How to write an emdly skill that passes review and actually works in an agent — structure, rules that bind, output contracts, and the mistakes reviewers bounce.

Writes first-touch e-mails that sound like a person — one observation, one sentence of value, one soft ask. Kills every buzzword on sight.

Drafts a reply to a support ticket from the conversation and the help center — answers the actual question first, promises only what the policy allows, and escalates when it should.

Turns a draft PRD into one a team can build from — problem before solution, success metric with a number, explicit non-goals, and open questions with owners.

Scores every Jira ticket for clarity, scope and effort before it hits the sprint. Flags vague acceptance criteria and suggests a rewrite.

Triages a vulnerability report (Dependabot, npm audit, composer audit) — reachable or not, exploitable in this app or not, and what to do this week versus later.

Reads yesterday's commits, tickets and threads, and writes each person's standup draft — blockers first, no ceremony.

A reviewing discipline for agents — read the diff twice, test the edge cases, comment on intent, never nitpick formatting a linter owns.

Turns a week of merged PRs into a changelog humans read — grouped by impact, written for users, breaking changes on top.

Reads a GA4 export and finds where the checkout funnel leaks — drop-off by step, device and source — with the one fix to try first.

Prepares a retrospective from the sprint's data — what shipped, what slipped, where time went — and turns it into three concrete questions for the team.

Edits a landing page section by section — one promise above the fold, proof before features, and a CTA that says what happens next.

Sketches a threat model for a feature from its description and data flow — assets, entry points, top threats by STRIDE, and the mitigation that exists or is missing.

Builds content briefs from a keyword — search intent, outline, entities to cover, and the questions the top ten never answer.

Holds any draft against your voice guide — banned words, sentence rhythm, claims policy — and returns a marked-up pass, not a rewrite.

Given kubectl output and alerts, ranks probable causes and proposes the next safe diagnostic step — read-only commands only, never a mutation without a human.

Explains what a query actually does — joins, filters, gotchas — in plain language, then flags the index it wishes existed.

Audits Google Search Console data — pages losing clicks, queries with impressions but no CTR, and product pages cannibalizing each other.

Writes an endpoint's reference page from its route, request validation and response shape — every field, every error, one working example.

Grades take-home submissions against a rubric you define, blind to names, and writes feedback the candidate can actually use.

Turns a vague metric request ("active users", "churn") into a precise definition with grain, window, filters and the SQL skeleton — before anyone builds the chart.

Sorts raw playtest notes into balance, UX, bugs and feel — ranked by how many testers hit it, with repro steps pulled from context.

Drafts blameless postmortems from a timeline and a channel export — impact, contributing factors, and actions with owners.

Reads stat tables and patch notes, then flags outliers — dominant strategies, dead perks, and curves that punish new players.