📈 GH Trending Explorer 🎲 Surprise me

← all repos

lisbeth718 / prod-check

git clone https://github.com/lisbeth718/prod-check.git

Claude Code skill: audit your app before you take money from customers. 180 checks, 3 tiers, evidence-backed.

+0
stars, 24h · #36,109
+27
stars, 7 days · #4,522
+27
stars, 30 days · #21,695
28
★ total · +27% this week
1
forks · 0 open issues

Star history (as recorded here)

2026-10-20 · 27 ★ 2026-10-22 · 28 ★
LicenseMIT
Created2026-10-16 (6d ago)
Last push2d ago
Watchers0
OwnerUser

Links

README excerpt
# prod-check
Find the bugs that end startups before your first customer finds them: leaked keys, other users' data, fake payments, no restorable backup.
Run `/prod-check` in your repo. In about 30 minutes your agent checks up to 68 high-risk items with evidence (file:line, command output); anything it cannot reach is reported UNKNOWN. It then hands you the top 5 fixes.
Read-only in production; replays write only to a test environment you provide; test alerts and messages are sent only with your OK. Built for agent-written apps: catches pasted keys, packages that don't exist, IDOR in server actions, Supabase tables open to the browser, and webhooks nobody verifies.
Going enterprise later? `full` covers all checks up to SSO, SOC 2 and DPAs. You don't need those yet.
Based on a checklist by Mari Luukkainen.
## Install
Claude Code:
 [code] 
Then run `/prod-check` in your app's repo.
Other agents: give the agent the path to `SKILL.md` and tell it to follow it. It reads the files under `references/` as needed.
Humans: work through `CHECKLIST.md`. Each line has an ID, a tier and a gate. `references/checks/` says how to verify each one.
## Usage
 [code] 
Quick mode runs and reports only a short list of the highest-risk checks. See `references/quick-profile.md`.
Fixes are proposed after the report and applied only when you say go.
## How it works
### Tiers and gates
A weekend landing page and a multi-tenant billing app do not need the same checks.
Each check has a tier, and some sections are gated on a feature.
- T0 Public site: no accounts, no payments, no customer data.
- T1 Paying: any of: accounts, customer data, payments.
- T2 Enterprise/regulated: SSO, questionnaires, contract terms, or regulated data (health, financial account, legal, member, union or political data). Ordinary PII such as CVs and emails stays T1.
- A (advisory): reported as notes, never a FAIL.
The auditor checks everything at or below your tier. Anything above is N/A, not FAIL.
Gates are features you may not have: accounts, multi-tenant, payments, usage billing, uploads, imports, transactional email,
background jobs, API or webhooks, AI, URL fetching, impersonation, subscriptions, free trials, multiple locales, health data, card data, database, self-hosted servers. A gated-off section is N/A.
`multi-tenant` means an org, team or workspace entity where several users share records; per-user ownership alone is not multi-tenant.
Some things only a person can answer (on-call, postmortems, pen tests, who owns the domain).
The auditor asks you once. An artifact it can verify is PASS, your word alone is ATTESTED, and an admitted gap is FAIL.
### Stages
T0 and T1 rows carry a stage; T2 and advisory rows use `-`. Stages: `launch` (before the first real user outside the team or the first invoice, whichever comes first), `10` (before 10 customers) or `100` (before 100 customers). All T0 rows are `launch`.
Quick mode is the high-risk T0 rows already in the quick list plus every T1 launch row;