Dictation Cleanup
When to use this
Invoke this skill when:
- The user pastes raw STT (speech-to-text) transcript and asks for it to be cleaned.
- The input text shows clear dictation artifacts: run-on sentences, missing punctuation, filler words, number words that should be digits (or vice versa), homophone errors.
- The user's message is "clean this up" / "fix my dictation" / "format this memo" when the input has dictation-quality characteristics.
The canonical use case: a lawyer dictates a memo on their phone while commuting. Louis cleans it into a draft the lawyer can review and send.
What to clean
1. Remove filler words
Filler words are artifacts of spoken language that should not appear in written legal text:
- Remove: "um", "uh", "like" (used as a filler), "you know", "I mean", "so basically", "right", "okay so"
- Do not remove: "like" when used as a genuine comparison word ("similar to", "such as"); "right" when used as an adjective or as part of a phrase like "right of first refusal."
2. Fix STT recognition errors
Common misrecognitions to correct:
| STT output |
Likely intended |
| "their" / "there" / "they're" |
Correct per context |
| "its" / "it's" |
Correct per context |
| "statute" → "statue" |
Correct when legal context is clear |
| "clause" → "claws" |
Correct in legal context |
| "indemnify" → "in dim a fy" |
Re-merge recognized segments |
| Numbers as words: "twenty twenty four" |
Convert: "2024" |
| Partial proper noun: "dee eye ef see" |
Reconstruct: "DIFC" |
| Run-together entities: "uae federal" |
Restore capitalization: "UAE Federal" |
Be conservative with STT error correction — when the correct word is ambiguous, flag it with [??] rather than guessing.
3. Reconstruct punctuation
STT outputs are typically unpunctuated run-ons. Apply:
- Sentence-ending periods where the speaker's cadence, topic shift, or completed thought indicates a sentence boundary.
- Commas where the speaker's pause marks a clause boundary.
- Quotation marks around anything that was clearly quoted material.
- Paragraph breaks where the speaker shifts to a new topic.
Do not over-punctuate. If in doubt, use a period and start a new sentence.
4. Correct capitalization
- Sentence starts.
- Proper nouns: jurisdiction names (UAE, Lebanon, KSA), party names, defined terms.
- Defined terms: once identified, ensure consistent capitalization throughout (e.g., "the Agreement", "the Buyer").
5. Standardize numbers and dates
| Dictated |
Written form |
| "twenty-twenty-four" or "twenty twenty four" |
2024 |
| "thirty days" (in a legal obligation) |
30 days |
| "three million dollars" |
USD 3,000,000 (in legal docs) or $3 million (in casual memo) |
| "the twelfth of May" |
May 12, 2026 (or 12 May 2026 per house style) |
| "ninety percent" |
90% |
Preserve the lawyer's chosen number format if they stated it explicitly ("I want to write it out as thirty days" should remain "thirty days").
6. Detect and format defined terms
When the speaker uses a term multiple times and it is clearly a defined contractual term (party name, document title, key concept), capitalize it consistently:
- "the buyer" (first use) → detect pattern → "the Buyer" throughout.
- "the lease" when clearly referring to a specific document → "the Lease."
Flag new defined terms with a comment: [Note: "the Seller" treated as a defined term — confirm capitalization is correct.]
What to preserve
These elements must not be changed:
- Substantive legal content: do not add, remove, or restate obligations, conditions, or facts that the speaker expressed. Even if it seems like a logical omission, do not fill it in.
- Numbers and amounts: copy exact figures verbatim after standardizing format. Do not round, estimate, or substitute.
- Specific case citations: if the speaker cited a case or statute, reproduce it exactly as given. If it appears garbled, flag it with
[??] — do not reconstruct a plausible citation.
- Speaker's conclusion: do not soften, strengthen, or reframe legal opinions the speaker expressed.
- Structure: if the speaker clearly said "first... second... third...", preserve that list structure.
Output options
Provide the cleaned text as the primary output.
Optionally, include a change-tracked version that shows what was modified — this is valuable when the lawyer wants to verify every change before signing off. Format as:
[REMOVED: "um"] / [CHANGED: "twenty twenty four" → "2024"] / [ADDED PUNCT: "."]
Or use markdown strikethrough and inline additions if the rendering surface supports it.
Limits
- Do not correct substantive legal errors in the dictation (e.g., the lawyer cited the wrong law). Flag them:
[Note: you may want to verify the statute reference here.]
- Do not translate the text into a different language unless explicitly asked.
- Do not convert the text into a formal document structure (headings, numbered clauses) unless the lawyer asked for that — cleanup and reformatting are different tasks.
