Find an opening before building, and track a launch after submitting - #6

Two new skills.

• find-opening: before you build. Which ChatGPT, Claude or Cursor tags and categories are growing or crowded, how locked their top apps are, where one store is thin next to another and which countries are thin, then up to five ranked openings with dated evidence. Every report says it counts listings, not demand or usage.
• track-launch: after you submit. Compares your wait with similar ChatGPT apps that went live (the start time is decoded from the app ID, so it works before the app is live), checks which of the 48 countries list it, sets its tags against the tags delisted apps carry more often, and counts returns. Hands off to optimize-listing once live, or to submit-plugin after a rejection or a wait past the 90th percentile.

Both are wired into the README (table, diagram, usage), the plugin-stack menu, the manifest descriptions, a hand-off line in submit-plugin, evals/confusion-pool.json (14 new queries, 2 relabelled) and evals/traps.md (3 rows).

Verified on live data (2026-11-06, CLI 0.1.13)

• Every operation and field the skills use was run signed out, with Pro where noted. Plan limits that differ from shared/data.md: signed out, new_list reaches 30 days, and charts_get at=-28d quietly returns a snapshot from 7 days back.
• track-launch on Hyatus (chatgpt:hyatus, Travel, went live 2026-11-06). Its wait was 10.1 days against a Travel median of 14.3 across 200 go-lives, which is early. It is listed in 48 of 48 countries. commerce shows up 1.62x more often among delisted apps. None of 92 delisted apps have come back since 2026-10-14.
• track-launch facts: the first 8 hex digits of the app ID match launched_at. new_list has no lag_days, so the skill computes the wait itself. Delisted rows have no category, so the skill matches on tags. There's no relisting event, so a return is inferred.
• find-opening on a fitness idea for ChatGPT and Claude. It found fitness crowded on ChatGPT, no extra room on Claude, and mental health and sleep as the opening it could measure. The run changed the skill: on ChatGPT the top apps now come from the category chart, not the tag.

Checks

• node scripts/check-skills.mjs: 8 skills pass. node scripts/check-contract.mjs: ok.
• scripts/test-installs.sh: every install path got 8 skills.
• evals/run_confusion.py on sonnet, 1 run each, over 42 queries (the new ones plus submit-plugin and watch-competitors). New skills: 13 of 13. The 3 misses went to plain text or Bash, so none were taken by the new skills.

Follow-up

shared/ doesn't cover tags_tree, market_get or charts_regions_get yet, and the checker's command regex misses them.

🤖 Generated with Claude Code

https://claude.ai/code/session_015NmuwgGtdf9fGS9mYJjUuP