@lalalic/markcut 2.2.8 → 2.3.0

package/skills/markcut/docs/templates/courseware/prompts/outline.md

@@ -1,58 +0,0 @@
-# Prompt: Course Outline
-
-> Fill every `{placeholder}`, then execute in your own (orchestrator) context.
-> Output feeds `prompts/scene.md` and must be confirmed by the user before scene writing.
-
----
-
-You are an instructional designer planning a {duration_min}-minute video course.
-
-Topic / source material:
-
-{material}
-
-Audience: {audience}
-Language(s): {languages}
-
-Requirements:
-
-- Structure: 1 hook + 1 title slide + 3–6 concept sections + 1 summary.
-- Total narration budget ≈ {duration_min} × 150 words. Allocate a word budget per section; the sum must fit the budget.
-- Per concept section provide:
-  - title (≤ 5 words)
-  - 3–6 key points, each as `**bold key phrase** — short elaboration`
-  - one concrete example (real product, scenario, or story) — gets its own key point
-  - whether the content is comparative → table (if yes, name the columns)
-  - whether a mermaid diagram would help (flowchart, architecture, timeline)
-  - per-bullet **image/video suggestion** (a concrete visual that illustrates this point)
-  - word budget
-- Summary: one comparison table spanning all concepts (name the columns) + a selection/decision guide ("If X → use Y").
-- Hook: one vivid visual idea usable as an image/video generation prompt, OR a provocative question if generation is unavailable.
-- Stay strictly within the given material plus well-established knowledge. No invented facts, numbers, or citations.
-
-Output format (markdown):
-
-```
-## Hook
-- visual: <generation prompt or question>
-
-## Concept 1: <title>
-- points:
-  - **<key phrase>** — <elaboration>
-  - ... (last point = the concrete example)
-- images:
-  - bullet 1: <description of visual — screenshot, diagram, photo, icon>
-  - bullet 2: <description>
-  - ...
-- table: no | yes (<col1>, <col2>, ...)
-- mermaid: no | yes (<type and description of diagram>)
-- words: <n>
-
-## Concept 2: ...
-
-## Summary
-- table columns: <dimension>, <per-concept columns...>
-- guide:
-  - <condition> → <concept>
-- words: <n>
-```

package/skills/markcut/docs/templates/courseware/prompts/scene.md

@@ -1,70 +0,0 @@
-# Prompt: Concept Scene (slide + script)
-
-> Fill every `{placeholder}`, then execute in your own (orchestrator) context.
-> Run once per outline section (TitleSlide, each Concept, Summary).
-> Constraints below mirror TEMPLATE.md §2/§3 — the reviewer checks against them.
-
----
-
-You are writing one scene of a {duration_min}-minute course for a {audience} audience.
-
-Outline section:
-
-{outline_section}
-
-Course context (for continuity — do not repeat content already covered):
-
-{outline_full}
-
-Produce, for language(s) {languages}:
-
-**A. Slide source** (markdown, rendered by `<Slide>`):
-
-- `##` heading = section title
-- ≤ 6 bullets, shape: `<emoji> **bold key phrase** — short elaboration`
-- consistent emoji set across the course: {emoji_set}
-- markdown table if the outline marks this section comparative
-- `>` quote block only for a single memorable takeaway line
-- no bullet may exceed one line at 1.3em / 1920px width (~90 chars)
-- Optional mermaid diagram: add a fenced mermaid code block after the bullet list (` ```mermaid ...``` `) for flowcharts or architecture visuals.
-- **For each bullet, suggest a visual** (image, diagram, screenshot, or icon) that could accompany it. The visual should show something concrete: a screenshot, a diagram step, a photo, etc.
-
-**B. Narration script** (per language, **one paragraph per bullet**):
-
-- **One paragraph per bullet** — never merge bullets into a single script. Each paragraph becomes a separate `- script "..."` node with `on:(start, <sceneId>.current=N)`.
-- Paragraph length ≈ **15–30 words** per beat (~6–12s at 2.5 w/s).
-- Each paragraph **expands** its bullet — never reads it verbatim.
-- Spoken register: short sentences, no markdown, no stage directions.
-- The concrete example gets its own bullet and its own paragraph. Place it last (final beat of the scene).
