title-research

Generate YouTube titles by mining what actually performs, not by guessing.

Pipeline: research → extract patterns → generate → score → report.

Depends on:

• firecrawl CLI (authenticated)
• the title-score skill, installed as a sibling directory of this one

Paths below are relative to this skill's base directory, not the user's project. Resolve them from the SKILL.md you are reading, e.g.:

SKILL_DIR=<absolute path of the directory containing this SKILL.md>
SCORE="$SKILL_DIR/../title-score/scripts/score.py"

If $SCORE does not exist, title-score is not installed: say so and stop before Step 4 rather than reporting unscored titles as ranked.

Step 1 — Research top titles

python3 "$SKILL_DIR/scripts/research.py" "web scraping AI" --sort views

The script scrapes YouTube search results sorted by view count, uses firecrawl interact to extract title | channel | views, and falls back to firecrawl search site:youtube.com ... when interact fails (common: YouTube scrape jobs die with "Job not found" or 401 consent pages).

Run 2-3 query variants for broader coverage (e.g. "web scraping AI", "AI agent web search", "scrape anything"). If both strategies fail, tell the user instead of inventing data.

Step 2 — Extract winning patterns

From the highest-view titles, note concrete formulas, e.g.:

• "Use AI To Scrape ANYTHING" → [VERB] + [TECH] + To + [VERB] + ANYTHING
• "This AI Agent Can Scrape ANYTHING (100% Automatic)" → This X Can Y (promise)
• "AI Web Scraping Is Insanely Good" → X Is Insanely Good
• "How To Scrape Any Website With X" → How To + task + With + tech
• "in 10 Minutes" / "in 100 Lines" → concrete effort qualifier
• CAPS emphasis words: ANY, ANYTHING, FREE, BLIND, INFINITE

Step 3 — Generate candidates

Apply the patterns to the user's video. Rules:

• Honor hard constraints (e.g. "must include Convex", no clickbait lies).
• Generate 10-15 candidates across at least 4 distinct patterns.
• Keep each title truthful to the actual video content.
• Reuse exact power words (ANYTHING, Insanely, FREE) and number qualifiers (minutes, lines, pages) from the real winners.

Step 4 — Score and rank

printf '%s\n' "Title 1" "Title 2" ... | python3 "$SCORE" -

Iterate once: tweak the weakest sub-scores of the top half (add a digit, a colon/question mark, a power word) and re-score. Keep the best 10.

Step 5 — Report

Table ranked by score:

| # | Score | Title | Inspired by (real video + views) |

Note which sub-scores help/hurt each title, flag that the score is a heuristic approximation (see title-score skill), and recommend the top 1-2 with a one-line rationale tying title to actual video content.

Honesty

• Never present invented titles as "real videos". Every inspiration must come from Step 1 output.
• If view counts are unavailable (search fallback), say so.
• The final score is the title-score approximation, not vidIQ.