LLM Workflow

A Second Agent as Reviewer — 11 of 34 Facts in a 25-Page AI Plan Were Fabricated | Agentic AI in Practice (XI)

Vendor PPTs, AI-drafted admissions emails, internal plans written with Claude — fact-fabrication rate sits at a 20-30% industry baseline. Five minutes in you can spot why 'let the LLM double-check itself' is pseudo-verification; ten minutes in you have the DRAFT → VERIFY → FINALIZE 3-phase gate template; twenty minutes in you have R1-R7 — the seven categories of fact errors that keep recurring (enum casing / fabricated emails / API paths / model IDs / deadlines) — turned into a PR checklist. Next time you review AI-drafted material, every claim traces back to a `file:line` or URL.