-- For TitleSlide: 1 paragraph (just the welcome/introduction line).
-- For Summary: 1 paragraph per row in the comparison table + 1 final closing paragraph ("Thank you for watching…").
-
-Output format:
-
-```
-### <SceneName (PascalCase, no spaces)>
-
-~~~md source
-<slide markdown>
-~~~
-<one extra fenced block per additional language: ~~~md <lang>-source>
-
-script (en):
-<p1: describes bullet 1>
-
-<p2: describes bullet 2>
-
-<p3: describes bullet 3 — concrete example>
-
-image suggestions:
-<bullet 1: brief description of visual — e.g. GitHub commit diff screenshot>
-<bullet 2: brief description of visual — e.g. unique hash diagram>
-<bullet 3: brief description of visual — e.g. PR timeline graphs>
-
-script (<lang>):
-<p1 translation>
-
-<p2 translation>
-
-<p3 translation>
-```

package/skills/markcut/docs/templates/deep-dive/TEMPLATE.md

@@ -1,354 +0,0 @@
----
-name: deep-dive
-description: 深度解读 — turn a complex topic, problem, or phenomenon into a thorough video-essay style analysis with evidence, multiple perspectives, citations, and synthesis.
-when-to-use: video essays, investigative deep dives, philosophical explorations, science explainers, cultural criticism, political analysis — 3-20 minute long-form analytical content
-engine: "@lalalic/markcut — run via `npx @lalalic/markcut`"
----
-
-# Deep Dive (深度解读) Template
-
-Follow this file top to bottom. Read the markcut skill (`SKILL.md` → `docs/markdown-descriptive.md`) first if you have not.
-
-```mermaid
-flowchart LR
-    Topic[📥 Topic / Problem / Phenomenon] --> Research[🔍 researcher agent<br/>evidence + sources]
-    Research --> Outline[📝 outline.md<br/>narrative arc]
-    Outline --> Script[✍️ script.md<br/>scene-by-scene]
-    Script --> Storyboard[📄 deep-dive.md]
-    BGM["🎵 BGM (optional)"] --> Storyboard
-    Media["🖼️ src:auto / archival / citations"] --> Storyboard
-    Storyboard --> Render[🎬 npx markcut render]
-    Render --> MP4[📺 final.mp4]
-    MP4 --> Reviewer[🔍 reviewer agent]
-    Reviewer -->|FAIL| Fix[🔧 fix]
-    Fix --> Storyboard
-    Reviewer -->|PASS| Done[✅ Done]
-```
-
-| Path | Runs in | Purpose |
-|---|---|---|
-| `TEMPLATE.md` | your context | everything |
-| `prompts/*.md` | your context | fill-in prompts you execute |
-| `agents/*.md` | separate session | subagent definitions |
-
-## 0. Prerequisites
-
-- `npx @lalalic/markcut` runnable
-- `ffmpeg`/`ffprobe/exiftool` on PATH
-- Web search capability (for research phase)
-- ITT/TTV CLI (for `src:auto` visuals) — optional but recommended
-- For reviewer: image-understanding capability, STT CLI
-
-## 1. Inputs — collect before starting
-
-| Input | Required | Default | Notes |
-|---|---|---|---|
-| Topic / problem / phenomenon | **yes** | — | what to analyze. Can be a question ("Why do democracies fail?"), a phenomenon ("The rise of remote work"), a problem ("Plastic pollution"), or a cultural object ("Oppenheimer") |
-| Source material | no | — | optional: article URL, book title, research paper, documentary to reference |
-| Language | no | en | narration language |
-| Tone | no | `analytical` | `analytical` (neutral, evidence-driven), `opinionated` (strong thesis, persuasive), `philosophical` (open-ended, reflective), `investigative` (detective-like, reveals) |
-| Target duration | no | 8 min | 3–20 min. Longer formats need deeper scene breakdown |
-| Depth | no | `moderate` | `overview` (surface-level, broad), `moderate` (some depth, 3-5 angles), `deep` (thorough, 5-10 angles, academic-level) |
-| Voice | no |  | mlx-audio voice per language |
-| BGM style | no | `ambient` | background music mood. Optional — deep dives can work without BGM |
-
-**Core rule: Every factual claim must be supported by a source.** If you cannot find a source for a claim, either qualify it ("some argue that...", "it's believed that...") or remove it. Cite sources on screen (see §3).
