Content Engine

Content Engine — VPC × Storytelling Pipeline

Commercial pipeline để biến hồ sơ khách hàng (Value Proposition Canvas) thành post IG/Reels chất lượng. 4 stage độc lập — buyer có thể chạy full pipeline hoặc dừng/edit/rerun từng stage.

When to invoke

VN trigger phrases:

• "lập hồ sơ khách hàng + viết content"
• "content engine"
• "build factory content"
• "viết content theo VPC"
• "tạo content từ insight khách hàng"

User can also invoke /content-engine directly.

Do NOT use for:

• Long-form blog/SEO articles (use other skill)
• Lead magnet ebooks (use lead-magnet-factory)
• Pure customer research without content output (use fb-insight-miner standalone)
• Non-Vietnamese content (skill is VN-only in v1)

Required inputs (ask user if missing)

Ask in 1 consolidated question. Required:

1. Persona/niche — e.g., "beauty creator chuyển từ affiliate sang course"
2. Data source — chọn 1:
   • fanpage_url (FB fanpage public để scrape qua fb-insight-miner)
   • brief (điền form templates/brief-form.md)
   • hybrid (cả 2)
3. Content goal trọng tâm — awareness / lead / sale (có thể đa mục tiêu, default lead)
4. Output format — ig_caption / reels_30s / reels_60s / fb_post / multi
5. Số bài cần generate — default 12 (mix formula)
6. CTA chính — sản phẩm/offer skill sẽ điều hướng tới (ví dụ: "khoá Content POV 4 tuần — link bio")
7. Brand voice notes (optional) — tone, từ cấm, signature phrases
8. Quality mode — economy (default) hoặc pro:
   • economy: Sonnet 4.6 mọi stage, không có editorial pass. ~$0.80/run. Đủ chất lượng cho daily content.
   • pro: Sonnet 4.6 stage 1-3, Opus 4.7 stage 4 + editorial pass. ~$2.50/run. Dùng cho launch/sale post quan trọng.

4-stage workflow

Use TodoWrite to track each stage visibly.

Stage 1 — Customer Profile (VPC)

Goal: sản xuất output/{persona_slug}/profile.json đạt schema schemas/profile.schema.json.

Read first: references/vpc-framework.md.

Branching by data_source:

• fanpage_url → invoke fb-insight-miner skill on the URL. After it produces an insight report, map:
  • Pain points → pains[]
  • Voice-of-customer phrases → voice.verbatim_bank + voice.repeated_words
  • Persona summary → demographics
  • Aspirations from comments → gains[]
  • "What they're trying to do" → jobs[]
• brief → load templates/brief-form.md, present to user, parse answers into the same fields.
• hybrid → run both; cross-reference; flag any field where brief contradicts scraped data.

Quality gates (from references/vpc-framework.md Section 3):

• jobs ≥ 3, pains ≥ 5, gains ≥ 5, verbatim_bank ≥ 5
• evidence_summary.observed_pct ≥ 60 → set ready_for_hook_stage = true
• If gates fail → tell user explicitly, suggest re-running fb-insight-miner with deeper depth, OR ask user to fill brief gaps.

Niche pack assist: if user's niche matches a file in niche-packs/ (e.g., beauty-creator.json), seed the profile with priors then verify each prior against actual data — never silently ship priors as observed.

Write output/{persona_slug}/profile.json and validate with scripts/validate.py profile.

Stage 2 — Insight Mining

Goal: sản xuất output/{persona_slug}/insights.json (10-25 insight) đạt schema schemas/insights.schema.json.

Read first: references/vpc-framework.md Section 2.1 (field-to-asset mapping) + references/storytelling-and-hooks.md Section 3 (mapping tables).

Procedure:

1. Load profile.json into context.
2. For each of 8 insight types (tension / contradiction / say_vs_do / surprise_myth_bust / aspiration / objection / transformation / identity_shift), generate 1-3 candidate insights by combining pains/gains/voice. Insight = reframing, not raw pain.
3. Score each by novelty_score (1-5). Drop anything ≤ 2 unless it's the only insight in that category.
4. For each insight: cite ≥1 verbatim quote from profile.voice.verbatim_bank.
5. Auto-suggest recommended_formulas using Table A from storytelling-and-hooks.md.
6. Tag applicable_goals.

Write to output/{persona_slug}/insights.json and validate.

Stage 3 — Hook Library

Goal: sản xuất output/{persona_slug}/hooks.json (50-100 hooks) đạt schema schemas/hooks.schema.json.

Read first: references/storytelling-and-hooks.md Section 2 (8 hook frameworks).

🆕 Hook bank integration (added v1.1):

• Trước khi generate, check niche của user → nếu match một bank trong ~/.claude/skills/hook-swipe-library/banks/, load hooks từ bank làm seed:
  • women-biz-vn niche → load women-biz-vn-72.json (72 hook proven theo Kallaway 18-topic × 4-temperature)
  • Future banks: thêm vào hook-swipe-library/banks/ khi có
• Tỷ lệ trộn: 30-50% hook từ bank (đã proven, chỉ cần remix theo voice persona) + 50-70% generated mới (theo profile.voice cụ thể)
• Lý do: bank = baseline an toàn; generated = customize sâu cho persona cụ thể.

🆕 Methodology integration (added v1.1): Khi generate hook mới, follow ~/.claude/skills/kallaway-hook-master/ framework:

• 6 Hook Frameworks (references/6-hook-frameworks.md): Fortune Teller / Experimenter / Teacher / Magician / Investigator / Contrarian
• 5 Dream Outcome Formats (references/5-dream-outcome-formats.md): About Me / If I / To You / Can You / He-She Just Did
• 4-Component Alignment (references/4-component-alignment.md): Spoken + Text + Visual + Audio cho format video

→ Mỗi hook output có thêm field kallaway_framework (1 trong 11 framework) và temperature (lanh/am/nong/banhang) để A/B test có hệ thống.

