speech_adaptation v1.0.0

name: speech_adaptation description: Adapting written text for spoken delivery: cutting sentences to breath length, replacing written signposts with spoken ones, marking pauses and emphasis, and timing the result against the target runtime. Use when a report, article, memo, doc, or web page has to become a talk, video script, podcast segment, voiceover, or audiobook read, or when a script reads fine on paper but sounds stiff, rushed, or hard to follow aloud. metadata: { "includeInPrompt": true }

Speech Adaptation

Purpose

Turn text that was written for the eye into text that works in the ear, on the first hearing, at a real human pace. Written and spoken language differ in set ways: spoken language uses shorter sentences, shorter words, more contractions, more repetition, and more self reference, because a listener cannot reread, pause the speaker to check a term, or see headings, brackets, and layout. This skill covers the rewrite itself: breath length, spoken signposting, delivery markup, spoken forms for numbers and abbreviations, and a timing check against the target runtime. Delivery coaching after the text is locked belongs to public speaking and teleprompter skills. Recording and mixing belong to voiceover production.

Workflow

1. Get the delivery facts first. Format (talk, video, podcast, voiceover, audiobook), target runtime in seconds, who speaks it (the author, another presenter, a voice actor, or a text to speech voice), and whether wording is locked anywhere (quotes, legal lines, prices, names). Locked lines stay exact and get markup only.
2. Run the checker on the source. Run bin/adapt_check.py <file> --format <format> --target-seconds <seconds> to get the word count, a first runtime estimate, long sentences, and written-only forms. Treat the estimate as a floor: marked pauses, emphasis, and applause or laugh beats add time the count cannot see.
3. Pick the pace for the format. Use these working rates unless the speaker has a measured rate of their own, which always wins: talk 130 wpm, keynote 120 wpm, explainer video 135 wpm, general video 140 wpm, podcast conversation 150 wpm, audiobook 150 to 160 wpm, formal voiceover 120 wpm. Rates vary a lot by speaker and task, so calibrate from a one minute timed read whenever the speaker is available.
4. Cut sentences to breath length. One idea per sentence. Aim for most sentences at 20 words or fewer, and split any sentence a reader cannot finish in one comfortable breath. Move the main verb early. A sentence that parks its verb at the end forces the listener to hold every clause in memory, and it fails aloud even when it reads well.
5. Rewrite for the ear. Add contractions where the register allows them. Swap long or Latinate words for shorter plain ones when the meaning holds. Cut hedges and filler that only padded the page. Repeat key nouns instead of leaning on pronouns across a pause. Add deliberate repetition of the main point: on paper repetition looks lazy, in audio it is how the point survives.
6. Replace written signposts with spoken ones. Headings, bold type, "as noted above", figure calls, and footnote markers are silent. Convert them into a spoken preview ("three things: cost, timing, and risk"), transitions that name both sides ("that covers cost. Timing is where it gets tight"), internal summaries after dense sections, and a clear closing cue. Keep signposts short and vary them so the talk does not sound like a list being read out.
7. Write numbers, dates, and abbreviations as spoken. Spell out any number whose spoken form is not obvious, round long figures unless the exact value is the point, and keep one number per sentence. Write abbreviations in full. Hyphenate initialisms that are read letter by letter, like F-B-I. Add a phonetic note for acronyms read as words and for names or terms the speaker may trip on.
8. Mark delivery in the text. Use a small, consistent markup set and state it at the top of the script: bold for stressed words, [PAUSE] for a full beat after a key point, [BEAT] for a short breath break, and line breaks where breath groups split in short-form video scripts. Mark less than you want to. A page covered in markup gets ignored.
9. Apply the rules for the source type. Read references/rewrite-rules-by-text-type.md and apply the section that matches the source: report, essay or article, documentation, academic text, marketing copy, or fiction. Each type fails aloud in different places, and the fixes differ.
10. Time it for real and fix the gap in the text. Read the adapted script aloud with a timer, at delivery pace and volume. Compare against the target. Over target: cut whole ideas first, then sentences, then words. Never fix an overrun by planning to speak faster. Under target by more than about 10 percent: add an example, a recap, or a pause, not padding. Re-run bin/adapt_check.py on the final text and record both the estimate and the timed read.
11. Check the worked examples when stuck. references/worked-examples.md shows a report paragraph turned into a talk opening and web copy turned into a video script, with each change explained and timed.

Output Contract

When asked to adapt text for spoken delivery, deliver:

• The adapted script, with the markup set stated at the top and locked lines left exact
• A timing report: word count, the rate used, the checker estimate, the timed read aloud result if a read was done, and the gap against the target runtime
• A short change list: what was cut, what was reordered, which written forms were converted, and any place where meaning had to be simplified to survive being heard once
• Anything that could not be adapted without changing meaning, flagged for the author to decide, instead of silently rewritten

Operating Rules

1. Accuracy rule (non-negotiable). Every factual claim must be verified via web search or a primary source. Never invent numbers, dates, quotes, or events. Unverifiable = cut it or frame as opinion.
2. Meaning outranks elegance. If the spoken version would change a fact, a commitment, or a legal or quoted line, keep the wording and fix only pacing and markup, then flag it.
3. The read aloud test is the gate. No adaptation is done until someone has read it aloud, or the speaker's timed sample has been checked. A script that only reads well silently is not adapted.
4. Never solve a timing problem with speed. Cutting text protects comprehension. Rushing spends it.
5. Numbers a listener must act on (prices, dates, phone numbers, doses) get repeated once and written exactly as spoken. Never leave the spoken form to the reader to guess.
6. Style: no em dashes anywhere in skill output. Plain, direct language.
7. If you write or modify Python CLIs, compile with python3 -m py_compile before reporting done.