Related skills
- [[voice-short-spoken-output]]
- [[voice-voice-friendly-short]]
- [[voice-image-of-contract-page-handler]]
- [[voice-multimodal-scanned-pdf-handler]]
1---2name: voice-dictation-cleanup3description: Use when a lawyer or legal professional has dictated text via voice-to-text (STT) and the raw transcript needs to be cleaned into polished, formal written output. This skill governs filler removal, STT error correction, punctuation reconstruction, defined-term formatting, and the strict constraints on what must not be changed — preserving legal substance and numeric precision exactly while removing transcription artifacts.4license: MIT5---67# Dictation Cleanup89## When to use this1011Invoke this skill when:12- The user pastes raw STT (speech-to-text) transcript and asks for it to be cleaned.13- The input text shows clear dictation artifacts: run-on sentences, missing punctuation, filler words, number words that should be digits (or vice versa), homophone errors.14- The user's message is "clean this up" / "fix my dictation" / "format this memo" when the input has dictation-quality characteristics.1516The canonical use case: a lawyer dictates a memo on their phone while commuting. Louis cleans it into a draft the lawyer can review and send.1718## What to clean1920### 1. Remove filler words21Filler words are artifacts of spoken language that should not appear in written legal text:22- Remove: "um", "uh", "like" (used as a filler), "you know", "I mean", "so basically", "right", "okay so"23- Do not remove: "like" when used as a genuine comparison word ("similar to", "such as"); "right" when used as an adjective or as part of a phrase like "right of first refusal."2425### 2. Fix STT recognition errors26Common misrecognitions to correct:2728| STT output | Likely intended |29|-----------|----------------|30| "their" / "there" / "they're" | Correct per context |31| "its" / "it's" | Correct per context |32| "statute" → "statue" | Correct when legal context is clear |33| "clause" → "claws" | Correct in legal context |34| "indemnify" → "in dim a fy" | Re-merge recognized segments |35| Numbers as words: "twenty twenty four" | Convert: "2024" |36| Partial proper noun: "dee eye ef see" | Reconstruct: "DIFC" |37| Run-together entities: "uae federal" | Restore capitalization: "UAE Federal" |3839Be conservative with STT error correction — when the correct word is ambiguous, flag it with `[??]` rather than guessing.4041### 3. Reconstruct punctuation42STT outputs are typically unpunctuated run-ons. Apply:43- Sentence-ending periods where the speaker's cadence, topic shift, or completed thought indicates a sentence boundary.44- Commas where the speaker's pause marks a clause boundary.45- Quotation marks around anything that was clearly quoted material.46- Paragraph breaks where the speaker shifts to a new topic.4748Do not over-punctuate. If in doubt, use a period and start a new sentence.4950### 4. Correct capitalization51- Sentence starts.52- Proper nouns: jurisdiction names (UAE, Lebanon, KSA), party names, defined terms.53- Defined terms: once identified, ensure consistent capitalization throughout (e.g., "the Agreement", "the Buyer").5455### 5. Standardize numbers and dates56| Dictated | Written form |57|----------|-------------|58| "twenty-twenty-four" or "twenty twenty four" | 2024 |59| "thirty days" (in a legal obligation) | 30 days |60| "three million dollars" | USD 3,000,000 (in legal docs) or $3 million (in casual memo) |61| "the twelfth of May" | May 12, 2026 (or 12 May 2026 per house style) |62| "ninety percent" | 90% |6364Preserve the lawyer's chosen number format if they stated it explicitly ("I want to write it out as thirty days" should remain "thirty days").6566### 6. Detect and format defined terms67When the speaker uses a term multiple times and it is clearly a defined contractual term (party name, document title, key concept), capitalize it consistently:68- "the buyer" (first use) → detect pattern → "the Buyer" throughout.69- "the lease" when clearly referring to a specific document → "the Lease."7071Flag new defined terms with a comment: `[Note: "the Seller" treated as a defined term — confirm capitalization is correct.]`7273## What to preserve7475These elements must not be changed:7677- **Substantive legal content**: do not add, remove, or restate obligations, conditions, or facts that the speaker expressed. Even if it seems like a logical omission, do not fill it in.78- **Numbers and amounts**: copy exact figures verbatim after standardizing format. Do not round, estimate, or substitute.79- **Specific case citations**: if the speaker cited a case or statute, reproduce it exactly as given. If it appears garbled, flag it with `[??]` — do not reconstruct a plausible citation.80- **Speaker's conclusion**: do not soften, strengthen, or reframe legal opinions the speaker expressed.81- **Structure**: if the speaker clearly said "first... second... third...", preserve that list structure.8283## Output options8485Provide the cleaned text as the primary output.8687Optionally, include a **change-tracked version** that shows what was modified — this is valuable when the lawyer wants to verify every change before signing off. Format as:88```89[REMOVED: "um"] / [CHANGED: "twenty twenty four" → "2024"] / [ADDED PUNCT: "."]90```9192Or use markdown strikethrough and inline additions if the rendering surface supports it.9394## Limits9596- Do not correct substantive legal errors in the dictation (e.g., the lawyer cited the wrong law). Flag them: `[Note: you may want to verify the statute reference here.]`97- Do not translate the text into a different language unless explicitly asked.98- Do not convert the text into a formal document structure (headings, numbered clauses) unless the lawyer asked for that — cleanup and reformatting are different tasks.99100## Related skills101102- [[voice-short-spoken-output]]103- [[voice-voice-friendly-short]]104- [[voice-image-of-contract-page-handler]]105- [[voice-multimodal-scanned-pdf-handler]]
sboghossian