-
-## 2. Scene grammar — narrative architecture
-
-### The deep dive arc
-
-A deep dive is not a lecture and not a highlight reel — it's a **narrative journey** through a topic. The canonical arc:
-
-```
-# video                          ← root: width:1920 height:1080 fps:30 layout:series
-│                                  subtitle:{fontSize:"24px"}
-│                                  transition:fade(0.5)
-├── ## Hook                      ← 15–30s. The question, the contradiction, the mystery
-├── ## Context                   ← 30–60s. Background the audience needs
-├── ## Thesis                    ← 15–30s. Your central argument / framing
-├── ## Angle 1                   ← 1–3 min. First perspective / evidence block
-│   │                              (may contain multiple sub-scenes)
-│   ├── {scene}                  ← Claim → Evidence → Source → Analysis
-│   ├── {scene}                  ← Repeat for each sub-point
-│   └── ...
-├── ## Angle 2                   ← 1–3 min. Second perspective
-├── ## Angle 3                   ← 1–3 min. (optional, depth-dependent)
-├── ## Counterpoint              ← 30–60s. The opposing view (fairly stated)
-├── ## Rebuttal / Synthesis      ← 30–60s. Why the counterpoint doesn't fully hold, or synthesis
-└── ## Conclusion                ← 30–60s. Restate thesis, broader implication, call to reflection
-```
-
-### Scene timing template
-
-| Section | Duration | Purpose |
-|---|---|---|
-| Hook | 15–30s | Arresting question, contradiction, or provocative claim. No context yet. |
-| Context | 30–60s | Background: history, definitions, prior work. What the audience MUST know. |
-| Thesis | 15–30s | Clear statement of your argument/framing. "I argue that..." |
-| Angle×N | 1–3 min each | Evidence blocks. Each angle = one perspective or dimension of the topic |
-| Counterpoint | 30–60s | Fair presentation of the strongest opposing view |
-| Synthesis | 30–60s | Why the evidence points your way, or a nuanced middle ground |
-| Conclusion | 30–60s | Broader implication, call to reflection, closing thought |
-
-### Scene types (within sections)
-
-Each section contains 1–5 scenes. A scene is one of these types:
-
-**Claim scene** — A point is stated, then supported:
-```
-## Main Idea
-layout:parallel
-- image src:auto prompt:"<metaphorical or illustrative visual>" duration:10
-- subtitle src:"<the claim — one clear sentence>" duration:10 type:Fade
-- script "<narration expanding the claim, citing sources>" 
-```
-
-**Evidence scene** — Data, quote, or case study:
-```
-## The Data
-layout:parallel
-- image src:auto prompt:"<visual related to the data>" duration:8
-- subtitle src:"<key statistic or quote>" duration:8 type:Typewriter
-  style:"fontSize:36px;color:#ffd700"
-- script "<narration explaining what this evidence means>"
-```
-
-**Source citation scene** — Attribution overlay:
-```
-## Source
-layout:parallel
-- image src:auto prompt:"<academic or documentary feel>" duration:4
-- subtitle src:"— Smith (2023), Journal of Political Science" duration:4 type:Fade
-  style:"fontSize:20px;color:#aaa;fontStyle:italic"
-```
-
-**Comparison scene** — Two-column or before/after:
-```
-## Two Views
-layout:parallel
-- component isBackground:true jsx:"<SplitComparison left='View A' right='View B' />"
-- script "<narration comparing the two perspectives>"
-```
-
-**Transition scene** — Brief pause, signaling a shift:
-```
-## Shifting Gears
-layout:parallel
-- image src:auto prompt:"<transition visual — doorway, turning page>" duration:3
-- subtitle src:"But there's another way to look at this..." duration:3 type:Fade
-```
-
-### Scene count guide
-
-| Depth | Sections | Total scenes | Duration |
-|---|---|---|---|
-| overview | Hook + Context + Thesis + 2 angles + Counterpoint + Synthesis + Conclusion | 10–18 | 3–5 min |
-| moderate | Hook + Context + Thesis + 3 angles + Counterpoint + Synthesis + Conclusion | 15–28 | 5–10 min |
-| deep | Hook + Context + Thesis + 4–6 angles + Counterpoint + Rebuttal + Synthesis + Conclusion | 25–50 | 10–20 min |
-
-### Visual style for each tone
-
-| Tone | Visual approach | Pacing | Music |
-|---|---|---|---|
-| analytical | clean diagrams, data viz, archival footage, minimal text | steady, 10–30s per scene | ambient, subtle |
-| opinionated | bold text overlays, dramatic imagery, reaction clips | dynamic, 5–15s per scene | building, persuasive |
-| philosophical | nature visuals, slow pans, negative space, poetic imagery | slow, 15–40s per scene | sparse or none |
-| investigative | document-style, maps, photos, timeline graphics, evidence close-ups | varied, 8–25s per scene | tense, mysterious |
-
-## 3. Authoring rules — the analytical bar
-
-### Research quality (see `prompts/outline.md` and `agents/researcher.md`)
-
-Every deep dive is built on evidence. Before writing any scene:
-1. **Research the topic** thoroughly using web search. Find: key facts, statistics, expert quotes, opposing views, historical context.
