Everything we publish must be traceable to what the product actually does and costs, and must read like a practitioner wrote it. The reader's interest comes first: teach something they can act on, and answer straight — including when the honest answer is "no" or "it costs money."
Check these against code before repeating any of them — they change: pricing and credits (
, the pricing page), free-plan limits (
src/shared/audit-limits.ts
), MCP capabilities (
— one file per tool), and any UI affordance copy tells the reader to use (the column, sort, or filter must exist in the client code).
Spawn subagents to run the review passes (voice/deslop, claims accuracy, directness) and have them return exact old → new proposals rather than editing directly. Do not accept their proposals blindly: verify each one against the actual file, and each factual claim against the code, before applying — subagent rewrites can introduce their own awkwardness or errors, and a proposal that mismatches the file means it reviewed stale text. After applying, sweep the changed surface yourself (patterns cluster — one em dash or hedge usually has neighbors), then run
npm --prefix web run types:check
and prettier on touched TS/TSX.