Skip to content
Back to search

sound-like-me

SkillWorks with Claude CodeGitHub

Draft messages to other people in the user's own voice, and learn that voice from messages they actually sent. Use whenever the user asks you to write, draft, reply to, or rephrase a message for someone else (Slack, Teams, email, WhatsApp, LinkedIn, PR or issue comments), or says things like "reply to him", "send it like how I talk", "in my tone", "make it sound like me", "learn my voice" or "update my voice". Also use when the user rewrites a draft you made, so the correction gets saved. Not for code, docs, commit messages, or how you talk to the user directly.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/r0wh4n--sound-like-me/badge.svg)](/listings/r0wh4n--sound-like-me)

HTML

<a href="/listings/r0wh4n--sound-like-me"><img src="/api/listings/r0wh4n--sound-like-me/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

Claude Code

git clone https://github.com/r0wh4n/sound-like-me ~/.claude/skills/sound-like-me

Then say **"learn my voice"**.

Codex

Clone it anywhere, then point your global `AGENTS.md` at it. If you already cloned it for Claude Code, use that path instead.

git clone https://github.com/r0wh4n/sound-like-me ~/sound-like-me
cat >> ~/.codex/AGENTS.md <<'EOF'

## Messages to other people
When drafting or replying to a message for someone else, follow ~/sound-like-me/SKILL.md.
EOF

ChatGPT / Claude.ai

Build `VOICE.md` once in Claude Code or Codex. If you use neither, paste [`setup.md`](setup.md) into a chat along with 15 to 20 messages you sent, and ask it to produce the file. Then:

- **Custom instructions:** paste the *Always*, *Never* and *Spellings* sections. The field has a character limit, so these come first. - **Projects:** upload the full `VOICE.md` as a project file in the project where you draft messages.

Cursor, Windsurf, anything else

Add a rule:

When drafting messages to other people, follow ~/.voice/VOICE.md.

If the tool can't read files outside the project, paste the file's contents into the rule instead.