-2. **For each factual claim**, note the source (URL, book title with page, paper with DOI). If no source exists for a claim, either remove it or soften it ("some suggest...").
-3. **Present counterpoints fairly**: the "steel man" version of opposing views, not a straw man. If you can't steel-man an opposing view, acknowledge the uncertainty.
-4. **Multiple perspectives** are the heart of a deep dive. A monological "this is correct" piece is a lecture, not a deep dive. Always include at least one counterpoint.
-
-### Narration script
-
-- **Conversational but precise**: Write for the ear, but don't sacrifice accuracy. Use everyday language for complex ideas.
-- **Pacing**: 2.5–3 words/second. A 30s scene ≈ 75–90 words.
-- **Scene-level hook**: Every scene should have a mini-hook in its first sentence — why should the viewer keep watching through this scene?
-- **Transitions**: End each scene/section with a hook into the next ("But that's only half the story...").
-- **No filler**: No "as we can see," "it's important to note that," "in order to understand X we must first..." — cut these.
-- **Citations in narration**: "According to a 2023 study in Nature..." not "A study said..."
-
-### On-screen citations
-
-Every time you cite a source in narration, add a brief citation scene or overlay:
-```markdown
-## Source
-layout:parallel
-- image src:auto prompt:"<related visual>" duration:4
-- subtitle src:"— Smith (2023), Journal of Political Science" duration:4 type:Fade
-  style:"fontSize:20px;color:#888;fontStyle:italic"
-```
-
-For dense deep dives, use root-level citation slides between sections rather than per-claim — but the narration must still name the source.
-
-### Subtitle / text overlay rules
-
-- **Font size**: 28–48px for main text (smaller than short-video — deep dive viewers read more).
-- **Key claims**: 36–48px, bold, centered.
-- **Citations**: 18–24px, italic, bottom-right or bottom-center, subtle color.
-- **Data / statistics**: 40–56px, prominent, colored (`#ffd700` for emphasis).
-- **Quotes**: 32–40px, with quotation marks, citation below.
-
-### Visual rhythm
-
-- **Variety is critical**: Don't use the same visual pattern for every scene. Mix: full-screen imagery, split screens, text-only, data visualization, b-roll, talking head.
-- **B-roll relevance**: Every visual should serve the argument. Avoid generic "filler" stock footage. If a concept is abstract, use metaphorical or illustrative visuals that provoke thought.
-- **Scene length variety**: Within a section, mix short scenes (3–5s for impact) with longer ones (15–30s for explanation). A constant rhythm is monotonous.
-
-### BGM
-
-Optional but recommended. Use `foreground:true` for ducking:
-```markdown
-- audio isBackground:true foreground:true src:bgm.mp3 volume:0.08
-```
-
-Volume lower than other formats (0.05–0.10) — deep dive TTS clarity is paramount. For philosophical or heavy topics, consider no BGM at all (silence can be powerful).
-
-### Multi-language (variants)
-
-- Same `# <lang>` variant block pattern as courseware.
-- For bilingual deep dives, ensure both translations match in nuance — translation errors in analytical content are serious.
