mkt-transcript-to-hooks-script

name

mkt-transcript-to-hooks-script

description

Transform a source video transcript into a full Vietnamese short-video script — 4 distinct hook variations (Bold Statement, Data/Shock, Counter-intuition, Myth Busting) plus a Before-After / Three Acts / Action body with inline [REF] markers. Output is structured JSON ready for review + Notion push. Sub-skill invoked by the `mkt-video-url-to-script-notion` orchestrator (and its `mkt-script-hook-writer` sub-agent). USE WHEN user says 'transcript to 4 hooks', 'viết 4 hook từ transcript', 'transcript to script với 4 hook', or a parent workflow hands you a transcript and needs a Vietnamese short-video script.

Transcript → 4 Hooks + Script

Single-purpose sub-skill: distill one video's transcript into a finished Vietnamese short-video script with 4 hook variations the editor can A/B test. Does NOT download videos, does NOT push Notion.

---

Input

Either inline arguments in the invoking prompt, or a JSON blob:

{
  "transcript_text": "full plaintext transcript of source video…",
  "source_title": "original uploader's title",
  "source_url": "https://www.youtube.com/shorts/XYZ",
  "duration_sec": 80,
  "angle": "optional: what angle the user wants",
  "avoid": "optional: anything not to mention (e.g. 'don't say day 21')"
}

---

Workflow

Step 1 — Distill the core idea

Read the transcript. Identify:

• 1-line topic
• Main claim / payoff
• Key supporting points (bullet list, max 6)
• Content type: listicle / demo / personal story / news / comparison
• Best structure (see references/v1-structures.md):
  • Before-After (default, most versatile)
  • Three Acts (if personal story with conflict)
  • Action (if demo/reaction/tension)

Step 2 — Draft Last Dab + CTA first

Per brand voice: write the closing punchline FIRST, then hook, then middle.

Default CTA (use verbatim unless user overrides):

Comment "Agent" mình gửi bạn link nhóm học Agents miễn phí nhé.

Step 3 — Write 4 hooks — one per type

Use references/hook-patterns.md. Each hook is 1–2 short spoken sentences, max 3 seconds aloud:

• Hook A — Bold Statement / Comparison Shock — compare brands, declare a paradigm shift, drop a strong claim.
• Hook B — Data / Specific Numbers — lead with concrete figures (6 update, 60 giây, 90+ công cụ, v.v.).
• Hook C — Counter-intuition / Personal Contrarian — "mình vừa làm X, nghe điên nhưng…", unusual personal move.
• Hook D — Myth Busting / Behavior Disruption — call out what the audience is currently doing wrong; invert a common assumption.

Every hook must:

• Use mình (never tui, never tôi), address bạn / các bạn.
• Put the concrete payoff in the first sentence — no warm-up.
• Contain at least one specific detail (number, tool name, action) — no generic hooks.
• Respect the avoid list if the caller supplied one.

Step 4 — Write the body in chosen structure

Follow the template in v1-structures.md. Body = TRƯỚC → CẦU NỐI → SAU → LAST DAB + CTA. Keep:

• Short spoken sentences, conversational.
• Brand names allowed (Claude, ChatGPT, OpenAI, Anthropic, Cursor, Slack, Notion, Jira, Google Docs, Figma, Midjourney, TikTok, Reels, etc.). No other English jargon.
• One [REF: url] marker after the line that visually matches each reference URL the caller gave. If only the source URL is available, place [REF: source_url] right after the hook.

Step 5 — Write the editor notes

One short block with:

• Aspect ratio
• Visual recommendations per hook (1 line each)
• Demo / screen recording guidance
• Nhạc (BPM)

Step 6 — Emit JSON output

Exactly this shape (for the parent orchestrator to collect):

{
  "title": "Script title in Vietnamese",
  "structure": "Before-After|Three Acts|Action",
  "duration_sec": 75,
  "hooks": {
    "A": "Hook A text…",
    "B": "Hook B text…",
    "C": "Hook C text…",
    "D": "Hook D text…"
  },
  "body": "Full body text with inline [REF: url] markers — TRƯỚC → CẦU NỐI → SAU → LAST DAB + CTA",
  "editor_notes": "Short editor notes…",
  "references": [
    {"url": "https://…", "note": "nguồn gốc / cảnh nào"}
  ],
  "source_url": "https://…",
  "source_title": "original title"
}

Also produce a human-readable markdown preview (title, 4 hooks, body, editor notes) so the user review step is easy to read. The JSON is the machine output; the markdown is for the human.

---

Hard Rules — enforced before returning

• First person = mình (never tui, never tôi).
• 4 distinct hooks, one per type, all reference a concrete detail.
• Body ends with the CTA (or caller-overridden CTA).
• Every caller-supplied reference URL appears at least once as [REF: …] in body.
• No English jargon outside allowed brand names.
• Respect avoid list (check final text).
• Target word count = duration_sec × 2.8 (Vietnamese speaking pace ~170 wpm).

If any rule fails, fix in place before emitting output.

---

Files

• references/hook-patterns.md — 4 hook types with templates + criteria
• references/v1-structures.md — Before-After / Three Acts / Action templates