email-compose-review — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited email-compose-review (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
You are drafting an email on behalf of a human. Before it reaches Mail, your draft must survive a multi-pass review. The goal: remove AI writing tells, tighten the prose, verify facts, and ensure the message lands clearly — all without adding filler or bloat.
This skill runs on LLM-drafted text only. If the user wrote the body themselves and asks for a review, apply it. If they hand you a body to send as-is, skip the review and open the draft directly.
review, revise — then open the draft.
Automatic sending of LLM-drafted email is not acceptable.
any MCP tool.
Write the email body internally. Don't call any MCP tools yet. Just write it.
In Claude Code (sub-agents available): spawn all five reviewers as parallel sub-agents in a single message. Each receives the draft and returns specific flagged issues plus suggested fixes. Collect all five outputs, then proceed to Step 3.
In Claude Desktop (no sub-agents): run the five reviews as sequential internal passes. Be honest and rigorous — don't just nod through each one.
AI-written email has recognizable tells. Find and remove them all.
Opening clichés to cut:
Sycophantic filler:
Hedge padding:
Corporate buzzwords:
going forward, at the end of the day, value-add, bandwidth, deliverable
Sign-off bloat:
unless the context is formal enough to warrant it
The test: read each sentence and ask "would a normal person write this?" If not, cut it.
Spelling, grammar, punctuation. Specifically:
Fix these silently — don't flag minor grammar corrections in the changelog unless they change meaning.
Convert weak passive constructions to active where it makes the writing stronger.
| Passive | Active |
|---|---|
| "The issue will be looked into by our team" | "Our team will look into it" |
| "It was decided that the meeting would be moved" | "We moved the meeting" |
| "Your request has been received" | "We received your request" |
Exception: passive is correct when the subject is unknown, irrelevant, or when the object is more important than the actor. Don't force active voice where it sounds unnatural.
Check every claim against the context you were given:
"[email protected]", use "Hi" or wait for confirmation.
a casual thread shouldn't open with "Dear Mr. Smith,")
Flag anything you're uncertain about rather than guessing. The user can fill in gaps — hallucinated details in email are worse than placeholders.
at the end?
as short as the content allows.
Combine all five reviews into:
conflict, prefer the version that is clearer and shorter.
the user should verify before sending).
Present the output in this format:
**Draft ready for review**
[revised email body]
---
**Changes made:**
- [bullet]
- [bullet]
**Please verify before sending:**
- [correctness flag if any — omit this section entirely if nothing to flag]Then ask: "Want me to open this as a draft in Mail?"
Wait for the user to confirm. If they ask for edits, apply them and show the revised version again. Do not open Mail until they say yes.
Once the user approves, call compose_email or reply_email with:
send: false — alwaysThe user reads it once more in Mail and clicks Send themselves. That's the human-in-the-loop gate that makes this safe to use for real correspondence.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.