trust-compliance
Legal, security, privacy, and deliverability requirements every delivered app must satisfy — legal pages, PII/deletion rights, CSP/Trusted-Types/Zod, AI-agent security, RFC7807 errors, email domain auth.
Install
npx skills add https://github.com/heymegabyte/agent-skills/tree/main/27-trust-compliance
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install heymegabyte-agent-skills@llmmart
git clone https://github.com/heymegabyte/agent-skills.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole heymegabyte/agent-skills collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
Trust, Security & Compliance — every app must satisfy
Legal & policy
- Ship real
/privacy,/terms,/accessibilitypages accurate to the ACTUAL data practices — never boilerplate that lies. Accessibility statement is ADA Title II / WCAG 2.2 AA-ready. - No dark patterns — honest defaults, no forced continuity, no confirm-shaming (Christ-like ethos).
Data & PII
- Collect the minimum PII needed; never log secrets or PII in plaintext; keep credentials server-side only.
- Honor user right-to-deletion and data export on request. State retention plainly.
Security controls (every app)
- CSP Level 3
strict-dynamic+ per-request nonce · Trusted Types · HSTS. No inline script without a nonce. - Zod validation at EVERY runtime boundary (env · API in/out · params · forms · webhooks · queue/DO messages · AI outputs) — validation is a security control, not just typing.
- Forms gated by Turnstile. All hyperlinks valid (no dead/hostile links).
- Tenant isolation:
org_idon every row + query; 404 (never 403) on cross-tenant access.
AI-agent security
- Treat all model/tool input as untrusted: defend against prompt injection; authorize every tool call; validate + schema-bind every AI output before use (Contract-First AI). Sandbox generated/risky code before it runs.
Errors (never leak internals)
- User-facing errors use RFC7807 envelopes:
code+correlationId+errors[]+ what-to-do-next. Never expose stack traces, SQL, or internal identifiers.
Email domain authentication
- Sending domain publishes SPF + DKIM + DMARC (BIMI where a VMC exists). These records live on the SENDING domain's zone (may differ from the site domain — surface any mismatch before staging).
Files (agent-skills)
-
SKILL.md 2.1 KB
--- name: "trust-compliance" description: "Legal, security, privacy, and deliverability requirements every delivered app must satisfy — legal pages, PII/deletion rights, CSP/Trusted-Types/Zod, AI-agent security, RFC7807 errors, email domain auth." triggers: - "security" - "privacy" - "legal" - "compliance" - "csp" priority: 2 pack: "security" stage: stable --- # Trust, Security & Compliance — every app must satisfy ## Legal & policy - Ship real `/privacy`, `/terms`, `/accessibility` pages accurate to the ACTUAL data practices — never boilerplate that lies. Accessibility statement is ADA Title II / WCAG 2.2 AA-ready. - No dark patterns — honest defaults, no forced continuity, no confirm-shaming (Christ-like ethos). ## Data & PII - Collect the minimum PII needed; never log secrets or PII in plaintext; keep credentials server-side only. - Honor user right-to-deletion and data export on request. State retention plainly. ## Security controls (every app) - CSP Level 3 `strict-dynamic` + per-request nonce · Trusted Types · HSTS. No inline script without a nonce. - Zod validation at EVERY runtime boundary (env · API in/out · params · forms · webhooks · queue/DO messages · AI outputs) — validation is a security control, not just typing. - Forms gated by Turnstile. All hyperlinks valid (no dead/hostile links). - Tenant isolation: `org_id` on every row + query; 404 (never 403) on cross-tenant access. ## AI-agent security - Treat all model/tool input as untrusted: defend against prompt injection; authorize every tool call; validate + schema-bind every AI output before use (Contract-First AI). Sandbox generated/risky code before it runs. ## Errors (never leak internals) - User-facing errors use RFC7807 envelopes: `code` + `correlationId` + `errors[]` + what-to-do-next. Never expose stack traces, SQL, or internal identifiers. ## Email domain authentication - Sending domain publishes SPF + DKIM + DMARC (BIMI where a VMC exists). These records live on the SENDING domain's zone (may differ from the site domain — surface any mismatch before staging).
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.