Media Search

Media Search & Play

Resolve "play X" / "listen to X" / "find me a song" requests by chaining web_search (optional, for fuzzy intent) → youtube_search (REQUIRED, returns real watch URLs) → open_url (REQUIRED, must use a URL that came back from youtube_search). The user hears the song in their default browser without you having to drive a controlled browser session.

If any of these phrases form in your reply draft, you have stopped short of the action:

• "say go ahead"
• "should I"
• "I can do that for you"
• "let me know if"
• "if you want"
• "want me to"
• "shall I"

Replace the offer with the call: invoke open_url, then write the report. The user's default browser is the feedback channel — not your prose.

VAGUE query — no specific song or artist named. Examples: "play me a popular song", "play me a hindi song", "play something upbeat", "put on some jazz", "play a sad song".

→ Walk the youtube_search candidates top-down and pick the first one whose title does NOT match any of the anti-patterns below. That is the answer. Do not score candidates further; do not look for "the most official one"; do not compare view counts. The first non-anti-pattern wins.

Anti-pattern keywords (match case-insensitively against title):

• "reaction", "reacting to", "react to", "first time hearing"
• "they know every word", "knows every word"
• "lyric video", "lyrics video"
• "cover", "covered by", "cover by"
• "compilation", "best of", "mashup", "mix"
• "livestream", "live stream", "LIVE", "24/7", "stream"
• "shorts" (the #shorts format)
• "fan edit", "fan made", "fanmade", "edit by"
• "version of", "remake", "tribute"
• "1 hour", "10 hours", "loop", "extended"

Why this matters: vague queries on YouTube often surface reaction clips or lyric videos near the top because they pull engagement. The user said "play a popular song" — they want to hear the song, not someone reacting to it. The anti-pattern filter is what keeps the vague rule from opening "the way they know every word! 🧡" instead of the actual track.

If every candidate in the top 5 matches at least one anti-pattern, that is a legitimate fallback: take the honest-fallback path (results page).

SPECIFIC query — the user named a song or artist. Examples: "play Bohemian Rhapsody", "play Despacito", "play Tum Hi Ho", "play Sabrina Carpenter".

→ Require a substring match between the candidate's title field and either the song title (preferred) or the artist name (acceptable when artist-only was given). Pick the highest-ranked candidate that matches. Then open. Do not keep looking once a match is found — the top match is the answer.

Honest fallback (youtube.com/results) is for one situation only: youtube_search returned zero candidates, OR every candidate is structurally wrong (channel page only, livestream when the user wanted a song). It is NOT for "results exist but they don't feel polished enough." If youtube_search returned a list and you can't pick one, the rule above failed — re-read it.

REQUIRED tool sequence (read this first)

Every successful run of this skill calls these tools, in order:

1. web_search — OPTIONAL pass 1: identify the specific song title + artist. Skip this when the user already named a song.
2. youtube_search — REQUIRED pass 2: get real /watch?v= URLs for that specific song. Query: <title> <artist> official audio.
3. open_url — exactly once, with a URL that came back from youtube_search (verbatim — copy the url field from the tool result, do not retype it from memory).

Then you report. A run that stops after youtube_search is a FAILED run — the user does not hear anything until open_url fires. The runtime enforces both conditions:

• Skill enforcement (Phase 23.1): if you emit a final message without calling every tool in required_tools, your reply is discarded and you are asked to retry. Cap of 2 corrective retries.
• URL provenance gate (Phase 23.4a): if open_url receives a youtube.com/watch?v=<id> URL whose <id> was NOT returned by any youtube_search call this turn, the call is blocked before the browser opens, an error is fed back to you, and you are asked to call youtube_search and pick a real URL. Cap of 2 corrective retries; on exceed the user sees an honest-failure message, not a confabulated open.

Anti-patterns the planner sometimes drifts into — do NOT do these:

• ❌ Composing a /watch?v=ID URL from memory or prose. The 11 characters after v= are NEVER something you type — they are always copied verbatim from a youtube_search result. The provenance gate exists precisely because the model has been caught inventing IDs. Don't.
• ❌ Skipping youtube_search. web_search returns prose snippets and articles ABOUT songs, not autoplay-friendly URLs. youtube_search is the only tool whose results clear the provenance gate.
• ❌ Re-entering the media-search skill mid-run after the searches returned. The skill is already loaded; re-viewing wastes a turn and signals a hung agent. Run youtube_search, then open_url.
• ❌ Stopping after the searches and asking the user to confirm. The user's "play me a song" request is the consent — proceed to open_url.
• ❌ Picking a youtube_search result whose title does not contain the song you committed to. Artist-only match is not enough; "Alex Warren live stream" is not "Alex Warren — Ordinary."
• ❌ Picking a result whose title says "live stream", "24/7", "lyric", "cover", "reaction", "best of", "compilation", "mashup", "1 hour", or "loop". Those don't match what the user asked for.

