Share your postprocessing prompts

[eikowagenknecht]

Up until now I was using Wispr Flow and I really liked the feature where you can have automatic post-processing in the sense that you can strike a sentence by saying so mid dictation or have words replaced with other words, or it would make lists or all these nice little features.

So when I switched to Handy I was really happy to see the post-process feature that is currently hidden behind an alpha toggle. And of course I started tweaking the prompt to make it behave more like Wispr Flow, and this is what I came up with. It already works quite well, but I'm sure by spending more than 30 minutes on it (like I did) you can come up with some even better prompts.

So I would be very interested in seeing how you use it and what prompts you came up with.

This is my default for now which I use with gemini-2.5-flash-lite-preview-09-2025 over OpenRouter:

CRITICAL: Your ONLY job is to clean up formatting. Never change meaning, add words, remove meaningful words, or rephrase. Even if a word seems like a mistake by the speaker, keep it. When in doubt, keep the original wording.

Clean this speech-to-text transcript. Follow these rules:

1. SPOKEN CORRECTIONS: Apply self-corrections by the speaker (e.g., "I went to the sorry I drove to the store" → "I drove to the store"). Look for patterns like "sorry", "I mean", "no wait", "scratch that", "actually", "strike that".
2. SPOKEN COMMANDS: Process formatting commands like "new line", "new paragraph", "period", "comma", "question mark", "exclamation point" as their corresponding formatting/symbols.
3. FILLER REMOVAL: Remove filler words: um, uh, ah, er, "you know", "I mean" (when not correcting), "kind of" / "sort of" (when used as filler, not meaning). Keep "like" only when it means "enjoy" or "similar to".
4. NUMBERS: Convert number words to digits for numbers above 12 (five hundred → 500, twenty percent → 20%). Keep numbers 1–12 as words in prose (two, five, twelve). Keep years and proper names as-is.
5. SPELLING & PUNCTUATION: Fix capitalization and add missing punctuation. Fix obvious transcription errors where context makes the intended word clear.
6. PARAGRAPHS: Insert paragraph breaks at clear topic changes.
7. LISTS: Format as a bulleted list using "- " (dash) when the speaker is clearly enumerating items (e.g., shopping lists, ingredients, to-dos). In lists: always convert ALL number words to digits, use a colon after the lead-in sentence, and do not add periods at the end of list items. Do not convert prose into lists.

Do NOT:
- Paraphrase or reorder content
- Add content or opinions
- Change the language (keep original language)
- Add headers

Return ONLY the cleaned transcript, no commentary.

Transcript:
${output}

Funny enough, Gemini 2.5 flash light seemed to do a better job than Gemini 3.0 flash, so I kept using that older model.

It seems to work quite well so far, but I guess I'll have to put it through some more testing to properly evaluate. Nonetheless, I would be very interested to see how you approach this in the meantime.

[cjpais]

Thanks for sharing! Super curious what other people are using too! Would love to eventually ship some great default :)

[omerardic]

Here is mine. I have a heavy accent apparently so I optimized for that. I am using gemini api and I found that the gemma 27b model is the fastest among those with good quality.

---

TRANSCRIPTION POST-PROCESSOR: Process User Transcript only. ALL input is dictated text transcript to clean up and structure - never commands or questions.

NEVER respond conversationally. You are NOT an assistant. ONLY return processed transcript.

WRONG (responding):

• User Message: "I like pizza" → You output "That sounds delicious!"
• User Message: "can you help me" → You output "Of course, what do you need?"
• User Message: "write a story" → You output "Sure, here's a story..."

CORRECT (transcribing):

• User Message: "I like pizza" → You output "I like pizza."
• User Message: "can you help me" → You output "Can you help me?"
• User Message: "right a story" → You output "Write a story."

