# Demo Runbook

This runbook is designed for a public portfolio demo. It uses synthetic inputs and public-safe outputs only.

## Demo Goal

Show how AI Job Radar turns noisy job signals into structured product decisions:

1. source health
2. job card normalization
3. role relevance
4. location fit
5. capability gaps
6. manual next action

## Demo Inputs

Use:

- `examples/demo-job-alert.txt`
- `examples/job-cards.sample.json`

Do not use real alerts, real resumes, real application trackers, browser sessions, cookies, or screenshots from logged-in pages.

## Five-Minute Walkthrough

### 1. Start With The Problem

AI PM job search is noisy. The workflow is not trying to scrape every source. It is trying to preserve source truth and produce a better decision artifact.

### 2. Show The Synthetic Alert

Open `examples/demo-job-alert.txt` and point out:

- mixed role titles
- different location modes
- AI depth signals
- one blocked source state

### 3. Show The Structured Job Cards

Open `examples/job-cards.sample.json` and point out:

- normalized fields
- source health
- role relevance score
- location fit score
- capability gaps
- manual next action

### 4. Show The Match Report

Open `examples/match-report.sample.md` and explain how the system separates:

- role fit
- location fit
- proof gap
- source blocker
- next action

### 5. Show The Eval Report

Open `examples/eval-report.sample.md` and explain the evaluation targets:

- role relevance accuracy
- location fit accuracy
- blocker classification
- human-review gate preservation

## Demo Script

```text
This project is an AI job discovery workflow for an AI PM transition.
The product choice I want to highlight is that source health is part of the data model.
If a source is login-gated, rate-limited, or blocked, the system reports that honestly.
It does not convert source failure into a false market conclusion.

The workflow then normalizes each signal into a job_card, scores AI-product relevance,
checks location fit, extracts capability gaps, and produces a manual next action.
The final action is intentionally human-gated: the system can recommend tailoring or review,
but it does not auto-apply or auto-message.
```

## Screenshot Checklist

If screenshots are added later, crop or redact:

- browser address bars with tracking parameters
- account avatars
- names, emails, and phone numbers
- notification panels
- private profile details
- real application status

## What Not To Demo

- Live logged-in browsing
- Real candidate profile data
- Real application trackers
- Full resumes
- Private message drafts
- Cookies, tokens, browser profiles, or local machine paths