A unified "Suggest improvement" affordance on posts, comments and changelog drafts. User-triggered only - never autonomous. Three modes: (1) Tighten wording, (2) Add structure (bullets, headings), (3) Summarise from linked GitHub PR/issue context. Output appears as a side-by-side diff the user must explicitly accept; AI never writes directly to a row.

Guardrails: every prompt is constrained to the source content + an explicit task instruction (no open-ended queries), no chain-of-tool-use, no external retrieval. Accepts only suggestions of size <= 1.5x the original. Surfaces the source model + token count on each suggestion so users can audit. Skipped entirely when AI provider settings are unset.