Build (or refresh) a voice profile of how the user writes email, from their own sent Gmail. The profile is what email-draft-in-voice consumes to ghostwrite email that sounds like them.
80%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Passed
No findings from the security scan
Fix and improve this skill with Tessl
tessl review fix ./skills-seed/email-voice-profile/SKILL.mdUse this when the user asks you to learn how they write email — "learn my voice",
"build my email voice profile", "study my sent mail so you can draft for me" — or when
email-draft-in-voice needs a profile that doesn't exist yet.
Personal DMs only. Sent mail is private; refuse to run this from a channel or group and offer to continue in the user's DM. The profile you produce is published to their personal Files, so only their own conversations and their own Inbox drafting can read it.
Computers are replaced without notice, so nothing durable lives on one. Before building,
search memory for "email voice profile" and GET $AGENT_API_URL/v1/files?q=email-voice-profile
(header x-agent-capability: $AGENT_API_TOKEN). An existing profile is refreshed, not
rebuilt blind: download it with GET $AGENT_API_URL/v1/files/<id>/content and carry its
validated hard rules forward unless the new corpus contradicts them.
Use the selected email access skill to retrieve up to 300 of the user's sent messages.
Strip quoted replies and ---delimited signatures, drop automated and trivial messages,
and write one JSON line per email to voice/corpus/corpus.jsonl, tagged internal or
external by recipient domain. Record corpus stats; if fewer than ~50 usable emails
survive, tell the user the profile will be rough and ask whether to continue.
Read a deliberate spread, not just the top of the file: oldest and newest, internal and external, one-liners and long emails. You are looking for what is distinctive — patterns a generic professional emailer would not share.
Watch for a signature block: only ---delimited signatures are stripped, so a plain
signature (name / title / phone) survives at the end of many bodies. Treat a verbatim
recurring trailing block as the mail client's signature, not a chosen sign-off — note
it in the profile so drafts never include it (Gmail appends it on send).
Write email-voice-profile.md (that exact name; the Inbox drafting loop looks it up by
name) with exactly these sections,
every claim backed by observed frequency ("uses 'Best,' in 80% of external mail"),
with 2–3 verbatim example lines each:
Hold out 3 real threads the profile wasn't built from. For each, draft the user's
reply using only the profile, then compare with what they actually sent. Where the
real email diverges, tighten the profile — usually a missing hard rule or register
distinction. Record validated: <date> at the top of the profile once the drafts read
as plausibly theirs.
Show the user the finished profile and ask what feels wrong; their corrections go in as hard rules.
Upload the finished file to Files: curl -fsS "$AGENT_API_URL/v1/files/upload-client" -H "x-agent-capability: $AGENT_API_TOKEN" -o /tmp/qm-upload.py && python3 /tmp/qm-upload.py email-voice-profile.md, then record one memory fact: "Email voice profile is Files
(built )". The Inbox drafting loop reads the newest email-voice-profile.md
in the user's Files on every pass, so publishing is what makes drafts sound like them.
The corpus stays on the computer; only the profile is durable. Re-run this skill any
time they want it refreshed — publish the new version and update the memory fact.
yc-software/qm
c94f65b
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.