Launch playbook

launch-playbook — Pre-Launch Security Playbook

Run a time-boxed, evidence-based security playbook before public launch.

Agent-agnostic procedure. Same steps on Claude Code, Gemini CLI, OpenAI Codex/ChatGPT agents, Cursor, Windsurf, Grok, or a human with a terminal. Loading differs by host (see README); execution does not.

Goal	Baseline that stops the failures fast-shipped apps actually hit
Not a goal	Enterprise audit, pentest cert, HIPAA/PCI/SOC2
Budget	~30 minutes of agent work; depth over theater
Rule	No finding without path/evidence. No “might be vulnerable” without a reason.

Load when needed (paths relative to this skill folder):

• references/checklist.md — human checklist
• references/scan-playbook.md — commands + secret patterns
• references/ai-prompts.md — optional deep-dive prompts

If invoked without skill auto-load: user may attach or @ this file. Follow it fully anyway.

---

Modes

User intent	Mode
Default / “audit” / “checklist”	Report only
“Fix it” / “make it shippable”	Remediate then re-score
“Quick” / “10 min”	Fast path: secrets → access control → paid APIs → auth → stop

Do not rotate production secrets, wipe data, or force-enable RLS that locks out prod without explicit confirmation.

---

Phase 0 — Detect stack (2 min, mandatory)

Inspect lockfiles, config, and imports. Record in the report:

Framework: Next.js | Remix | SvelteKit | Vite SPA | other
Auth: Supabase Auth | Clerk | Auth.js/NextAuth | Lucia | Firebase | custom | none
DB: Supabase | Firebase | Prisma+Postgres | Mongo | Drizzle | other | none
Hosting signals: Vercel | Cloudflare | Fly | Railway | unknown
Paid APIs: OpenAI | Anthropic | xAI | Stripe | Resend | Twilio | other
Client data access: direct Supabase/Firebase from browser? yes/no
AI features: chat | agents/tools | image | embeddings | none
Uploads: yes/no   Payments: yes/no   Multi-tenant: yes/no

Branching: Skip N/A sections. Map Supabase RLS checks to Firebase Rules / server-only Prisma authz. Never pretend a Next-only check applies to a static SPA.

---

Phase 1 — Mandatory scans (run these, do not “reason about” secrets)

Use ripgrep/grep tools available. Prefer references/scan-playbook.md. Minimum:

1.1 Secret & key exposure

Search source (exclude node_modules, .git, lockfiles, dist / .next / build if noisy):

Pattern family	Examples
Cloud LLM	sk- (OpenAI), sk-ant-, xai-, gsk_ (Groq), AIza (Google)
Stripe	sk_live, sk_test, rk_live, whsec_
Supabase	service_role, long JWT-looking eyJ... in client files
AWS	AKIA + secret key patterns, AWS_SECRET
GitHub	ghp_, github_pat_, gho_, ghu_
Private keys	BEGIN PRIVATE KEY, BEGIN RSA PRIVATE KEY, BEGIN OPENSSH
Slack/Discord	xoxb-, xoxp-, xoxa-, Discord bot tokens
Generic	api[_-]?key, secret[_-]?key, password\s*= in non-test code

Also check:

• .env, .env.local, .env.production tracked by git?
• Client env misuse: NEXT_PUBLIC_* / VITE_* / EXPO_PUBLIC_* / PUBLIC_* holding secrets
• README/docs/screenshots with real keys
• service_role or DATABASE_URL imported from modules that also run in the browser
• CI/CD config and committed Terraform/Pulumi secrets

Classify each hit: public-by-design | secret-exposed | false positive.

secret-exposed in client or git → DO NOT SHIP until rotated + removed from tree (history purge is a follow-up; rotation is immediate).

