Choosing a topic

How subjects are scored for demand, what that evidence says, and the checks a topic must pass before a script is written.

Two questions, not one

.claude/skills/video-script-production/references/s1-rival-brief.md states the standing arrangement: "The user normally names the topic; this report is the evidence for the choice, not a step you repeat." This page lays out that evidence and adds the gates that the Moser Baer and Cafe Coffee Day runs actually applied.

Part 1: the topic study

What it is

A second series spec, series/fall-of-x.json ("The Fall of X"), lists subjects instead of formats: 33 Indian companies (Kingfisher, Jet Airways, Satyam, Byju's, Cafe Coffee Day, Yes Bank, Go First, Vodafone Idea, DHFL, IL&FS, Future Group, Videocon, and others), global companies, people and one economy. The series commands plan, search and enrich collect videos; then:

node src/series-analysis/topic-study.js --series=fall-of-x [--out=reports/series-analysis]

It writes reports/series-analysis/fall-of-x-topics.md and .csv (gitignored, so regenerate when needed). Run on 2026-10-25 at 04:35.

Rules it follows

• On-topic video: the title frames a collapse and names the subject. Shorts and fiction clips are not the topic.
• Attribution: each video counts once, for the subject its title names first.
• Eligible subject: at least two on-topic videos. Result: 5,175 videos searched, 2,610 on-topic, 103 subjects. Only 500 of the 5,175 were enriched (exact dates and likes); the rest use search-page view counts and relative date text.
• Score = 0.45 demand + 0.30 recency + 0.10 engagement + 0.15 headroom, each part a percentile within the eligible subjects:
  • Demand: median views (0.5), best single video (0.3), total views (0.2).
  • Recency: views in the last 12 months (0.6) and median views per day (0.4).
  • Engagement: median likes per view, enriched videos only; neutral when there are no likes.
  • Headroom: total demand plus the share of videos that did not reach 100K views. High demand with few strong videos is a gap.
• Not measured: shares (the data does not expose them), thumbnails, click-through, retention.

The top 25

Views are search-page counts unless the video was enriched. "Strong" means a video above 100K views.

The bottom of the 103: Juicero, Subhiksha, Argentina economy, Gitanjali Gems, Parmalat, Ramalinga Raju, Byju Raveendran (the person), Webvan, Rotomac Pens, Kenneth Lay (scores 0.26 down to 0.09).

By type, and by title frame

Robustness: does the ranking survive different weights?

In all four lists: Harshad Mehta, Pan Am, Nokia, BlackBerry. Cafe Coffee Day appears in two (headroom-heavy and equal).

How to read it (my reading; the study states none of this)

• The Indian subjects with the strongest evidence are Harshad Mehta (a person, not a company), Nirav Modi PNB, Cafe Coffee Day, Future Group, Jet Airways, Byju's and Ola Electric. Ola Electric has the least supply (17 videos, 3 strong) and 912K views in the last 12 months, which the score rewards as headroom.
• Indian subjects have higher typical views per video than global ones (median of medians 1.6K against 452), while global subjects have far higher best videos (1.4M against 403K). Both facts fit an Indian-language audience that mostly watches Indian stories.
• The title-frame table is a median over all on-topic videos regardless of channel size, so it mixes frame, subject and channel. "The fall of X" has a median of 1.4K and "case-study" 10.5K. That is a tension with the series name that nobody has resolved; a frame is not shown to cause the difference.
• Moser Baer does not appear because it was not in the spec's company list. Absence is not a low score.

Names the study found but did not search

Titles that name a subject at least three times and were not in the spec (the "next round" list): Ambassador (5 videos, 525K views total), Go First (3, 901K), Woolworth, the PNB fraud, Kozmo.com, dead social networks. Add the real ones to the spec and rerun search.

Part 2: what other evidence says about subjects

The seven within-channel studies (transcripts/_cross-channel-study.md) were written for a different question, but their strongest findings concern the subject and the promise. They are leads: seven separate hypothesis sets by one analyst, not a pooled test.

HinExp's own headline (from its study): the channel wins when the title promises to open up something the viewer already touches in their own life and shows where the money goes; it loses on other people's drama or fantasy. Its top video is "The Economics of Owning a Bus Business" (56,667 views per day, 37 times its median). The series premise borrows HinExp's style but not its best-performing subject type. This is worth deciding on purpose: a collapse story works best, on this evidence, when the company is an object the viewer has touched. Moser Baer's hook uses a blank disc; Cafe Coffee Day's uses a café. That fits, and was a design choice made in the hook, not a selection rule.

From the 20-video teardown: 19 of 20 titles carry a household brand or person. From the blind study (post-hoc): the weakest new performers were polished Kodak and Sears videos about American companies. From the user (2026-10-20): off-company collapses are fine as adjacent finds and should not be discarded.

Part 3: the gates a topic must pass

The first column says where the gate is written down. "Synthesis" means the gate is my summary of how the Moser Baer and Cafe Coffee Day runs actually proceeded; it was not a named gate before this page.

The two topics so far

Candidate list for the next topic

This is a reading of the study, not a recommendation. Indian subjects ranked in the top 20 that have no script yet: Nirav Modi PNB Scam (9), Future Group (11), Jet Airways (13), Byju's (14), Ola Electric (15), Vodafone Idea (20). Harshad Mehta (1) and Subrata Roy (17) are people; the "Why X died" cohort rule excluded a person profile (Elizabeth Holmes), while the user said off-company collapses are fine as adjacent finds, so a person-led topic is a decision to ask about. Each still needs gates T2 and T5 to T9. Jet Airways and Future Group already have rival videos in the 20-video teardown (Rajnikant Sharma's Big Bazaar at 2.7M, two Jet Airways case studies), so a gap would have to be found in their weak points.

How to choose the next topic, step by step

1. Rerun search and enrich for fall-of-x if the study is older than a few weeks, add the "next round" names, rerun topic-study.js.
2. Take the top Indian company subjects. For each, read T3 and T4 honestly: is it an object the viewer has touched, and is it one company?
3. For the top two or three, run S1 on a short setting (--max=4) to see the rival set, or just read search.json candidates, before committing to a full run.
4. Do a thirty-minute S3 pass on the person and the primary documents (T6, T7). Drop topics where the cause cannot be sourced.
5. Give the user a short comparison with the evidence and the unknowns, and let them choose. State the title tension and the sensitivity items at that point, not at the hand-off.

Written 2026-10-27 from the repository files and the project memory notes. If a page and a file disagree, the file is right: fix the page.