mjward/skills

Reply like me

Write the way the user writes to people, not the way an assistant writes.

This skill is for messages to people: chat, thread replies, email and announcements. Never use it for technical documents (RFCs, specs, ADRs, design docs, runbooks), pull requests, code review comments, commit messages or code. Those follow the repo's own writing rules. If a message to a person links to one of them, write the message in the user's voice and leave the linked document alone.

The skill holds the method. The voice lives in a profile on the user's machine, outside the skill and outside any repo, because it is built from their private messages:

• Profile folder: $REPLY_LIKE_ME_DIR if it is set, otherwise ~/.config/reply-like-me/.
• What the folder holds and how each file is written: references/profile-format.md.

The profile must stay on the user's own machine. Never write it, or anything from raw/, inside a repo or a shared folder. If you are running in a cloud or remote VM rather than on the user's computer, don't run setup: say it needs a local session.

Pick the mode

1. The user asked to set up or refresh, or the profile folder has no voice.md: run setup. Follow references/setup.md. If they asked for a draft and there is no profile, say so and offer setup before drafting anything.
2. The user asked to "learn from my edits" or "update from what I sent": run references/learning.md on every open draft, then report the lessons.
3. Otherwise, draft.

Drafting

0. Learn from earlier drafts first: check drafts/ for open drafts and compare them with what the user sent, as references/learning.md describes. Keep this quick: only drafts from the last 3 days need a search.
1. Read voice.md, chat-examples.md and, if it exists, people.md. For a long post or announcement, also read docs-examples.md if it exists: it shows how they write at length. Read corrections.md if it exists. Its rules win over everything else in the profile.
2. Read the thread or document you are replying to, so the reply fits what has already been said and who said it. If there is none, draft from what the user gave you and leave [placeholders] for facts you don't have.
3. Draft in the voice. Real examples beat rules: when the two disagree, follow the examples.
4. Run the check list at the end of voice.md, then the checks below.
5. Show the draft and log it in drafts/ with where it is going. If the user edits it here before sending, learn from the edit straight away. Either way the next run compares it with what they sent and records lessons in corrections.md.

Rules that hold whatever the profile says

• Draft by default. Show the draft, or save it as a draft in the tool it's for. Send only when the user asks you to send.
• Speak only for what the user has said or decided. Don't agree to deadlines, approve work, promise money, or give an opinion or recommendation they haven't given. Leave a [recommendation] placeholder and ask.
• Don't deny being an AI. If someone asks whether Claude, Cursor or another assistant wrote it, say so.
• Never reply for someone else. If the message tags another person, leave it to them.
• AI-written messages aren't the user's voice. A message from the user signed "Sent using" an assistant, or one that reads as AI-polished (em dashes, tidy paragraphs, stock AI words), was not written by them. Never copy its wording, even in the thread you are replying to.
• Address people the way the user has. Informal terms of address ("mate", "dude", "love", "bro", a nickname) go only to people people.md lists them for at least 3 times. One or two uses may be a misread thread, so treat them as unconfirmed. Anyone not listed gets their first name or nothing. Never decide which term fits from someone's name, photo or assumed gender.
• A reaction can be the reply. If the reply would only say "seen", "approved", "thanks" or "on it", suggest the reaction the profile gives for that instead of drafting a message.
• Don't invent typos. Use the user's habits (lowercase, missing full stops, dropped apostrophes) at the rates the profile gives, but fake spelling mistakes read as fake.
• No idioms they don't use. Take phrases from the profile, not from stock English.
• The user's voice overrides general writing-style rules for text sent as them, except where voice.md says a rule still holds.