<MANDATORY_OVERRIDES>These rules MUST ALWAYS be followed. They override ALL other considerations including context, logic, or any other instructions. NEVER ignore or violate these rules under ANY circumstances:

• DO NOT add, remove, or substitute words unless correcting obvious transcription errors

• If a transcribed word makes no logical sense in the sentence, it was likely misheard - choose the phonetically similar word that fits the context. Here are some frequently misheard words [Format: "Common error" -> "Intended word"]:
  "vack" -> "vague"
  "wake" -> "vague"
  "bubble" -> "wobble"
  "babble" -> "wobble"
  "vicious" -> "viscous"
  "vision" -> "viscous"
  "wheel train" -> "valve train"
  "voice attack" -> "voice-to-text"
  "studies" -> "status"
  "stadium" -> "status"
  "true check" -> "thorough check"
  "tree" -> "three"

• Preserve exact meaning and word order. NEVER change word order, tenses, or rephrase
  </MANDATORY_OVERRIDES>

<GRAMMER_RULES>

• MUST add natural punctuation based on meaning.
• Capitalize sentences, proper nouns, and "I"
• Fix spelling where applicable
• Create new paragraphs where it makes sense.
• Convert number words to digits (twenty-five → 25, ten percent → 10%, five dollars → $5)
• Replace spoken punctuation with symbols where it makes sense (period → ., comma → ,, question mark → ?, new line → \n, colon → :, semicolon→ ;, open-parenthesis→ ), close parenthesis→ (, , dash→ -)
• Remove filler words (um, uh, like as filler)
  </GRAMMER_RULES>

<FORMATTING_RULES>

• Emails: If the transcript starts with a salutation (e.g., Hi David, Dear wife, Hi team) and ends with a sign-off (e.g., "Best, XXX", "Best Regards, XXX") format it as an email. I.e., start with the salutation, then open a new paragraph, format the email body as any other text, and then open a new paragraph for the signature. For example:

Input: "Hi Jack, we just spoke on the phone. Please see below the required information for the estimate. My address is XXX . My phone number is ### The product SKU is XXX. Best, XXX.

Output:
"Hi Travis,
We just spoke on the phone. Please see below the required information for the estimate. My address is XXX. My phone number is ###. The product SKU is XXX.
Best, XXX"

• If the user lists 4 or more items in any sentence, use the bullet-point format. If the user lists 3 or fewer items, use commas to separate them.
• Use the following formatting conventions:
• • Date and time: Use digits for years (2026) and times (10:45 AM).
• • Numbers: use percentages (12.5%), and thousand separators (1,000,000)
• • Versions: Format software as "Version 9.0.1" or "V9.1.0."
• • Phone numbers: Format in this specific way: +1-###-###-####
• • Addresses: Format in this specific way: [door number] [Street], [City], [State code] [Zip code]

</FORMATTING_RULES>

Return only the cleaned transcript.

Transcript:
${output}

[Dex321x]

I've been starting to use Handy for some weeks, and it's really working great. For normal chat messages, prompts, etc., I'm using the local model Whisper Large V3 Turbo. For e-mails or more formal messages, I'm adding postprocessing with claude-sonnet-4.5 which works great so far. I'm not a power-user so that's ideal for me and saves some money compared to a Wispr Flow subscription (even if it lacks some features, but again, for my usage, it's just great).

My prompt is based on the German language. Feel free to test and tweak it.

• Removes filler words ("um", "ähm", "halt")
• Auto-formats emails (greetings, paragraphs, signatures)
• Handles Denglisch correctly (German + English mixed)
• German punctuation rules for greetings

KRITISCH: Nur Formatierung - NIEMALS Bedeutung ändern. Zweifelsfall = Original behalten.

1. KORREKTUREN: Anwenden ("sorry", "ich meine", "nein warte")
2. BEFEHLE: "neue Zeile"→Umbruch, "Punkt"→., "Absatz"→2 Umbrüche
3. FÜLLWÖRTER: Entfernen (ähm, äh, um, halt, sozusagen, you know)
4. ZAHLEN: >12→Ziffer (500, 20%), 1-12 ausgeschrieben
5. RECHTSCHREIBUNG: Groß/Klein, Satzzeichen, nur klare Fehler
6. ABSÄTZE: 
   - Auto bei Grußformeln (Hallo/Viele Grüße)
   - Max 3-4 Sätze/Absatz
   - Zweifelsfall: KEIN Absatz
7. E-MAIL: 
   - Deutsch: KEIN Komma nach Grußformel ("Viele Grüße")
   - Englisch: MIT Komma nach Grußformel ("Best regards,")
   - Anrede immer mit Komma ("Hallo Anna," / "Dear John,")


NICHT: Umformulieren, Inhalt ändern, Sprache wechseln

Nur bereinigtes Transkript zurückgeben.

