# Beginner-friendly audio/caption alignment pattern

Use this when a short-form AI-systems video starts feeling too abstract, too expert-coded, or when the visible card text does not match the spoken/captioned text.

## Problem observed

A technically correct preview can still feel wrong when it mixes multiple copy sources:

```text
old approval/timing card text
+ newer final voiceover
+ separately chunked subtitles
```

The viewer sees one idea, hears another, and reads a third. This creates cognitive friction even if the render pipeline is functioning.

A second failure mode is abstraction drift: language like `exception path`, `input boundary`, `output standard`, or `decision structure` is precise for builders but can lose viewers who are still early with AI tools.

## Preferred fix

For public-facing AI-systems shorts, choose one concrete everyday failure example and make it the single source of truth:

```text
FinalVoiceoverScript
→ ScriptDraft / retention beats
→ ShotPlan
→ top card text
→ scene cards
→ subtitle track
→ TTS audio
→ social copy
```

Do not let preview cards, subtitles, and voiceover evolve independently.

## Copy pattern

Prefer beginner-friendly, concrete hooks:

```text
Your AI automation works perfectly in the demo.
Then one weird email arrives.
```

Then explain the system concept through the example:

```text
The name is missing.
The format is different.
The request is unclear.
And suddenly the automation does the wrong thing confidently.
```

Translate abstract system ideas into plain-language splits:

```text
What should go through?
What should be reviewed?
What should stop?
```

Only after the concrete example lands should the deeper principle appear:

```text
If your AI does not know what to do with weird inputs,
it will break the moment real life shows up.
```

## Quality checks

Before sending a preview:

1. Check the first card visually: the top text must match the spoken hook and final subtitles.
2. Check for mixed-language labels in final-near English previews (`Time`/`Motion`, not German `Zeit`/`Bewegung`).
3. Probe actual TTS duration and build subtitle cues against the generated audio length, not only an estimated target.
4. Use explicit duration mode:
   - `preserve_video` for approval timing previews;
   - `trim_to_audio` only when intentionally producing final-near clips.
5. Avoid resending several visually similar Telegram media previews without labels, path, duration, and hash.

## Anti-patterns

Avoid starting broad-audience shorts with terms like:

```text
exception path
input boundary
output standard
action boundary
decision structure
```

These can still exist as internal architecture vocabulary, but public scripts should say what a beginner can picture immediately: weird email, missing name, changed format, unclear request, wrong confident action.