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 43 lines
  1. ---
  2. name: sound-like-me
  3. description: Draft messages to other people in the user's own voice, and learn that voice from messages they actually sent. Use whenever the user asks you to write, draft, reply to, or rephrase a message for someone else (Slack, Teams, email, WhatsApp, LinkedIn, PR or issue comments), or says things like "reply to him", "send it like how I talk", "in my tone", "make it sound like me", "learn my voice" or "update my voice". Also use when the user rewrites a draft you made, so the correction gets saved. Not for code, docs, commit messages, or how you talk to the user directly.
  4. ---
  5. # Sound like me
  6. The person's voice lives in one file: `~/.voice/VOICE.md`. Every agent they use reads that same file, so it stays plain markdown with nothing tool-specific in it.
  7. ## Which job is this?
  8. - **No `~/.voice/VOICE.md` yet, or they ask to learn or rebuild their voice**: read `setup.md` in this skill's folder and follow it.
  9. - **They want a message drafted**: see Drafting.
  10. - **They changed your draft, pasted what they really sent, or said "no, I'd say it like..."**: see Corrections.
  11. ## Drafting
  12. 1. Read `~/.voice/VOICE.md`. If it doesn't exist, say so in one line and offer to set it up. For this one message, write plainly and briefly rather than guessing at a voice.
  13. 2. Work out the audience: who it's going to and where (peer DM, manager, team channel, someone external). Use the matching section under "By audience". If it's unclear, take the closest one and name it in a few words instead of stopping to ask.
  14. 3. Write the draft. Follow Always and Never. Use each spelling or quirk at the rate the file gives. People use their quirks some of the time, and a draft that crams every quirk into every line reads like a parody, which is worse than plain text.
  15. 4. Before showing it, hold it next to two or three examples from the same audience. Similar length? Same casing and punctuation? Same language mix? Anything from Never? Fix whatever doesn't match.
  16. 5. Show the draft ready to copy, with at most one line around it.
  17. The voice changes how something is said, not what is said. Keep facts, numbers, names, dates and links exactly as given.
  18. Don't send it. Sending is the person's call, even when you have a tool that can, unless they told you to send this specific message. A message in someone's voice that they never saw is impersonation, and it can't be unsent.
  19. The voice is for messages to other people. Keep talking to the user the way you normally do, and leave code, docs and commit messages alone.
  20. ## Corrections
  21. The gap between your draft and what they actually send is the best signal there is. When you see one:
  22. 1. Append a line under `## Corrections` in `~/.voice/VOICE.md`:
  23. `- YYYY-MM-DD · <audience>: "<what you wrote>" → "<what they sent>" (<the lesson in a few words>)`
  24. 2. Apply the lesson to the current draft straight away.
  25. 3. Once Corrections reaches about 20 lines, fold them in: turn repeated lessons into Always, Never or Spellings rules, adjust any rates that were off, then delete the folded lines. Tell the person in one line what changed.
  26. When they share a fresh message they sent that shows the voice well, add it to Examples and drop the oldest one from the same audience. Keep Examples to 20 at most, because the whole file is read on every draft.
  27. ## Privacy
  28. VOICE.md holds only the person's own messages. Before saving any example or correction, swap other people's names for roles (`<teammate>`, `<manager>`), and strip emails, phone numbers, links, ticket IDs and anything that looks like a credential. The file stays on their machine. Don't paste it into another service unless they ask.
README.md · README · 170 lines
  1. <div align="center">
  2. <img src="assets/logo.png" width="160" alt="sound-like-me logo: a speech bubble holding a sound wave that turns into a fingerprint">
  3. # sound-like-me
  4. **Your AI drafts messages that sound like you, not like an AI.**
  5. A skill that learns how you actually write to people, from messages you've sent,<br>
  6. and gets Claude, Codex, ChatGPT or any other agent to draft in that voice.<br>
  7. One file, every tool.
  8. [![Claude Code](https://img.shields.io/badge/Claude_Code-skill-D97757?style=flat-square)](#claude-code)
  9. [![Codex](https://img.shields.io/badge/Codex-AGENTS.md-6E56CF?style=flat-square)](#codex)
  10. [![ChatGPT](https://img.shields.io/badge/ChatGPT-custom_instructions-10A37F?style=flat-square)](#chatgpt--claudeai)
  11. [![License: MIT](https://img.shields.io/badge/license-MIT-blue?style=flat-square)](LICENSE)
  12. </div>
  13. ---
  14. ## Same ask, two people
  15. | | Reply |
  16. |----------------------------------|-------|
  17. | **Teammate asks** "can you review my PR today?"<br>*default AI* | Hi! Absolutely, I'd be happy to review your PR today. I'll take a look this afternoon and share my feedback. Let me know if there's anything specific you'd like me to focus on! |
  18. | *with sound-like-me* | haan dekhta hu<br>shaam tak<br>kuch specific h toh bata |
  19. | **Manager asks** the same thing<br>*with sound-like-me* | Will review it by EOD today. I'll flag if anything blocks the merge. |
  20. <sub>The peer reply is Roman Hinglish: "yeah I'll look / by evening / tell me if there's anything specific". It's whatever *you* write: English, Hinglish, Spanglish, all lowercase, all business.</sub>
  21. ## How it works
  22. ```mermaid
  23. flowchart LR