-
-## 4. Components & styles
-
-### Optional: Comparison / Split-screen component
-
-When comparing two ideas, images, or perspectives:
-
-```jsx
-export function SplitComparison({ left, right, leftLabel = '', rightLabel = '' }) {
-  return (
-    <div style={{
-      display: 'flex', width: '100%', height: '100%',
-      position: 'absolute', inset: 0,
-    }}>
-      <div style={{
-        flex: 1, display: 'flex', alignItems: 'center', justifyContent: 'center',
-        background: 'linear-gradient(135deg, #1a1a2e 0%, #16213e 100%)',
-        fontSize: 36, color: '#e0e0e6', padding: 40, textAlign: 'center',
-        flexDirection: 'column', gap: 12,
-      }}>
-        {left}
-        {leftLabel && <span style={{ fontSize: 18, color: '#888', marginTop: 8 }}>{leftLabel}</span>}
-      </div>
-      <div style={{
-        width: 2, background: 'rgba(255,255,255,.1)',
-      }} />
-      <div style={{
-        flex: 1, display: 'flex', alignItems: 'center', justifyContent: 'center',
-        background: 'linear-gradient(135deg, #16213e 0%, #0f3460 100%)',
-        fontSize: 36, color: '#f5f5f7', padding: 40, textAlign: 'center',
-        flexDirection: 'column', gap: 12,
-      }}>
-        {right}
-        {rightLabel && <span style={{ fontSize: 18, color: '#888', marginTop: 8 }}>{rightLabel}</span>}
-      </div>
-    </div>
-  )
-}
-```
-
-### Stylesheet (optional)
-
-If you want a consistent overlay style for citations and footnotes:
-```css
-~~~css stylesheet
-.citation {
-  font-size: 20px;
-  color: #888;
-  font-style: italic;
-  text-align: right;
-  position: absolute;
-  bottom: 40px;
-  right: 40px;
-}
-.stat {
-  font-size: 52px;
-  color: #ffd700;
-  font-weight: 800;
-  text-align: center;
-}
-.quote-text {
-  font-size: 38px;
-  color: #f5f5f7;
-  font-style: italic;
-  text-align: center;
-  padding: 0 60px;
-  line-height: 1.5;
-}
-~~~
-```
-
-## 5. Workflow
-
-### Phase 0: Research
-
-1. **Understand the topic** — Read the topic/source material. Identify:
-   - Core question or problem
-   - Key dimensions / angles
-   - Known facts vs contested areas
-   - Potential sources and evidence
-
-2. **Research** — Fill `prompts/outline.md` with the topic. This produces:
-   - Research summary with key facts, sources, statistics
-   - Narrative arc: hook angle → angles → counterpoint → synthesis
-   - Evidence map: what supports each point
-
-   **Alternative (deeper)**: Run `agents/researcher.md` in a separate session. It returns a structured research document with sources for every claim.
-
-3. **Review research**: Every factual claim must have a source. If sources are weak, note uncertainty in the narration.
-
-### Phase 1: Outline
-
-4. **Narrative arc** — Using the research, write the scene outline: sections, scene types, key claims, and evidence per scene. Use `prompts/script.md` or write directly.
-
-### Phase 2: Storyboard
-
-5. **Write scenes** — For each section in the outline, write the markcut markdown:
-   - Choose scene types (claim, evidence, source, comparison, transition)
-   - Write narration (conversational but precise, sourced)
-   - Add on-screen citations for every source
-   - Choose visuals (`src:auto` prompts or archival imagery)
-   - Add effects for transitions between scenes
-
-6. **Assemble deep-dive.md** — Compile all sections into a single markcut file following §2 grammar. Root config includes TTS voice and subtitle defaults. Add optional BGM.
-
-### Phase 3: Render & Review
-
-7. **Render** — `npx @lalalic/markcut render deep-dive.md`. On errors: fix and re-render, max 3 per error.
-
-8. **Review (quality gate)** — Run `agents/reviewer.md` in a fresh separate session. Pass: `deep-dive.md`, rendered MP4, `TEMPLATE.md`, target duration, language, tone, depth. Returns `{verdict, findings[]}`.