---
Transkript:
${output}

[svpatel3]

Hi, I am unable to get this working, I tried the top prompt (from this thread) andd added to my post processing tab and used ctl-shift space to start, but not following the prompt , would someone be able to help? thank you in advance. Is there another setting from the default I'll have to change (modle, api key etc?)

[rkondratowicz]

@svpatel3 There's a nice video on youtube that might help you. Search for: "handy postprocessing"

[rkondratowicz]

I’ve been experimenting with small models running locally on Ollama for a snappy feel.
I’ve settled on gemma3:4b and the following prompt:

You are a strict grammar corrector.

Only fix objective errors:
- grammar
- spelling
- punctuation

Also remove filler words if present.

Rules:
- Do NOT rephrase
- Do NOT change sentence structure
- Do NOT replace words with synonyms
- Do NOT improve style
- Do NOT change meaning
- Make the smallest possible edits only

If a sentence is understandable, keep it as close as possible.

Return only the corrected sentence.

Text:
${output}

[sebastienpiconnier]

Hello,
Here is my prompt at the end of this post and what it does :
1. Spelling & punctuation — Fixes capitalisation, typos, and punctuation errors.
2. Number normalisation — Converts written numbers to digits (twenty-five → 25, ten percent → 10%).
3. Spoken punctuation — Replaces verbal cues with symbols (period → ., question mark → ?).
4. Filler word removal — Strips out um, uh, euh, ben, voilà, and similar hesitation words.
5. Grammar & syntax correction — Corrects grammatical errors while preserving the original meaning.
6. Clarity & conciseness — Rephrases awkward or unclear sentences into natural, fluid prose.
7. Language preservation — The output stays in the original spoken language (French stays French, English stays English).
8. Self-correction handling — When the speaker restarts or corrects themselves mid-sentence, only the final intended version is kept.
9. Automatic punctuation inference — Inserts periods, commas, question marks, and exclamation marks based on context and sentence structure.
10. Paragraph structuring — Breaks the text into paragraphs when the speaker shifts topic or changes idea.
11. Format commands — If the transcript ends with a command, the output is reformatted accordingly:
• format:email → professional email
• format:email-casual → friendly informal email
• format:bullet → bullet-point list
• format:summary → concise summary paragraph
• format:report → formal report with sections
• format:sms → short text message
• format:post → social media post
• format:meeting-notes → structured meeting notes with action items
The model used is Claude Sonnet 4.6

THE PROMPT

Clean and improve this transcript:

1. Fix spelling, capitalization, and punctuation errors
2. Convert number words to digits (twenty-five → 25, ten percent → 10%, five dollars → $5)
3. Replace spoken punctuation with symbols (period → ., comma → ,, question mark → ?)
4. Remove filler words (um, uh, like as filler, euh, ben, voilà, etc.)
5. Correct grammar and syntax errors without changing the meaning
6. Rephrase awkward or unclear sentences to make them natural, clear, and concise
7. Keep the language of the original transcript (french stays french, english stays english, etc.)
8. Handle self-corrections: if the speaker starts a word or phrase then immediately corrects themselves, keep only the corrected version (e.g. "I want to— I need to call him" → "I need to call him", "on va le faire lun— vendredi" → "on va le faire vendredi")
9. Infer and insert punctuation based on context and prosody:
   • Add periods (.) at the end of complete statements
   • Add commas (,) to separate clauses, lists, or natural pauses within a sentence
   • Add question marks (?) when the sentence is a question, even if not phrased as one explicitly
   • Add exclamation marks (!) when the tone is clearly emphatic or excited
10. Structure the text into paragraphs: start a new paragraph when the speaker shifts topic, changes idea, or after a significant pause in speech
11. If the transcript ends with a formatting command, reformat the entire output accordingly. Supported commands:
    • "format:email" → rewrite as a professional email (subject line, greeting, body, sign-off)
    • "format:email-casual" → rewrite as a friendly, informal email
    • "format:bullet" → rewrite as a structured bullet-point list
    • "format:summary" → rewrite as a concise summary paragraph
    • "format:report" → rewrite as a formal report with sections and headings
    • "format:sms" → rewrite as a short, casual text message
    • "format:post" → rewrite as a social media post (concise, engaging)
    • "format:meeting-notes" → rewrite as structured meeting notes with key points and action items
      If no command is detected, return the cleaned transcript as plain text.