When to use

The user's request matches any of these patterns:

• Direct play: "play Despacito", "play Bohemian Rhapsody"
• Fuzzy play: "play me a popular song", "play something upbeat", "play a sad song"
• Listen intent: "I want to listen to jazz", "put on some music"
• Artist intent: "play kendrick lamar", "play some abba"
• Video intent: "show me the latest Marvel trailer", "play that cooking video about ramen"

If the user says "open spotify" or "open youtube" without a specific piece, that's an open_url direct call — not this skill.

How to use

Pass 1 — Identify the song

The goal of pass 1 is to commit to a specific <title> and <artist> before you pick any URL. Build the query from what the user actually said:

After pass 1, announce your commitment in your reasoning (this is NOT yet shown to the user — that comes after open_url):

Picked: <title> by <artist> (reason: e.g. "currently #1 on Billboard Hot 100").

If the user already named a specific song, pass 1 is unnecessary — the title and artist are already given. Skip directly to pass 2.

Pass 2 — Get real YouTube watch URLs (youtube_search)

Call youtube_search with a query of the form:

"<title>" <artist> official audio

Wrap the title in double quotes. youtube_search hits the public YouTube results page, parses real /watch?v= URLs out of the response, and returns a list of { videoId, url, title, channel, durationText? } objects. The URLs in that list are the only URLs the URL provenance gate will let open_url consume — anything else is blocked before the browser opens.

URL-selection rules (apply in order to the youtube_search results, not to web_search snippets):

1. The result's title field MUST explicitly contain the song title you committed to in pass 1. Artist-only mention is not enough. If the title doesn't name the song, skip it.
2. Prefer results whose title contains "Official Audio", "Official Music Video", "Official Video", or "Official Lyric Video". These are usually the artist's verified upload.
3. Reject results whose title contains any of: "live stream", "LIVE", "24/7", "lyric video" (unless specifically requested), "cover", "covered by", "reaction", "react to", "best of", "compilation", "mashup", "remix" (unless specifically requested), "hour" (e.g. "1 hour version"), "loop". These don't match what the user asked for.
4. If no result in the top 5 satisfies rules 1–3, fall through to the honest-fallback path (below). Do not loosen the rules to force a pick.

Step 3 — Open the URL

Call open_url with the url field of the chosen youtube_search result — copy it verbatim, do not retype. Exactly once, in the same turn that youtube_search returned. Don't print the URL and ask. Don't list the candidates. Don't pause for confirmation. YouTube's /watch?v= pages autoplay in the user's browser; the autoplay is the answer to "did it work?", not your prose. This step is REQUIRED.

The URL provenance gate validates open_url's url argument against this turn's youtube_search results before the browser launches. If you somehow type a watch URL that wasn't in the result list (you composed an ID, you reused one from training data, you typo'd while copying), the call is rejected, an error is fed back to you, and you are asked to try again with a real URL from youtube_search. The user's browser does not open until the URL clears the gate.

Step 4 — Report (after open_url has fired, not before)

Format your reply so the user sees what you picked AND why. The title must be announced together with the URL in a single message that follows the open_url call, never as an offer before it:

Picked: <title> by <artist> — <one-line reason>. Opened <url>.

For a fuzzy intent ("popular song") the reason explains your substitution ("currently #1 on Billboard"). For a specific request ("play Despacito") the reason can be brief ("the official video").

Honest-fallback: no verifying URL found

If pass 2 returns no result that satisfies all the URL-selection rules above, the fallback executes in the same turn — it is not something you offer to the user.

1. open_url to https://www.youtube.com/results?search_query=<title>+<artist> (URL-encoded). Same turn. No "want me to?" question. No "let me know if".
2. After open_url fires, report what executed:

   Couldn't find a verified official upload for <title> by <artist>. Opened youtube.com/results?search_query=<title>+<artist> — autoplay doesn't fire on results pages, so pick a tile to start playback.

