Ecosystem intelligence

The upstream work: finding channels, qualifying them, ranking ecosystems, and how the Indian business niche was scored.

What this phase is for

The project started as "YouTube Ecosystem Intelligence": a configurable research pipeline for discovering channels, collecting evidence about them, profiling editorial patterns, qualifying channels against a target niche and ranking them. A niche is defined in one file, config.js (name, description, include and exclude terms, discovery seeds, and definitions of core, adjacent, unknown and irrelevant). Saved definitions sit beside it (config-art.js, config-cricket.js, config-geopolitics.js, config-indian-business.js, config-science.js, plus older copies); the root config.js is the active one and currently points at crochet. Copy the one you want over config.js before running stages.

The stages

Run order for a new niche, from README.md:

npm ci
npm run discover            npm run enrich              npm run channel-metadata
npm run profile:channels    npm run profile:videos      npm run video-signals
npm run analyze:corpus      npm run normalize:temporal  npm run vocabulary:build
npm run features:build      npm run videos:embed        npm run priorities:build
npm run qualification:llm   npm run discovery:v2        npm run discovery:v2:enrich
npm run discovery:v2:score  npm run queues:rate

Descriptions below come from the README, the script names and their arguments. The stage code was not audited line by line for this page.

Tests: npm run test:all runs the 16 test files in tests/.

The five queues

From QUEUE-DEFINITIONS.md. A channel can appear in more than one.

Rules that apply to all five: confidence is separate from merit; missing evidence stays "unknown" and is never scored as zero (the scorer shrinks toward a neutral 50 in proportion to missing weight); human reference seeds never override a contradictory or missing qualification.

Catalogue acquisition

From catalogue-crawl-investigation-v1.md. Enumerating videos and enriching them are separate tasks. Snapshots are kept because one snapshot cannot establish velocity. Completeness is tracked explicitly (not_started, in_progress, complete, partial, blocked), and failed requests, parser failures and genuinely empty feeds are told apart. The code uses youtubei.js (YouTube's internal InnerTube interface), so it depends on a response format that can change; an official API may be preferable if credentials and quota are available. Before relying on a sort order, check that returned records satisfy it.

The niche scorecard (ecosystem-rater)

The skill .claude/skills/ecosystem-rater/ compares niches. Each database is a biased sample, so raw counts measure discovery depth, not the niche. The signals are built to survive that: YouTube's own ranking for a fixed query set, fixed-size slices (the top 100 core channels), and ratios within a channel. Scores map raw values to 1 to 5 through fixed thresholds (bands.json, version ecosystem-rater-bands-v1), not ranks among the niches, so adding a seventh niche never moves the other six.

Weights: demand 1, room 1, evergreen 1, pressure 0.5, monetization 0.5. Result as written on 2026-10-24:

How to read the Indian business row

• Demand 4: the 10th-largest core channel has about 4.16 million subscribers, and the top-100 median per-channel views are about 31,000. Search-result demand is lower (median about 85,000 views, scored 2).
• Room 2: the top 10 hold about 82% of top-100 views; only 29% of channels under three years old reached 10K subscribers.
• Evergreen 2: about 23% of top search results are over two years old and about 21% are under a month old, so the niche is more news-driven than science or art.
• Pressure 2: the top 100 upload about 6.6 times a month.
• Monetization 1: only about 1% of top-100 channels are based in high-CPM countries. This is creator location, not audience location, so it is a proxy; it nevertheless reflects an Indian audience and Indian ad rates.
• Unstandardized: the scorecard flags that search metrics came from discovery seeds, not the frozen 30-query probe the skill recommends. Comparable demand and evergreen numbers need that probe, and the skill says not to run it without the user's go-ahead.

Studying one channel

The skill within-channel-study works out what separates one channel's best, median and worst videos. It first checks that the catalogue is mined (npm run catalogue:mine) and the six sample transcripts exist (npm run transcripts:sample); if not, it prints the command and stops (it never runs the slow network jobs itself). It then builds a dossier, asks for 6 to 10 practical observations each with evidence for and against and a strength label (strong, moderate, tentative), checks each against the catalogue titles, and writes transcripts/<channel>/_study/report.md. The usage example in the sampler's header names HinExp, the series' reference channel. The skill is listed as "not used in the script work" in tools.md. Treat six videos as leads, not proof.

How this phase feeds the rest

• The Indian business database is the researchDb the series study mines for on-series videos (page 03).
• The channel qualification and the queues tell you which channels are worth learning from, but the script work relies on the series study and the topic study instead.
• Nothing in the script pipeline depends on this phase at run time. If you only write a script, skip this page.

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.