-patching rule now also checks Node.js (fixed in 22.23.4, 24.22.1, 26.12.1) and React Server Components packages (19.0.9, 19.1.10, 19.2.9, 19.3.1) against the official advisories, and adds npm install hygiene after the tensorlake@0.5.144 attack (npm ci, lockfile, --ignore-scripts for untrusted installs, no unasked dependencies, never revoke tokens yourself after a suspected malicious package). Rules reduce risk; they do not guarantee you are patched.