| name | audio-native-article |
| description | Transform articles, blog posts, documentation pages, newsletters, or essays into audio-native TTS packages instead of literal transcripts. Use when preparing narration scripts, section timing metadata, image/table/chart/code descriptions, listening-first summaries, voice-test excerpts, or audio player data for written content with headings, images, diagrams, tables, charts, screenshots, code blocks, or other non-textual material. |
Audio-Native Article
Create listener-first narration that preserves the article's meaning while adapting the page for audio. Do not read the page literally when a listener needs context, transitions, or a spoken explanation of visual material.
Workflow
-
Inventory the article.
- Read the source content and frontmatter/metadata.
- Capture title, author/byline, date, dek/description, headings, links, images, figures, tables, charts, code blocks, callouts, and embeds.
- If working from Markdown or HTML, you may run
scripts/extract_article_context.mjs <article-file> for a first-pass JSON inventory.
-
Decide what the listener needs.
- Keep claims, facts, dates, numbers, names, and attribution faithful.
- Convert visual or structural material into concise spoken context.
- Preserve the article's point of view, but remove web-only friction such as raw URLs, redundant nav text, decorative alt text, and repeated captions.
-
Write the audio-native script.
- Start with title, byline when useful, publication date when useful, and a one-sentence orientation.
- Use headings as spoken section transitions, not as flat labels.
- Turn tables/charts into the takeaway, then mention the strongest supporting values.
- Turn screenshots/images into what they prove or illustrate.
- Turn code into intent and short commands only when the exact syntax matters.
- Add brief transitions when the page jumps visually but audio needs continuity.
-
Create playback metadata.
- Produce
sections.json with heading labels in source order for current-section UI and hover previews.
- If generating audio, produce or request waveform peaks as normalized
0..1 values.
- Keep visual descriptions in a separate notes file if they are too detailed for the final script.
-
Validate before TTS.
- Read the script aloud mentally for pacing and listener fatigue.
- Check that every image/table/chart/code block was either narrated, summarized, or intentionally skipped as decorative.
- Compare against the source for factual drift.
- Estimate duration at 140-165 spoken words per minute.
Output Shape
Prefer this package when the user asks to prepare a full audio version:
audio/
script.txt
sections.json
visual-notes.md
manifest.json
script.txt is what goes to TTS. visual-notes.md records why visual material was included, condensed, or skipped. manifest.json tracks source path/URL, voice/provider settings if known, word count, estimated duration, and generated audio filenames.
Non-Text Material
Use references/audio-native-writing.md when the article contains images, charts, tables, screenshots, code, formulas, or diagrams.
Do not say "image shows" repeatedly. Vary the phrasing and connect visuals to the argument:
- "The screenshot makes the failure mode concrete..."
- "The table's important pattern is..."
- "The chart is not about the exact curve; it shows..."
- "The command example is short enough to read..."
Guardrails
- Do not invent facts hidden outside the source.
- Do not preserve source markdown or HTML syntax unless syntax is the topic.
- Do not read full URLs unless the URL itself matters.
- Do not include secrets, private keys, local-only paths, or provider API keys in outputs.
- Do not optimize for a TTS vendor's branding or embed widget; keep the narration and metadata provider-neutral.