Packaging: titles and thumbnails

WHEN

Generating or evaluating titles, diagnosing a thumbnail, reviewing the title+thumbnail pair, writing the description and tags.

EXECUTION ORDER

1. Score the titles — mandatory, never opine first python3 tools/score_titles.py --title "A || B || C" Returns 0-100 with a per-axis breakdown and the memory justifying each point. It is heuristic: you must declare it is an own rubric, not a CTR prediction, and that it is unvalidated until the Studio CSV is ingested.

1b. Two rules about the number, before you look at it

Do not chase the score. If an honest title scores 78 and an exaggerated one scores 95, recommend the 78 and say why. The rubric measures resemblance to what worked, not whether the title is true; well-built clickbait scores high and damages the channel.

Scores are only comparable within the same call. Score the current title ALONGSIDE the candidates, in one run (--title "current || A || B || C"). Comparing today's number with another session's is not valid. And only propose a candidate that beats the current one strictly: a tie is not an improvement.

2. Demand backing

   python3 tools/yt_search_terms.py --days 90        # REAL demand, first
   python3 tools/kw_research.py --kw "candidate 1; candidate 2"

   The first one wins: those are terms people already reached the channel with. The second is a proxy and only ranks candidates against each other.
3. Real references, not intuition python3 tools/yt_outliers_channels.py --min-ratio 2.5 Look at how the niche's live outliers are titling right now.
4. The thumbnail

   python3 tools/yt_thumbnails.py --video ID
   python3 tools/score_thumbnail.py --video ID --title "the chosen title"

   score_thumbnail only scores the automatic 40%. It returns the other 60% as unanswered questions, and you must answer them by opening the JPG with vision. Inventing that 60% is the failure this tool is designed to prevent.
5. If the thumbnail has to be CREATED, not just judged This skill scores what exists. To generate it: image-generation-core (base rules) + thumbnail-best-practices (design), and likeness-preservation if a face appears. If there is a specific reference to work from, thumbnail-inspiration. A generated image comes back as source: generated: you cannot score it and then present that score as a CTR forecast.
6. Anti-redundancy contrast — the step most often skipped Check that title and thumbnail are not communicating the same concept. If the hard number is in the title, the thumbnail shows the visual result. If memory/sop/niche_bend_sop.md exists, its lessons outrank any general rule.

SOURCES

Steps 1 and 2b heuristic — always declare it. Step 2a youtube_api. Steps 3 and 4a derived. Step 4b is mixed: derived in the pixels, the agent's judgement for the rest.

OUTPUT

1. The titles ranked by score, with the per-axis breakdown.
2. One line per title explaining which axis lifts or sinks it, citing the memory.
3. The thumbnail diagnosis with the judgement axes already answered after looking at the image.
4. The title-thumbnail contrast: which angle each one covers.

Never say a title "is going to work". Say which real video it resembles, and why.