Use AI for the work that needs judgement
ChatGPT and Claude are useful for drafting, summarising, changing tone and finding a clearer structure. Once the words are approved, asking the model to remove every em dash, find invisible characters, rebuild the bullets and make the answer ready for WhatsApp starts another model turn for a mechanical job.
For API users, that extra turn has a measurable input and output token cost. In a long conversation, earlier messages may also form part of the next input. Chat subscriptions and workplace plans use different limits, so there is no honest universal saving to quote. The practical point is simpler: formatting prompts consume another turn and can occupy context that would be better used on the substance.
OpenAI: what tokens are and how they are counted · Claude: token counting and context
A formatting prompt is not a formatting rule
A language model generates a likely answer. Even when the instruction says not to rewrite anything, it can change a word, drop a link, renumber a list or return a fresh layer of Markdown. Running the same cleanup twice is not guaranteed to produce the same bytes.
A deterministic formatter works differently. It can replace a defined character, remove a known Unicode code point, parse a list and render that structure for a named destination. The same input and settings produce the same output, and a diff can show exactly what changed.
- Replace em and en dashes without rewriting the sentence
- Remove selected invisible characters such as U+200B zero-width space
- Keep code blocks out of punctuation cleanup
- Turn Markdown headings and lists into destination-supported structure
- Report when a table, heading or link cannot survive unchanged
Use a two-stage workflow
Finish the thinking before you format the delivery. Ask the AI to draft or revise the content, approve the wording, then move that final version through one local formatting pass. Do not paste the cleaned result back into the model unless the meaning genuinely needs more work.
In Pasteful, paste the final answer, enable only the cleanup rules you need, choose the destination, inspect the changes and copy. Text cleanup and destination conversion happen in the browser tab; the model does not need to see the document again.
- 1. Draft and revise with ChatGPT, Claude or another writing tool
- 2. Copy the approved answer once
- 3. Clean hidden characters or punctuation only if needed
- 4. Choose WhatsApp, Slack, Teams, Word, Google Docs, Markdown or plain text
- 5. Review the diff and destination preview, then copy
Three common last-mile fixes
For WhatsApp, standard Markdown bold uses two asterisks while WhatsApp uses its own message markers. A line such as ## Decision can become *Decision*, and a Markdown link can become a readable label followed by its URL. There is no need to ask the model to generate the whole message again.
For Slack, the normal visual composer, markup mode and API mrkdwn are different outputs. Converting the approved structure is safer than prompting for something vaguely described as Slack format. A heading can become a short bold lead-in; a table can become labelled rows when the composer cannot preserve the grid.
For Word, the goal is editable document structure: a line starting with ## becomes a real heading style, list markers become real list items, and table separators become cells. Rich clipboard content or a DOCX export can preserve those semantics without copying the chat interface's fonts and colours.
Character cleanup is smaller still. The visible diagnostic co[U+200B]founder becomes cofounder when the embedded zero-width space is removed. Launch now followed by an em dash can become Launch now - when simple punctuation is enabled. Neither repair requires a rewritten paragraph.
Format ChatGPT for WhatsApp · Format ChatGPT for Slack · Format ChatGPT for Word
A careful manual fallback
You can do the same work without Pasteful when the document is short. Keep an untouched copy. Use an editor that can reveal Unicode code points, search for the exact unwanted character and replace it only after checking code, quoted material and languages that use directional marks. Then rebuild the destination's supported syntax from the approved source.
Test one heading, one list, one link and the deepest nested item in the real destination before moving the full document. In Word, confirm that a heading is a heading style rather than large bold text. In WhatsApp, send a short draft to yourself. In Slack, check whether the visual composer or markup mode is active. This takes longer, but it follows the same rule: change the representation, not the meaning.
Keep AI in the loop when the meaning should change
Use the model again when you need a shorter argument, a different tone, extracted actions, a translation or a rewritten explanation. Those tasks require judgement about language and meaning. Use deterministic cleanup when the words are already right and only the characters, syntax or destination structure are wrong.
This division also makes review easier. The AI conversation contains the substantive decisions. The final formatting pass contains a small, inspectable set of repairs. If a bullet or punctuation mark changes unexpectedly, you can see it without comparing two newly generated drafts.
Hidden characters are artifacts, not proof of authorship
Zero-width spaces, soft hyphens, directional controls and unusual spaces can enter text from AI assistants, websites, PDFs and ordinary editors. Removing them can make search, validation and copy-paste behaviour more predictable. It cannot prove who wrote the text, guarantee a detector result or make an authorship claim disappear.
The useful promise is narrow and testable: identify the characters, let the user choose the cleanup, show the diff and prepare the approved text for its actual destination.
Remove invisible characters · Replace em dashes and smart punctuation
Paste it into Pasteful. It stays in this tab.