Why AI text is full of ** and ##
Chat assistants write their answers in Markdown, a plain-text shorthand for formatting. Two asterisks on each side of a word mean bold. A line that starts with ## is a heading. A dash at the start of a line is a bullet. The chat window reads those symbols and shows you the formatted version, so you never see them there.
The trouble starts when the text leaves the chat. Depending on the app and how you copy, you get either the formatted version or the raw Markdown underneath it. Gmail, Outlook, Word, LinkedIn and most other places do not read Markdown, so they show the symbols exactly as typed: **Next steps** instead of bold text, ## Summary instead of a heading, and a dash or an asterisk in front of every list item.
| What you see | What it was meant to be |
|---|---|
**text** or __text__ | Bold |
*text* or _text_ | Italic |
#, ## or ### at the start of a line | A heading, from biggest to smallest |
- or * at the start of a line | A bulleted list |
1. at the start of a line | A numbered list |
`text` | Code or a technical term |
[text](https://…) | A link |
Lines of | a | b | | A table |
The cleaner reads that Markdown the same way the chat window does, then writes it out in the form your destination understands. It removes symbols, never words.
What each destination gets
There is no single right way to clean the text, because every app accepts formatting differently. Pick where you are pasting and the output changes to match.
Plain text
For anywhere formatting will not survive: a text message, a form field, a support ticket, a code comment. Symbols go, bullets become •, numbered lists keep their numbers, and a link becomes its text followed by the address in brackets, or just the text if you switch that off. Code blocks stay exactly as they were.
Pasted
## Next steps - **Ship** the pricing page - Read [the notes](https://example.com)
Cleaned
Next steps • Ship the pricing page • Read the notes (https://example.com)
Gmail and Outlook
Email understands formatting, so this tab copies real formatting: bold stays bold, lists stay lists, tables arrive as tables and links stay clickable. Headings become bold lines a little larger than the body text, because a full-size heading in an email reads like shouting. Copy puts both a formatted and a plain version on the clipboard, and the email app uses the one it can.
Pasted
**Status:** on track | Task | Owner | |---|---| | QA pass | Lee |
Cleaned
Status: on track
| Task | Owner |
|---|---|
| QA pass | Lee |
Word and Google Docs
Documents get real headings. A ## line becomes a heading that Word and Google Docs recognize, so it takes their heading style and appears in the navigation pane or the document outline. Tables become editable tables. If you would rather have a file, Download .docx writes the same structure into a Word document on your device.
Pasted
## Summary Revenue grew **12%** this quarter.
Cleaned
Summary
Revenue grew 12% this quarter.
A LinkedIn post is plain text, so there is no real bold. The usual workaround is Unicode bold letters: separate characters that look bold. The LinkedIn tab converts bold words and headings to them; switch that off if you prefer normal letters. Use them sparingly, because some screen readers read these characters oddly and search may not match them. The counter shows how much of LinkedIn's 3,000-character limit you have used, and it counts each bold letter as two characters to stay on the safe side.
Pasted
**Big news:** we launched today
Cleaned
𝗕𝗶𝗴 𝗻𝗲𝘄𝘀: we launched today
Slack
Slack has its own formatting code, close to Markdown but not the same. Bold uses one asterisk instead of two, there are no headings, and links are written as <address|text>. The Slack tab rewrites the Markdown into Slack's version. Tables, which Slack cannot show, become aligned columns in a code block.
Pasted
## Heads up **Deploy** at 3pm, see [runbook](https://example.com)
Cleaned
*Heads up* *Deploy* at 3pm, see <https://example.com|runbook>
How to get AI to skip Markdown in the first place
You can ask for plain text up front. A specific instruction works better than “no formatting”:
Write in plain text with no Markdown: no asterisks, no pound signs, no bullet symbols and no tables. Use short paragraphs, and put each list item on its own line starting with a number.
- ChatGPT: add the instruction to your custom instructions so it applies to every new chat, or paste it at the top of a single prompt.
- Claude: put it in your profile preferences, or in a project's instructions if you only want it there.
- Gemini: include it in the prompt, or save it in a Gem's instructions for the tasks where you need plain text.
Expect some drift. In long conversations, models tend to slide back into Markdown, especially for lists and comparisons. Repeating the instruction helps, and so does running the result through this page.
Invisible characters, and why they matter
Text copied from chat apps, web pages and PDFs can carry characters you cannot see. A zero-width space sits between two letters and takes up no room at all. A word joiner or a byte-order mark does the same. A narrow no-break space looks like an ordinary space but is a different character.
They cause quiet problems. Search will not find “on track” if a zero-width space hides inside it. A link, an email address or a spreadsheet formula with one inside stops working. Code fails with an error that points at nothing. Word counts and character limits come out wrong.
The cleaner removes zero-width characters and hidden direction marks, turns unusual spaces into normal ones, and tells you how many it found, so you know whether anything was there. It leaves alone the joiners that hold emoji together (a family emoji is several emoji joined by invisible characters) and the ones that languages such as Persian and Hindi need to spell words correctly.
What the cleaner never changes
It does not rewrite, shorten or “humanize” anything: every word you paste comes out the same. The options that do touch characters, replacing em dashes, straightening curly quotes, removing emojis and removing citation markers such as [1], stay off until you turn them on. Em dashes are a matter of style, and some readers now take them as a sign of AI writing, so that option is there if you want it.
Everything happens on your device. The page loads once; after that, the text you paste is processed by code running in your browser and is never uploaded, stored or logged.