Draft
Turn what the conversation already holds into a message the user sends as
themselves.
Input
The source is whatever is in context: an agent's research, a support
investigation, notes, bullet points, a pasted thread. When another skill already
drafted a reply, treat that reply and the evidence behind it as the source.
Work out from the conversation:
- Recipient: who, and how they relate to the user (colleague, friend, family,
customer).
- Channel: chat (Teams, Slack, SMS), email, or a doc.
- Language: the language of the message being answered, otherwise the
language of the conversation.
Ask one question covering only what is still missing.
Steps
- Read
voice.md next to this file.
- Pick the voice.
- Sent as the user personally: follow
voice.md.
- Sent in a company's name to its customers or partners: follow that company's
tone guide if one is loaded in this session, and keep only the bans from
voice.md. If none is loaded, say so and use voice.md.
- List the claims the message will make, each with where it comes from in
the conversation. Keep the list to yourself.
- Write the message. For chat, use no markdown emphasis (
**bold**,
_italic_). Teams and Slack show the markers literally. Carry emphasis through
short lines instead.
- Check it against the accuracy rules below. Then search it for em-dashes and
en-dashes.
- Output the message on its own, ready to paste. Under it, add one line per
claim you dropped or softened, so the user can check. Skip that part when
nothing was dropped.
- Copy to clipboard:
pbcopy on macOS, wl-copy or xclip -selection clipboard on Linux, clip on Windows. Confirm only after the command
succeeds. If it fails, say the message needs manual copying.
Accuracy
Research is long and hedged. A short message in a confident voice strips the
hedges. That is the main way this skill goes wrong.
- Every claim traces to the conversation. Add no facts, numbers, dates,
promises or next steps the source does not hold.
- Uncertainty survives. If the source says "likely" or "not verified", the
message says so, once and plainly: "Probably the customer discount. Not
verified yet."
- A guess never becomes a cause. Open questions stay open.
- Cut detail, never the strength of the evidence.
- Only what the recipient should see. Conversations hold private context,
internal IDs and other people's names. Leave out what this recipient has no
reason to read.
1---2name: draft3description: Write a message, reply or short write-up that the user sends under their own name, in their voice, from whatever the conversation holds: research, a support investigation, notes or bullet points. Use when the user says "write a message to <name> about this", "draft a reply", "reply to <name>", "turn this into a Teams message", "email <name> about this", "write this up in my voice", or "/draft". Not for text the user wrote or dictated (polish-text) or for proofreading a post (proofread-post).4---56# Draft78Turn what the conversation already holds into a message the user sends as9themselves.1011## Input1213The source is whatever is in context: an agent's research, a support14investigation, notes, bullet points, a pasted thread. When another skill already15drafted a reply, treat that reply and the evidence behind it as the source.1617Work out from the conversation:1819- **Recipient:** who, and how they relate to the user (colleague, friend, family,20 customer).21- **Channel:** chat (Teams, Slack, SMS), email, or a doc.22- **Language:** the language of the message being answered, otherwise the23 language of the conversation.2425Ask one question covering only what is still missing.2627## Steps28291. **Read `voice.md`** next to this file.302. **Pick the voice.**31 - Sent as the user personally: follow `voice.md`.32 - Sent in a company's name to its customers or partners: follow that company's33 tone guide if one is loaded in this session, and keep only the bans from34 `voice.md`. If none is loaded, say so and use `voice.md`.353. **List the claims** the message will make, each with where it comes from in36 the conversation. Keep the list to yourself.374. **Write the message.** For chat, use no markdown emphasis (`**bold**`,38 `_italic_`). Teams and Slack show the markers literally. Carry emphasis through39 short lines instead.405. **Check it** against the accuracy rules below. Then search it for em-dashes and41 en-dashes.426. **Output** the message on its own, ready to paste. Under it, add one line per43 claim you dropped or softened, so the user can check. Skip that part when44 nothing was dropped.457. **Copy to clipboard**: `pbcopy` on macOS, `wl-copy` or `xclip -selection46 clipboard` on Linux, `clip` on Windows. Confirm only after the command47 succeeds. If it fails, say the message needs manual copying.4849## Accuracy5051Research is long and hedged. A short message in a confident voice strips the52hedges. That is the main way this skill goes wrong.5354- **Every claim traces to the conversation.** Add no facts, numbers, dates,55 promises or next steps the source does not hold.56- **Uncertainty survives.** If the source says "likely" or "not verified", the57 message says so, once and plainly: "Probably the customer discount. Not58 verified yet."59- **A guess never becomes a cause.** Open questions stay open.60- **Cut detail, never the strength of the evidence.**61- **Only what the recipient should see.** Conversations hold private context,62 internal IDs and other people's names. Leave out what this recipient has no63 reason to read.
alexoberneyer