docs/demo-script.md

# The Compliment Forest Demo Script

Target length: 85-95 seconds.

Use a pre-generated completed forest for the final reveal. Record one real
generation separately, then edit through the waiting periods so the video shows
the complete product without pretending inference is instantaneous.

## 0:00-0:08 - Hook

**Screen:** Open on a completed Layered Paper Cut forest. Slowly scroll past two
different clearings.

**Voiceover:**

> Most AI encouragement sounds kind, but says very little. The Compliment
> Forest turns one real worry into a grounded, illustrated path forward.

## 0:08-0:22 - Start With a Real Worry

**Screen:** Return to the start. Enter:

- Name: `Ka`
- Situation: `I worry that one test score means I am not as smart as I thought.`

Submit and show two adaptive questions with different selected answers.

**Voiceover:**

> I begin with one sentence. The forest asks five adaptive questions about
> what happened, what feels at stake, and what useful progress would look like.
> It does not ask me to choose a generic mood.

## 0:22-0:31 - Choose the Visual Language

**Screen:** Move quickly through the remaining answers. Select **Layered Paper
Cut** and click **Grow my forest**.

**Voiceover:**

> Then I choose one of four LoRA-trained visual styles. Text and images run on
> separate Modal GPU services, while the Hugging Face Space streams progress.

## 0:31-0:58 - Show the Five-Chapter Arc

**Screen:** Cut to the completed forest. Highlight each role title as the page
scrolls:

1. arrive
2. steady
3. widen
4. step
5. carry

Pause longest on `step`. Show a sentence containing a concrete action such as
reviewing one missed question or identifying one topic to practice.

**Voiceover:**

> MiniCPM plans from exact phrases in my input, writes five chapters, and
> critiques them. The path first acknowledges the feeling, then separates facts
> from fear, offers realistic options, gives one small action, and ends with a
> simple plan.

## 0:58-1:12 - Explain What Makes It Reliable

**Screen:** Overlay a simple diagram:

`adaptive intake -> planner -> author -> critic -> validators -> FLUX images`

Then briefly show a test terminal with `155 passed`.

**Voiceover:**

> Local validators reject invented biography, repeated sentences, unsupported
> dates, vague fallback language, and advice with no practical step. Bad
> chapters are repaired selectively. If repair fails, the app returns an honest
> retry instead of canned prose.

## 1:12-1:24 - Show the Small-Model Work

**Screen:** Show the Hugging Face model, dataset, LoRA, and trace cards in a
quick four-panel montage.

**Voiceover:**

> The project also publishes a 1.08-billion-parameter MiniCPM fine-tune and
> llama.cpp GGUF, four FLUX style adapters, training data, and sanitized
> planner-author-critic traces.

## 1:24-1:32 - Close

**Screen:** Return to the full forest and its final mantra.

**Voiceover:**

> The Compliment Forest is whimsical encouragement with engineering boundaries:
> small models, honest uncertainty, and one useful step back into the day.

**End card:**

`huggingface.co/spaces/build-small-hackathon/compliment-forest`

## Recording Checklist

- Record at 1440p or 1080p, 16:9.
- Keep browser zoom near 90% so the custom interface remains readable.
- Hide bookmarks, personal tabs, tokens, and terminal paths.
- Use captions for every voiceover line.
- Do not show a crisis phrase in the main demo; mention the safety boundary in
  the article and README so the product story stays focused.
- Export a thumbnail from the completed paper-cut forest.