Reverse-engineering the editorial grammar of a Filipino documentary series with AI

The Philippines is underrepresented in documentary storytelling. Events that shaped the country in the 1990s and 2000s, the ILOVEYOU virus, the Ormoc flood, the Ampatuan massacre, rarely get film or reenactment coverage outside a handful of institutions. One of those institutions is ABS-CBN's Philippines' Most Shocking Stories (PMSS), a documentary series with a consistent, recognisable editorial voice.

I wanted to understand that voice structurally. Not to clone it, but to test whether AI can analyse an institutional editorial style with enough fidelity to describe its rules, and eventually to draft new work in the same register. This is the research note after ten episodes.

The pipeline

Instead of watching episodes and taking notes, I built a pipeline:

download → scene-change keyframes → transcript
  → AI analysis → structured episode sidecar
    → aggregate style guide

The key decision was scene-change detection instead of fixed-rate frame sampling. In a documentary every cut carries information about pacing, emphasis and visual grammar. Sampling at a fixed frame rate misses dense cut sequences and wastes budget on static shots. I modified my video-watching tool to extract frames at scene boundaries, raised the frame budget to documentary length, and ran ten episodes across two seasons through it: tech history, natural disasters, political crime, true crime and domestic murder.

Each episode produced a structured sidecar. The sidecars were then aggregated into a versioned style guide.

What ten episodes show

Nine hook types

The cold open varies by story category, but the function is always the same: create an information gap that only the episode can close.

1. Historical anchor ("It was 1987 when...")
2. Crime setup
3. Number as opening ("Fifty-eight.")
4. Discovery scene to questions
5. Departure and return
6. False-normalcy shatter (a named, ordinary place, then the break)
7. Miracle-survival reveal
8. False-victim reveal, immediately reversed
9. Discovery in retrograde, backtracking to reconstruct the crime

Rules that held in all ten

These read less like production choices and more like the show's contract with its audience.

• Persistent logo and episode subtitle chrome on every frame
• A fragment-punch sentence rhythm in the narration
• A scale statement with Philippines-first geography ("the worst peacetime maritime disaster in world history")
• A promise close that opens the arc with a declarative guarantee
• A rhetorical question, self-answered, used as a structural joint
• A named reporter sign-off
• A structural spine: one named subject returned to at least twice, confirmed in all nine non-tech episodes

The visual grammar

Nine registers appear across the corpus: archival footage, actor reenactment, AI-generated visualisation where no usable archival exists, data infographics, character cards, verdict graphics, court document zooms, host studio segments and location footage. Season 2 upgrades production with an investigation-board studio and a named sign-off.

The style guide as a living artifact

After ten episodes the analysis stabilised into a style guide at version 0.10 with ten sections: hook formula, sentence rhythm, withhold rules, direct-address inventory, visual grammar, narration register, closing structure, and a table of confirmed versus unconfirmed patterns. Every new episode either confirms a rule or adds a variant. One episode introduced a location-specific false-normalcy shatter, a named commercial venue as the anchor. Another confirmed number-as-opening and added closing types specific to political crime.

Corpus so far: ten episodes, more than 29,900 transcript segments, more than 370 scene-change frames read individually.

What comes next

The corpus MVP is done. The real test is applied:

1. Pick a Philippine event PMSS has not covered.
2. Write a treatment to the format: hook type, structural spine, closing type, visual grammar.
3. Draft narration to the style guide.
4. Compare against an existing episode to check voice fidelity.

The deliverable is a style guide as a data artifact plus one original treatment in the register, as proof that editorial systems can be described precisely enough to be reused.