What problem does it solve? LLM features fail in production for predictable reasons: hand-assembled JSON breaks schema validators, raw JSON dumps render as unreadable <pre> blobs in dashboards, single-shot prompts cannot use tools, and inline JavaScript in Python-served dashboards fails silently. This Skill provides tested patterns for each of these failure modes. ## Core Features & Use Cases - Machine-validated JSON output: Write the payload to a file, serialize with json.dumps, and validate with json.loads before emitting, eliminating 'Expecting , delimiter' errors. - Structured dashboard cards: Replace raw JSON <pre> blocks with an IIFE that renders verdict badges, evidence lists, recommended actions, and MITRE tags, with correct template-literal escaping. - Static-to-agentic upgrade: A three-phase pattern (ledger, deterministic enrichment, function-calling loop) for converting a single LLM call into a tool-using agent with full audit trail. - Web UI debugging: Diagnose silent JS failures in FastAPI/Flask dashboards using new Function compile checks, browser console inspection, and binary search. - Use Case: You are adding an LLM investigation feature to a security alert dashboard. Use this Skill to emit schema-valid verdict JSON, render it as a readable card, upgrade the prompt to an agentic tool-calling loop, and debug the modal that stopped opening. ## Quick Start Use the llm-app-patterns skill to convert my single-shot LLM analysis endpoint into a tool-calling agentic loop with an investigation ledger.
