Docs/Skills and code/skill script-cohort-analysis

.claude/skills/script-cohort-analysis/references/rubric.md

Markdown53 lines3.7 KBSkill: script-cohort-analysis

Rubric v2: definitions

The scored checks are in the run's checklist.md, which is generated from scripts/checklist.js. This file defines the structures you extract before answering them.

Judge only what a transcript can show: words, order and timing. Ignore visuals, editing, delivery and music, and ignore auto-caption errors.

Sections

Split the runtime into contiguous sections, each covering one sub-topic or step. Sections are usually 45s to 3 min long, and they must tile the whole runtime with no gaps. Give each one:

• id: "A", "B", "C", … in order.
• start / end: mm:ss. Each section's end is the next section's start.
• summary: one English sentence saying what the section says, not how good it is.
• function: one of hook, promise, context, value, example, story, complication, payoff, recap, transition, cta, sponsor, filler, repetition. filler = no new information. repetition = restates earlier material without adding anything.
• depends_on: the ids of EARLIER sections whose information this section needs in order to make sense. For example, it uses a number, a character or a claim set up there. Use [] if it stands alone.
• references_unintroduced: things this section mentions as if already known that no earlier part of the transcript introduced. Examples: "as we saw", a character's name used without introduction, "this second model" when no first model came before. Usually [].
• interest: an integer 1–5. How strongly a typical target viewer would want to keep watching during this section. 1 = would skip, 3 = fine, 5 = gripped.

Loops

A loop is a question the script makes the viewer want answered, whether it's asked explicitly or only implied. Record:

• opened_at: when the question becomes live.
• closed_at: when it's answered, or null if it never is.
• size: major if it drives a section or the whole video, minor if it's answered within about a minute.

Promise

The core question or promise of the video, with the timestamp where it first becomes clear.

Topic card

One neutral sentence of 25 words or fewer saying what the video is about: its subject and the question it addresses. Describe subject matter only:

• No judgment of quality, style or format ("gripping", "animated", "well-researched").
• No title-style hype.
• Write it so that two videos on the same subject would get nearly the same card, however well each was made.

Holistic estimate

Imagine the channel made 50 videos. Give the percentile (0–100) you'd expect this script to rank at on script quality alone, with your confidence.

Notes for the v3 checks (2026-10-24)

• H5 and H6 read only the first 60 seconds. For H5, quote the concrete confirmation (a scene, fact or number the title implied) and, separately, the line that promises more than the title did; both must exist. For H6, "direction" means the route or cause. If the opening states the cause outright, or says nothing about where the answer is going, H6 fails.
• C5, K4 and K5 use your section list. A section "ends on" its last one or two sentences. C5 passes for a section only if the question or unresolved consequence is one the next section actually answers. For K4, a join is flat if you could insert "and then" between the sections without changing the meaning, or if it is a list-style signpost ("here is the second thing"); name each one.
• G1 to G4 count figures: a figure is a spoken quantity of money, share, count or duration. Plain dates do not count. For G1, a figure is "framed" if a person, loss, question, prediction or contrast appears in the same or the preceding sentence. For G2, scan the transcript in 30-second windows. For G4, "derived" means the speaker says they computed it.

# Rubric v2: definitions

The scored checks are in the run's `checklist.md`, which is generated from
`scripts/checklist.js`. This file defines the structures you extract before answering them.

Judge only what a transcript can show: words, order and timing. Ignore visuals, editing,
delivery and music, and ignore auto-caption errors.

## Sections
Split the runtime into contiguous **sections**, each covering one sub-topic or step.
Sections are usually 45s to 3 min long, and they must tile the whole runtime with no gaps.
Give each one:
- `id`: "A", "B", "C", … in order.
- `start` / `end`: `mm:ss`. Each section's `end` is the next section's `start`.
- `summary`: one English sentence saying what the section says, not how good it is.
- `function`: one of `hook, promise, context, value, example, story, complication, payoff, recap, transition, cta, sponsor, filler, repetition`.
  `filler` = no new information. `repetition` = restates earlier material without adding anything.
- `depends_on`: the ids of EARLIER sections whose information this section needs in order
  to make sense. For example, it uses a number, a character or a claim set up there.
  Use `[]` if it stands alone.
- `references_unintroduced`: things this section mentions **as if already known** that no
  earlier part of the transcript introduced. Examples: "as we saw", a character's name
  used without introduction, "this second model" when no first model came before. Usually `[]`.
- `interest`: an integer 1–5. How strongly a typical target viewer would want to keep
  watching during this section. 1 = would skip, 3 = fine, 5 = gripped.

## Loops
A loop is a question the script makes the viewer want answered, whether it's asked
explicitly or only implied. Record:
- `opened_at`: when the question becomes live.
- `closed_at`: when it's answered, or `null` if it never is.
- `size`: `major` if it drives a section or the whole video, `minor` if it's answered within about a minute.

## Promise
The core question or promise of the video, with the timestamp where it first becomes clear.

## Topic card
One neutral sentence of 25 words or fewer saying **what the video is about**: its subject
and the question it addresses. Describe subject matter only:
- No judgment of quality, style or format ("gripping", "animated", "well-researched").
- No title-style hype.
- Write it so that two videos on the same subject would get nearly the same card,
  however well each was made.

## Holistic estimate
Imagine the channel made 50 videos. Give the percentile (0–100) you'd expect this script
to rank at on script quality alone, with your confidence.

## Notes for the v3 checks (2026-10-24)
- **H5 and H6** read only the first 60 seconds. For H5, quote the concrete confirmation (a scene, fact or number the title implied) and, separately, the line that promises more than the title did; both must exist. For H6, "direction" means the route or cause. If the opening states the cause outright, or says nothing about where the answer is going, H6 fails.
- **C5, K4 and K5** use your section list. A section "ends on" its last one or two sentences. C5 passes for a section only if the question or unresolved consequence is one the next section actually answers. For K4, a join is flat if you could insert "and then" between the sections without changing the meaning, or if it is a list-style signpost ("here is the second thing"); name each one.
- **G1 to G4** count figures: a figure is a spoken quantity of money, share, count or duration. Plain dates do not count. For G1, a figure is "framed" if a person, loss, question, prediction or contrast appears in the same or the preceding sentence. For G2, scan the transcript in 30-second windows. For G4, "derived" means the speaker says they computed it.

Shown as it is in the repository at build time (2026-11-03); the home directory appears as <home>. If this page and the file differ, the file is right. Back to the catalogue.