-
-9. **Fix loop** — On FAIL: fill `prompts/fix.md` with findings, edit, re-render, re-review. Max 3 iterations, then escalate.
-
-## 6. Quality gate — exit criteria
-
-Done only when ALL hold:
-
-- [ ] reviewer verdict = `PASS` (zero blocker/major findings)
-- [ ] total duration within ±15% of target
-- [ ] structure matches §2 arc (hook → context → thesis → angles×N → counterpoint → synthesis → conclusion)
-- [ ] hook is engaging (question, contradiction, or provocative claim — not a bland introduction)
-- [ ] at least one counterpoint presented fairly (steel-manned, not straw-manned)
-- [ ] every factual claim in narration is sourced (spot-check 5 claims: each has a citation scene or in-narration attribution)
-- [ ] no fabricated sources, quotes, statistics, or studies
-- [ ] no blank/black frames; visuals are relevant (not generic filler)
-- [ ] STT transcript matches script lines (≥85% content match)
-- [ ] scene types are varied (not all claim scenes — mix of evidence, citation, transition, comparison)
-- [ ] BGM (if present) ducked under voice, volume ≤0.10
-- [ ] tone matches chosen style (analytical/opinionated/philosophical/investigative)
-- [ ] total scenes and section count appropriate for depth level (see §2 table)
-
-## 7. Reference
-
-- Golden example: See `tests/fixtures/templates/courseware.md` for the template format.
-- For comparison scenes: use the `SplitComparison` component from §4.
-- For citation formatting: follow APA/MLA style adapted for on-screen readability.
-- Effect reference: `fadeIn`, `zoomIn`, `slideInLeft`, `slideInRight` for scene transitions.

package/skills/markcut/docs/templates/deep-dive/agents/researcher.md

@@ -1,69 +0,0 @@
----
-name: deep-dive-researcher
-description: Research a topic thoroughly for a deep-dive video essay. Returns structured evidence with sources for every claim. Runs in a separate session with web search capability.
-context: fresh
-mode: read-only
-tools: web search, read
----
-
-# System prompt
-
-You are a research assistant for a video essay. Your job is to find **evidence** for every dimension of a topic. You are neutral — you present facts and sources, not opinions.
-
-## Task
-
-Given a topic and optional source material, research thoroughly:
-
-1. **Core facts** — what is definitively known? Include sources (URLs) for each.
-2. **Key statistics** — notable numbers with their sources.
-3. **Expert opinions** — notable people and their positions, with attributions.
-4. **Historical timeline** — key events, discoveries, or turning points.
-5. **Debates / controversies** — what are the main disagreements? Who holds each view? What evidence does each side cite?
-6. **Current state** — latest developments, recent research, ongoing debates.
-7. **Broader implications** — why this matters beyond the topic itself.
-
-For every factual claim, provide a source URL. If you cannot find a source, mark it as `[unverified]`.
-
-## Output format
-
-```json
-{
-  "topic": "...",
-  "research_date": "...",
-  "summary": "2-3 sentence overview of findings",
-  "key_facts": [
-    { "claim": "...", "source": "...", "confidence": "high|medium|low" }
-  ],
-  "key_statistics": [
-    { "stat": "...", "source": "...", "context": "..." }
-  ],
-  "expert_quotes": [
-    { "quote": "...", "attribution": "...", "source": "...", "position": "pro|con|neutral" }
-  ],
-  "timeline": [
-    { "date": "...", "event": "...", "significance": "..." }
-  ],
-  "debates": [
-    {
-      "question": "...",
-      "position_a": { "view": "...", "proponents": ["..."], "evidence": ["..."] },
-      "position_b": { "view": "...", "proponents": ["..."], "evidence": ["..."] }
-    }
-  ],
-  "current_state": "...",
-  "implications": ["..."]
-}
-```
-
-# Task template
-
-```
-Research the following topic for a {tone} deep-dive video essay.
-
-Topic: {topic}
-Source material: {source_material}
-Depth: {depth} (overview | moderate | deep)
-Language: {language}
-
-Use web search to find evidence for every dimension listed in your procedure. Return the structured JSON with sources for every claim.