Procedure:

1. Load profile.json + insights.json.
2. (NEW) Check hook-swipe-library/banks/ for matching niche → load seed hooks.
3. For each insight, generate 3-6 hook variants spanning ≥3 different hook types (curiosity_gap / contradiction / stat_led / question_led / story_open / callout / list / transformation) HOẶC dùng 1 trong 11 Kallaway frameworks (recommend cho video format).
4. Each hook MUST:
   • Use ≥1 word from profile.voice.repeated_words (record in uses_voice_words)
   • Be ≤25 words
   • Reference its paired_insight_id
   • Have a paired_formula (use Table A mapping)
   • Have a goal
   • (NEW for video format) Have kallaway_framework + temperature tags
5. Spread goal distribution across awareness/lead/sale roughly matching user's stated content_goal.
6. (NEW) Spread temperature distribution: ~30% lanh, ~30% am, ~30% nong, ~10% banhang (adjust theo content_goal).

Write + validate.

Stage 4 — Content Generation

Goal: sản xuất output/{persona_slug}/content.json (N posts) đạt schema schemas/content.schema.json.

Read first: references/storytelling-and-hooks.md Section 1 (formula specs + word counts).

Procedure:

1. Load profile.json + insights.json + hooks.json.
2. Pick N hooks balanced across formulas + goals (default 12 posts: 3 awareness / 5 lead / 4 sale; varied formulas).
3. For each pick, fill the formula's beat slots using:
   • Hook → from hooks.json entry
   • Body → from paired insight's statement_vi, hidden_cost, reframe_angle
   • Solution/Proof → from profile.gains (required/expected) + user's cta.offer
   • Voice → use customer vocabulary from profile.voice
4. Respect word count from formula spec + format (ig_caption / reels_30s / reels_60s / fb_post).
5. Generate hashtags: 3 broad + 5 niche + 2 branded (do not exceed 12).
6. Validate via scripts/validate.py content — every post must pass:
   • word count in range
   • uses_voice_vocabulary = true
   • no fabricated stats
   • no fabricated testimonials (use [testimonial_X_placeholder] if needed)
   • hook within first line
   • cta present
7. Write per-post markdown files to output/{persona_slug}/posts/{post_id}_{formula}.md for human review.

Directory layout

content-engine/
├── SKILL.md                          # this file
├── README.md                         # buyer-facing intro
├── schemas/
│   ├── profile.schema.json
│   ├── insights.schema.json
│   ├── hooks.schema.json
│   └── content.schema.json
├── references/
│   ├── vpc-framework.md              # Stage 1 reference
│   └── storytelling-and-hooks.md     # Stages 3+4 reference
├── templates/                        # (build after foundation approved)
│   ├── prompt-stage1-profile.md
│   ├── prompt-stage2-insight.md
│   ├── prompt-stage3-hook.md
│   ├── prompt-stage4-content.md
│   └── brief-form.md
├── niche-packs/                      # (build after foundation approved)
│   ├── _universal.json
│   ├── beauty-creator.json
│   ├── coach-online.json
│   └── ...                           # 10 niches total
├── scripts/
│   ├── validate.py                   # JSON schema validator
│   └── render_post.py                # content.json → markdown post files
├── examples/
│   └── linh-beauty-28/               # full sample run for buyers
│       ├── profile.json
│       ├── insights.json
│       ├── hooks.json
│       └── posts/
└── output/                           # per-run artifacts
    └── {persona_slug}/
        ├── profile.json
        ├── insights.json
        ├── hooks.json
        ├── content.json
        └── posts/

Commercial reliability rules

1. Never fabricate testimonials. Use [testimonial_N_placeholder] placeholders.
2. Never hallucinate stats. If profile has no observed stat, write qualitatively.
3. Never paraphrase verbatim quotes in profile.voice.verbatim_bank — keep original spelling/typos.
4. Stage gates are hard. If profile.evidence_summary.ready_for_hook_stage = false, do NOT proceed to Stage 2 without explicit user override.
5. Language lock. v1 = Vietnamese only. No code-switching unless brand voice explicitly allows.
6. Source traceability. Every pain/gain/job entry must carry source_ref or be marked confidence: "assumed".
7. CTA consistency. Every Stage 4 post's CTA must align with the user-provided cta.offer from inputs. No drift.

Cost transparency

• fb-insight-miner data path — quote user the fb-insight-miner cost (depth-dependent: $2–$6).
• Brief-only path — ~free (just AI tokens).
• Per-stage AI tokens by quality_mode:

Cost-cap (hard stop): if a single run exceeds $5 in token cost, stop and ask user to confirm before continuing. Show running total after each stage.

Failure modes

• Profile thin (gates fail) → Stop Stage 1. Recommend deeper data collection. Don't fake the gates.
• Insights all generic (novelty_score ≤ 2 across the board) → Likely Stage 1 voice/vocab is shallow. Reload profile, focus on voice.verbatim_bank.
• Hooks all sound the same → Pin hook type distribution explicitly; force ≥3 different types per insight.
• Posts use generic words instead of customer voice → Validate fail on uses_voice_vocabulary. Re-generate with stricter prompt citing profile.voice block.
• CTA drift in posts → Validate fail on cta_present + voice. Inject CTA from user input at template level, not by hoping AI remembers.
• Buyer wants English output → v1 doesn't support; tell user explicitly.

Output language

Vietnamese only in v1. All examples, prompts, and posts in VN.

Files to load at each stage

Status

v1.0 — foundation + templates + scripts + niche pack (coach personal brand) + 1 full example shipped.