← ClaudeAtlas

jev-writing-checklisted

Use before sending client-facing or team-facing writing — status updates, emails to a reviewer or client, PR descriptions, release notes, ticket comments, docs — especially when it summarizes work, makes promises, or reports results
calelamb/jevskillz · ★ 0 · AI & Automation · score 70
Install: claude install-skill calelamb/jevskillz
# Check writing for honesty and fit Writing fails in two ways: it overstates the evidence, or it doesn't answer what the reader asked. Check both before sending. **Honesty (claims mode):** one item per factual/promissory sentence, `evidence` = the proof. Any FAIL is an overclaim — soften it to what's shown or add the proof. Measured: "fixes all accessibility issues" vs a diff touching 3 of 61 controls → 0.03. **Fit (score mode):** state = `{request, draft}`; dimensions such as: - `answers`: "How completely does the draft answer every question in `request`?" - `clarity`: "How easily can the intended reader act on this without follow-up questions?" - `tone`: levels from "defensive or vague" → "direct, specific, courteous". Revise the lowest dimension first; rerun. For a reviewer reply (e.g. an app-store resubmission), make one `claims` item per finding you say is resolved, with the evidence you'll attach. **REQUIRED BACKGROUND:** jev-core.