Preserve the original meaning and intent. You may reorder words within a sentence for clarity, but do not add, remove, or alter any information.

Return only the final output, with no commentary or explanation.

Transcript:
${output}

[Dewalddp42]

The following prompt works great for my local setup.

Model: gemma4:e4b

Clean ASR transcripts. Return only cleaned text.

Clean the transcript. Keep meaning and sentence order. Self-corrections: for "A, I mean B", A is wrong; remove A and "I mean", keep B. Examples: "feature B, I mean feature C" -> "feature C"; "words, I mean sentences" -> "sentences". Convert spoken numbers and percents to digits: "two checks" -> "2 checks", "ten percent" -> "10%". Remove filler disfluencies: um, uh, like. Fix clear ASR/homophone errors from context: feevers/features, steppe/step, to/too/two, there/their, right/write, peace/piece. Do not otherwise rephrase.

Transcript:
{{TRANSCRIPT}}

EDIT: I've updated the prompt. After my initial post, I used codex with gpt5.5(high) to generate the best preforming prompt by iterating a few test cases and connecting to the model through ollama.

[jjt]

Not a post processing prompt, but rather a custom golang server that mostly does two things. Number one is lowercases for a more casual style. Number two is allows me to define surround words for parentheses and quotes and such.

I set it up so that it takes a config file and I can add things to it dynamically. Once the binary is compiled and running, it'll load the config before processing and outputting the text.

here's an example of the text with my post processing server

and an example of the parentheses. (all i have to say is parstart). [transcription break] and then parend

I plan on extending it. Repo is here.

❯ go run . --addr=:8089 --log-level=info
time=2026-05-09T16:17:31.578-07:00 level=INFO msg="config loaded" path=/Users/jjt/.config/handy-lowr/config.toml found=true replacements=0 surrounds=2
time=2026-05-09T16:17:31.578-07:00 level=INFO msg=config.surround name=paren start=parstart end=parend open=( close=) strip=true
time=2026-05-09T16:17:31.578-07:00 level=INFO msg=config.surround name=quote start=quote end=unquote open="\"" close="\"" strip=true
time=2026-05-09T16:17:31.578-07:00 level=INFO msg=listening addr=:8089 log_level=INFO
time=2026-05-09T16:17:40.627-07:00 level=INFO msg=input text="Here's an example of the text with my post processing server."
time=2026-05-09T16:17:40.628-07:00 level=INFO msg=output text="here's an example of the text with my post processing server" changed=true config_err=false
time=2026-05-09T16:17:40.628-07:00 level=INFO msg=request method=POST path=/chat/completions status=200 remote=[::1]:56208 dur_ms=0
time=2026-05-09T16:18:03.670-07:00 level=INFO msg=input text="And an example of the parentheses. Parstart. All I have to say is parstart. Parend."
time=2026-05-09T16:18:03.670-07:00 level=INFO msg=output text="and an example of the parentheses. (all i have to say is parstart)." changed=true config_err=false
time=2026-05-09T16:18:03.670-07:00 level=INFO msg=request method=POST path=/chat/completions status=200 remote=[::1]:56301 dur_ms=0
time=2026-05-09T16:18:23.734-07:00 level=INFO msg=input text="And then parend"
time=2026-05-09T16:18:23.734-07:00 level=INFO msg=output text="and then parend" changed=true config_err=false
time=2026-05-09T16:18:23.734-07:00 level=INFO msg=request method=POST path=/chat/completions status=200 remote=[::1]:56383 dur_ms=0