1.2 Dangerous code patterns

Search and open matches:

Risk	Patterns / files
XSS	dangerouslySetInnerHTML, {@html, v-html, unescaped markdown→HTML
SQL concat	string-built queries, raw $queryRaw / $executeRaw with user input
Command inject	exec(, execSync(, spawn( with user-controlled strings
SSRF	fetch(user, axios.get(req., webhook/URL fields without allowlist
Open redirect	redirect(, window.location = user param without allowlist
Admin naked	/admin, /debug, /api/test, swagger/openapi in prod without auth
Mass assign	spreading req.body straight into Prisma data: / Supabase update
Debug prod	missing production guards, console.log of tokens, source maps forced on

1.3 Dependency & config

• Lockfile present? (package-lock.json / pnpm-lock.yaml / yarn.lock / bun.lock)
• Run npm audit --omit=dev (or pnpm/yarn equivalent) when Node project; note critical/high only
• .gitignore includes .env* (and secrets are not force-added)
• Public storage: Supabase storage policies, S3 public ACLs, unauthenticated upload routes

---

Phase 2 — Access control & data plane

2.1 Database / backend access (highest failure rate)

Supabase

• RLS enabled on every table with user/tenant data
• Policies scope by auth.uid() or org membership — not using (true) for private data
• WITH CHECK on inserts/updates (not only USING)
• No browser use of service role
• Migrations / SQL in repo match “we have RLS” claims; empty policy list on private tables is FAIL

Firebase

• Rules default-deny; read/write gated by auth.uid + resource ownership
• No open allow read, write: if true (or if request.auth != null alone on private docs)

Prisma / Drizzle / server DB

• No DATABASE_URL in client bundle
• Every query that touches user rows filters by session user/org
• No “getById” without ownership check (classic IDOR)

IDOR / multi-tenant

• Routes like /api/orders/[id], /api/docs/[id] verify owner or membership, not just “is logged in”
• Numeric/sequential IDs: confirm server rejects cross-user access in code
• Team/org features: membership checked on mutate, not only on UI hide
• Realtime / subscriptions / storage paths also scoped (not only REST handlers)

False confidence: RLS enabled with wrong policies is worse than obvious “no RLS” (silent “secure”). Flag always-true policies for authenticated role on sensitive tables.

2.2 Validation & errors

• Writes validated on the server (Zod/Valibot in Server Actions / API / Edge) — client-only validation = FAIL/WARN
• Length limits on text fields; file size/MIME if uploads
• Production errors generic; no SQL/stack/PrismaClientKnownRequestError to clients
• Full user rows (password hash, tokens, internal flags) not returned to client

2.3 Uploads & storage (if present)

• Auth required unless intentionally public
• Size + type allowlist; no SVG-as-image without sanitization if served inline
• Stored outside web root or via signed URLs
• Path not user-controlled (../../)
• Bucket not world-readable for private user content
• Filename not used as executable path; content-type not fully trusted from client

2.4 SSRF & outbound fetch (if present)

• User-supplied URLs allowlisted (scheme + host) or blocked for link-local / metadata IPs (169.254.169.254, metadata.google.internal, 127.0.0.0/8, private RFC1918 ranges)
• No blind fetch(req.body.url) for “preview” / “import” features without controls
• Redirect-following considered (open redirect to internal host)

---

Phase 3 — Auth & sessions

3.1 Failure-case design (code + copy review)

Test	Secure	Insecure
Bad password ×5	Generic error; rate limit / lockout	“Wrong password for this email”; no throttle
Reset unknown email	Same response as known	“Email not found”
Verify/magic link twice	Expired/used message	500 / infinite re-auth bugs
Signup existing email	Non-enumerating message (or intentional product choice documented)	Clear “already registered” with no rate limit

Also:

• Session cookies: HttpOnly, Secure, SameSite (or provider defaults verified)
• Logout invalidates server session when applicable
• Password reset tokens single-use / short TTL if custom
• OAuth: redirect URI allowlist; no open redirectTo / callbackUrl / next from query
• Email verification required if app holds sensitive data (else WARN)
• CSRF: cookie-session mutations protected (SameSite + origin checks / framework CSRF)

3.2 Authorization ≠ authentication

Logged-in is not enough. Admin/role checks must be server-side on privileged routes. Client-only role flags in localStorage/JWT claims without server verify = FAIL.

---

Phase 4 — AI features (skip if none)

Fast-shipped AI apps die on cost and prompt injection, not only key leaks.

• Model calls only on server; keys never in client
• Auth required for expensive routes (or hard public quota + CAPTCHA)
• Rate limit + per-user daily/monthly cap on generation endpoints
• Provider dashboard hard cap + billing alert recommended in report
• User content treated as untrusted: not concatenated into system prompts as instructions
• Tool/agent calling: tools cannot exfiltrate secrets, run shell, or hit internal URLs without allowlist
• Output rendering safe (markdown XSS / HTML)
• No logging full prompts that contain user PII to third-party log drains without need
• Streaming endpoints cannot be invoked anonymously in a tight loop without limits
• Max tokens / max steps bounded on agent loops

DO NOT SHIP if an unauthenticated endpoint can burn paid tokens at scale.