-```

package/skills/markcut/docs/templates/deep-dive/agents/reviewer.md

@@ -1,108 +0,0 @@
----
-name: deep-dive-reviewer
-description: Read-only quality gate for deep-dive video essays. Checks narrative arc completeness, evidence quality, citation presence, visual variety, and factual integrity.
-context: fresh
-mode: read-only
-tools: read file, bash (ffmpeg/ffprobe, STT CLI), image understanding, web search
----
-
-# System prompt
-
-You are a strict, evidence-based QA reviewer for analytical video essays (深度解读). You review; you **never edit** any file. Every finding must cite concrete evidence.
-
-## Inputs (provided in the task)
-- `deep_dive_md` — absolute path to the generated markdown
-- `mp4` — absolute path to the rendered video
-- `template_md` — absolute path to this `TEMPLATE.md`
-- `duration_min` — target duration
-- `language` — expected language
-- `tone` — analytical | opinionated | philosophical | investigative
-- `depth` — overview | moderate | deep
-
-## Procedure
-
-### 1. Static review (source vs rulebook)
-
-**Narrative arc** (from §2):
-- Hook present as first section (not intro, not title card — a real hook).
-- Context section follows (background the audience needs).
-- Thesis or central framing present.
-- At least 2 angle sections for `overview`, 3 for `moderate`, 4+ for `deep`.
-- Counterpoint section present and fairly stated.
-- Synthesis or conclusion section present.
-- Total scene count matches depth level (see §2 table).
-
-**Evidence & citations**:
-- Spot-check 3 scenes with factual claims. Each must have either:
-  - An in-narration source mention ("According to a 2023 study..."), OR
-  - A dedicated source citation scene/subtitle nearby.
-- If any claim lacks attribution → major finding.
-- If any source appears fabricated (check top-level claim plausibility) → blocker.
-
-**Variety**:
-- Are scene types mixed? Not all claim scenes. There should be evidence, citation, comparison, or transition scenes interspersed.
-- Are effects used? At least some scenes have `effect:` (fadeIn, slideInRight, etc.).
-- Visual prompts: are they specific and relevant? Not generic "beautiful landscape".
-
-**Counterpoint fairness**:
-- Read the counterpoint section. Is it a real opposing view, or a straw man? If the counterpoint is obviously weak, flag it.
-
-**BGM/audio**:
-- If BGM present: `foreground:true` and volume ≤ 0.10.
-
-### 2. Dynamic review
-- `ffprobe`: duration, resolution, audio streams.
-- Extract 10 frames evenly spaced. For each:
-  - not blank/black
-  - has visible content matching the scene type (text for claim scenes, imagery for evidence, etc.)
-- Extract audio → transcribe with STT. Compare against `script` lines. ≥85% match.
-- Spot-check 3 `src:auto` prompts rendered correctly (no broken image icons).
-
-### 3. Report
-
-Severity: `blocker` (fabricated sources, wrong aspect, blank frames, no hook) · `major` (missing counterpoint, unsourced claims, low variety, no citations, duration >20% off) · `minor` (font size, timing, emoji usage)
-
-Output exactly this JSON:
-
-```json
-{
-  "verdict": "PASS | FAIL",
-  "measured": {
-    "duration_s": 0,
-    "target_s": 0,
-    "resolution": "",
-    "stt_match_pct": 0,
-    "section_count": 0,
-    "scene_count": 0
-  },
-  "findings": [
-    {
-      "id": "F1",
-      "severity": "blocker | major | minor",
-      "scene": "<section name or 'global'>",
-      "check": "<which rule>",
-      "issue": "<what is wrong>",
-      "evidence": "<source line / frame path / STT excerpt>",
-      "fix_hint": "<one-line suggestion>"
-    }
-  ]
-}
-```
-
-`PASS` requires zero `blocker` and zero `major` findings.
-
-# Task template
-
-```
-Review a deep-dive video essay against its template.
-
-deep_dive_md: {deep_dive_md}
-mp4: {mp4}
-template_md: {template_md}
-duration_min: {duration_min}
-language: {language}
-tone: {tone}
-depth: {depth}
-
-Follow your procedure. Use whisper-cli (whisper.cpp) for STT if python whisper is unavailable. Save frames under a review directory. Spot-check top-level claims for source plausibility if web search is available. Output the verdict JSON only.
-```