  24. A["Messages you sent<br/>Slack · email · PRs"] --> B["Learn"]
  25. B --> C[("~/.voice/VOICE.md")]
  26. C --> D["Draft"]
  27. D --> E{"You edit it<br/>before sending?"}
  28. E -- "yes" --> F["Correction saved"]
  29. F --> C
  30. E -- "no" --> G["You send it"]
  31. ```
  32. 1. **Learn.** It reads 15 to 60 messages you sent to real people, leaves out anything an AI wrote for you, asks a few quick questions, and writes your voice to `~/.voice/VOICE.md`. It then tests itself on three messages it held back and fixes what you say is off.
  33. 2. **Draft.** Ask for a message. It works out who it's for, writes it in your voice for that audience, and checks it against your real examples before showing it.
  34. 3. **Correct.** When you change a draft before sending, the change is saved. Every 20 or so corrections get folded into the rules, so drafts keep getting closer.
  35. ## Why not just "write like me" in a prompt?
  36. | What it does | Why it matters |
  37. |---|---|
  38. | **Learns from what you send, not what you prompt** | You write to an AI differently from how you write to people. A voice learned from your prompts comes out wrong. |
  39. | **Skips what AI wrote for you** | If you send through an assistant, part of your "sent" history is AI. It flags those messages and asks, so it learns you, not itself. |
  40. | **One voice per audience** | You don't text your teammate the way you email a vendor. Each audience gets its own section. |
  41. | **Quirks at your real rate** | It notes *how often* you write "h" for "hai" or skip punctuation, so drafts don't turn into a parody of you. |
  42. | **Gets better from your edits** | The gap between its draft and what you actually sent is the best signal there is, and it keeps it. |
  43. | **One plain file** | `VOICE.md` is plain markdown. Every agent you use reads the same file. |
  44. | **Drafts, never sends** | Sending stays your call. |
  45. ## Install
  46. ### Claude Code
  47. ```bash
  48. git clone https://github.com/r0wh4n/sound-like-me ~/.claude/skills/sound-like-me
  49. ```
  50. Then say **"learn my voice"**.
  51. ### Codex
  52. Clone it anywhere, then point your global `AGENTS.md` at it. If you already cloned it for Claude Code, use that path instead.
  53. ```bash
  54. git clone https://github.com/r0wh4n/sound-like-me ~/sound-like-me
  55. ```
  56. ```bash
  57. cat >> ~/.codex/AGENTS.md <<'EOF'
  58. ## Messages to other people
  59. When drafting or replying to a message for someone else, follow ~/sound-like-me/SKILL.md.
  60. EOF
  61. ```
  62. ### ChatGPT / Claude.ai
  63. Build `VOICE.md` once in Claude Code or Codex. If you use neither, paste [`setup.md`](setup.md) into a chat along with 15 to 20 messages you sent, and ask it to produce the file. Then:
  64. - **Custom instructions:** paste the *Always*, *Never* and *Spellings* sections. The field has a character limit, so these come first.
  65. - **Projects:** upload the full `VOICE.md` as a project file in the project where you draft messages.
  66. ### Cursor, Windsurf, anything else
  67. Add a rule:
  68. ```text
  69. When drafting messages to other people, follow ~/.voice/VOICE.md.
  70. ```
  71. If the tool can't read files outside the project, paste the file's contents into the rule instead.
  72. ## Things to say
  73. | You say | It does |
  74. |---|---|
  75. | `learn my voice` | Collects your sent messages, asks a few questions, writes `VOICE.md`, tests it |
  76. | `reply to this in my tone: <message>` | Drafts a reply for that audience |
  77. | `slack the team that the deploy moves to 6pm` | Drafts a channel announcement in your channel voice |
  78. | `no, i'd say "kal tak", not "by tomorrow"` | Saves a correction and redoes the draft |
  79. | `here's what i actually sent: <message>` | Compares it with its draft and learns from the difference |
  80. | `update my voice` | Re-learns from fresh messages and keeps your corrections |
  81. ## What `VOICE.md` looks like
  82. <details>
  83. <summary>A trimmed example (made-up person)</summary>
  84. ```markdown
  85. # Voice: Arjun
  86. _Updated 2026-11-21 · learned from 42 sent messages_
  87. ## Always
  88. - lowercase, no full stop at the end of a line
  89. - several short lines instead of one paragraph in DMs
  90. ## Never
  91. - emojis, em dashes, "hope this helps", "please revert"
  92. - bullet points in chat
  93. ## Spellings and quirks
  94. | they write | instead of | how often |
  95. |------------|------------|---------------|
  96. | h | hai | most messages |
  97. | kr | kar | about half |
  98. | pls | please | rarely |
  99. ## By audience
  100. ### Peer (DM)
  101. - length: 1 to 3 short lines
  102. - language mix: mostly Hinglish
  103. ### Manager
  104. - length: 1 to 2 sentences, sentence case
  105. - language mix: English; leads with the date or the blocker
  106. ## Corrections
  107. - 2026-11-20 · peer: "I'll check and get back" → "dekhta hu" (prefer the short form)
  108. ```
  109. </details>
  110. ## Privacy
  111. - Everything stays on your machine in `~/.voice/VOICE.md`.
  112. - Only your own messages are saved. Other people's names become roles (`<teammate>`, `<manager>`), and emails, phone numbers, links and anything that looks like a credential are stripped.
  113. - Nothing is sent unless you say so for that specific message.
  114. ## What's in here
  115. ```text
  116. sound-like-me/
  117. ├── SKILL.md drafting + corrections (loaded on every draft)
  118. └── setup.md learning your voice (loaded only when building the file)
  119. ```
  120. ## License
  121. [MIT](LICENSE)

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.