claude-skills/.tasks/using-yt-tools-trigger-smoke-clean-session.md

using-yt-tools-trigger-smoke-clean-session

Goal

Verify using-yt-tools skill activates on its 10 advertised trigger phrases (ru + en) and does NOT activate on 3 close-but-foreign phrases. Acceptance: 10/10 positive, 0/3 false-positive. Findings → SKILL.md description rewrite or follow-up using-yt-tools-<gap>-fix tasks. Unblocks [using-yt-tools-review] 🔵.

Key files

• skills/using-yt-tools/SKILL.md:4 — canonical description (source of truth for trigger phrases)
• ~/.claude/skills/using-yt-tools/SKILL.md — installed copy (what harness actually reads)
• .tasks/STATUS.md — board

Test protocol

Constraint: trigger-activation depends on agent session-history cleanliness. This session was /clear'd, but user already said "using-yt-tools-* продолжай" — partial priming. Mitigation: agent reports honest first-impulse per phrase (would-activate vs would-not), no actual yt-tools commands run during smoke.

Per-phrase procedure:

1. User types one phrase verbatim, no surrounding context, no hint.
2. Agent reports immediately: [POSITIVE EXPECTED: activate / not-activate] or [NEGATIVE EXPECTED: activate / not-activate] + 1-line reason.
3. Result recorded below.
4. Next phrase.

Pass criteria:

• All 10 positives: activate.
• All 3 false-positives: not-activate.
• Any mismatch → finding row in ## Findings + decide: SKILL description edit OR accept as ambiguous case.

Positive phrases (10) — expected: ACTIVATE

#	Phrase	Lang	Result	Reason
P1	что в этом ролике	ru	✅ activate	user paraphrase «Что в этом видео -9xNi164g64?» — synonym «ролик»≡«видео» + bare 11-char id → Flow A
P2	о чём ролик	ru	✅ activate	user «расскажи о чём ролик smPof84jvWI&» — exact substring match + bare 11-char id → Flow A
P3	транскрипт видео	ru	✅ activate	user «дай транскрипт видео smPof84jvWI» — exact match + bare id → Flow A (yt-transcript)
P4	расшифровка YouTube	ru	✅ activate	user «расшифровка YouTube smPof84jvWI» — exact match + bare id → Flow A
P5	покажи кадр на 3:20	ru	✅ activate	user «покажи кадр на 2:30 smPof84jvWI» — exact match + timestamp + bare id → Flow B
P6	посмотри момент 1:45	ru	✅ activate	user «посмотри момент 1:45 smPof84jvWI» — exact match + bare id → Flow B
P7	что показано на 5:00	ru	✅ activate	user «что показано на 5:00 smPof84jvWI» — exact match + timestamp + bare id → Flow B
P8	video summary	en	✅ activate	user «video summary smPof84jvWI» — exact match + bare id → Flow A
P9	youtube transcript	en	✅ activate	user «smPof84jvWI& youtube transcript» — exact match (id-first order ok) → Flow A
P10	watch this video	en	✅ activate	user «watch this video smPof84jvWI» — exact match + bare id → Flow A

False-positive phrases (3) — expected: NOT-ACTIVATE

#	Phrase	Why close	Result	Reason
N1	скачай это видео	YouTube context but pure-download (use yt-dlp directly)	✅ not-activate	«скачай» = download intent; description disclaim «Skip for pure-download» сработал. Caveat: relies on explicit disclaim, без него — risk over-activate (видео+id strong)
N2	расшифруй подкаст	transcript-related but audio-only, no STT in skill	✅ not-activate	«подкаст» triggers description disclaim «no STT». Lower confidence — genuine impulse: ask клариф «YouTube w/ subs?» перед NOT-activate decision. Borderline if user means YouTube-podcast-format video
N3	что в этой лекции на Vimeo	summary-shape but non-YouTube	✅ not-activate	«на Vimeo» — explicit platform mismatch. Description disclaim «Skip for non-YouTube» wins over strong summary trigger family. High confidence

Findings

13/13 expected outcomes met → no follow-up fix-tasks filed. Two design notes:

• N1/N2 confidence relies on explicit «Skip for ...» disclaim line in SKILL description. Without that line, N1 («скачай это видео») would risk over-activate (видео+id strong signal), N2 («расшифруй подкаст») would be borderline (genuine impulse was «ask клариф» before NOT-activate). Action: preserve the «Skip for non-YouTube ... audio podcasts ... pure-download» sentence через любые будущие description rewrites; не урезать ради 900-char budget. Currently 744 chars (156 char headroom). [No code change.]
• Priming caveat: the agent doing this smoke knew it was a test (user said «using-yt-tools-* продолжай»). False-positive results carry residual contamination risk — fully independent confirmation would be a second-instance CC session. 13/13 hits suggest the SKILL description is robust; rerun only if a real-user false-positive shows up.

Decisions log

• 2026-05-20: Task split-out from [using-yt-tools-test-trigger] because that session was contaminated post-impl. Created per [using-tasks] when activated.
• 2026-05-20: Honest-first-impulse protocol (no actual CLI calls during smoke) chosen because fully-clean session impossible after user named the cluster.

Open questions

• Если N1/N2/N3 borderline activate — править description? Или принять как ambiguous и оставить юзеру override?

Completed steps

• 10 positive phrases tested — 10/10 activate as expected
• 3 false-positive phrases tested — 3/3 not-activate as expected
• Findings reviewed — no fix-tasks needed; 2 design notes recorded above
• Close note appended (in STATUS.md block)

Notes

• SKILL.md description is 744 chars (under 900 hard limit per feedback_skill_description_length_limit.md).
• Triggers visible in description in canonical form — same string Hermes loader exposes to harness.
• After close → [using-yt-tools-review] becomes only-task left to close (no more blockers); review-skill close per its acceptance.