[rmkr-dev]

Adds a Skillbook entry and a matching Promptbook prompt for hardening LLM features (chat, RAG, tool-calling agents, LLM gateways) against prompt injection and unsafe output handling.

Skill: llm-app-prompt-injection-guardrails

• Trust map of every prompt input (trusted / user / untrusted), with untrusted content kept in its own delimited message
• System prompt treated as non-secret: no credentials, internal endpoints, or authorization logic in it
• Retrieval filtered by the caller's access before prompting; tools run in the caller's identity, never a broad service principal
• Narrow typed tool schemas with server-side argument validation; no generic fetch / SQL / exec tools
• Code-enforced approval for high-impact actions, plus round, token/cost, and time caps on agent loops
• Output encoded for each sink, strict structured-output parsing, and Markdown image exfiltration blocked
• Input/output guards with logged verdicts, redacted logging, and a red-team eval set in CI
• Copilot / Claude Code / Codex placement notes and a SKILL.md stub

Prompt: harden-llm-app-against-prompt-injection

• Inventory, trust map, ranked findings, code-first fixes, red-team eval cases, and tool / sink tables with before/after attack success rates

Catalogs go to 38 skills / 38 prompts. Catalog validation from scripts/site-check.mjs passes locally on both files.