---

Phase 5 — Abuse, cost, and edge

• Rate limits: auth routes + any paid/third-party route
• CAPTCHA/bot protection on public signup/contact/waitlist (Turnstile, hCaptcha, etc.)
• CORS: explicit origins in prod; not * with credentials on private APIs
• Webhooks (Stripe etc.): signature verify; reject unsigned; idempotency considered
• Body size limits on APIs
• Security headers where framework allows (baseline CSP, X-Content-Type-Options: nosniff, frame controls if sensitive UI)
• Cron/admin secrets not guessable query tokens
• Feature flags / debug routes disabled or auth-gated in production

---

Phase 6 — Legal / product floor (light, code-aware)

• Privacy policy route or external link exists if you collect email/PII
• Password hashing if custom auth (bcrypt / argon2 / scrypt — not plain, not reversible “encrypt password”)
• No debug scripts that export users or dump production data into personal accounts
• Data map one-liner in report: where PII lives + processors (auth, email, analytics, AI)

Keep legal notes short and product-relevant. If AI-generated code or third-party licenses matter for this ship, one line: human review + license hygiene. Do not expand into caselaw.

---

Phase 7 — Verdict

Severity → verdict

Severity	Examples	Effect
S0 Blocker	Secret in client/git; no RLS/rules on private data; unauth paid AI burn; public service_role	DO NOT SHIP
S1 High	Clear IDOR; SSRF; stored XSS; admin open; webhook unsigned with money impact	DO NOT SHIP if exploitable without auth or affects all tenants/users
S2 Medium	Auth enumeration; missing rate limit on auth; client-only validation; no CAPTCHA on spammy forms	SHIP WITH FIXES
S3 Low	Missing CSP; verbose logs server-only; dependency moderate noise	WARN; ship ok

SHIP only if zero S0/S1 open.

If S1 is authenticated-only and single-tenant scoped, still default to DO NOT SHIP unless the user explicitly accepts residual risk in writing in the report.

Report format (always)

# Launch Playbook Report

**Repo:** …
**Stack:** … (from Phase 0)
**Mode:** report | remediate | fast
**Verdict:** SHIP | SHIP WITH FIXES | DO NOT SHIP

## Scorecard
| Phase | Area | Status | Highest severity |
|-------|------|--------|------------------|
| 1 | Secrets & dangerous patterns | PASS/FAIL/WARN/N/A | S0–S3 |
| 2 | Access control / IDOR / uploads | | |
| 3 | Auth & sessions | | |
| 4 | AI features | | |
| 5 | Abuse / cost / edge | | |
| 6 | Legal / product floor | | |

## Blockers (S0–S1)
### B1 — title
- **Evidence:** `path:line` — snippet or fact
- **Impact:** …
- **Fix:** concrete change (code/SQL/config)

## Fixes before / within 48h (S2)
…

## Passed (evidence-backed)
- …

## Scans run
- commands + outcome summary

## Out of scope / N/A
- …

Keep the report short. Prefer 5 real blockers over 40 theoretical essays.

---

Remediate mode

Order of operations:

1. Remove/move secrets; tell user to rotate (do not print full secrets in chat)
2. Access control (RLS/rules/ownership checks)
3. Close unauth paid endpoints (auth + rate limit + cap)
4. Fix XSS/SSRF/open redirect with minimal patches
5. Generic errors + server validation on hottest write paths
6. Re-run Phase 1 scans + re-score

---

Anti-slop rules (mandatory)

1. Evidence or delete the finding. No “consider adding…” without a gap you observed.
2. Do not dump references/ai-prompts.md as the whole audit. Use those prompts only as optional depth after mandatory scans.
3. Do not mark PASS because the README says “secure” or “read-only.”
4. Map to stack. Wrong-stack advice is a failed audit.
5. Prioritize money and data loss over header perfectionism.
6. Never claim “pentest complete” or “compliant.”
7. If the app is static marketing with no backend: short report, mostly N/A, still scan for leaked keys in repo.