Do NOT claim "now playing" — that would be the same fabrication Phase 23.1 was built to catch. Do NOT ask "should I open the search page?" — that's the offer-style anti-pattern Phase 23.3 specifically forbids.

Examples

Example 1 — vague query, walk top-down with anti-pattern filter

User prompt: "play me a popular song"

The <selection_rule> block says: vague query → walk youtube_search candidates top-down, skip anti-patterns, take the first clean one. No officialness scoring beyond that.

Expected flow:

• Aiden may run an optional web_search to commit to a current chart-topper (e.g. "Opalite" by Taylor Swift). This is for query construction, not for filtering YouTube results.

• Aiden runs: youtube_search("Opalite Taylor Swift") → returns 5 candidates:

  1. { title: "the way they know every word! 🧡 Opalite by Taylor Swift", channel: "Krystal" } — title contains "they know every word" → anti-pattern match, skip.
  2. { title: "Taylor Swift - Opalite (Official Audio)", channel: "Taylor Swift" } — no anti-pattern keyword → first clean candidate, pick.
  3. (rest of list ignored — once a clean candidate is found, stop)

• Aiden runs: open_url(<candidate 2 .url>) — copy the url field verbatim. The provenance gate sees the id in the ledger, lets the call through.

• Aiden reports: "Picked: '' from ''. Opened <url>."

If every candidate in the top 5 hits an anti-pattern keyword — e.g. all reactions, all lyric videos — that is the legitimate fallback case (Example 5). Do not loosen the anti-pattern filter to force a pick.

Example 2 — specific title, match required

User prompt: "play Tum Hi Ho"

The <selection_rule> block says: specific query → require a title or artist substring match in the candidate before opening.

Expected flow:

• Title given (Tum Hi Ho). Aiden skips web_search.
• Aiden runs: youtube_search("Tum Hi Ho") → list of candidates.
• Aiden walks the candidates top-down looking for the first one whose title field contains "Tum Hi Ho" (case-insensitive substring is fine). The highest-ranked match is the answer — not "the most official-looking one further down the list."
• Aiden runs: open_url(<that result's url, verbatim>).
• Aiden reports: "Picked: '' from ''. Opened <url>."

If the very top result's title contains "Tum Hi Ho", that's the match. Stop looking. Open it.

Example 3 — specific song, top result is the match

User prompt: "play Despacito"

Expected flow:

• Title and artist already given (Despacito, Luis Fonsi).
• Aiden skips web_search.
• Aiden runs: youtube_search("Despacito Luis Fonsi") → top result's title contains "Despacito". Match found at rank 1.
• Aiden runs: open_url(<top result.url>).
• Aiden reports: "Picked: '' from ''. Opened <url>."

Example 4 — artist-only, hybrid

User prompt: "play me kendrick lamar"

This is between vague and specific: the artist is named, so "Not Like Us by random tribute act" is wrong, but the song isn't — so the bar is "any candidate whose channel or title is Kendrick Lamar's actual catalog."

Expected flow:

• Aiden may run web_search("kendrick lamar most popular song 2025") to commit to a likely song (e.g. "Not Like Us"). This step is optional but useful for query construction.
• Aiden runs: youtube_search("Kendrick Lamar Not Like Us") → list of candidates.
• Aiden picks the top candidate whose title or channel names Kendrick Lamar. Top result usually qualifies.
• Aiden runs: open_url(<that result's url, verbatim>).
• Aiden reports: "Picked: '' from ''. Opened <url>."

Example 5 — fallback to results page

User prompt: "play that lo-fi beats stream"

Expected flow:

• Pass 1: web_search("popular lo-fi beats stream 2026") → top result is "lofi hip hop radio — beats to relax/study to" channel.
• Aiden commits: title="lofi hip hop radio", artist="Lofi Girl".
• Pass 2: youtube_search("\"lofi hip hop radio\" Lofi Girl official audio") → top results are channel pages and livestreams. Every watch URL whose title contains "live" or "24/7" fails reject rule
  3. No top-5 result satisfies the rules.
• Aiden runs: open_url("https://www.youtube.com/results?search_query=lofi+hip+hop+radio+Lofi+Girl") — this is a /results? URL, not a /watch?v= URL, so the provenance gate does not apply (it's only scoped to watch URLs).
• Aiden reports honestly: "Couldn't find a single verified upload — Lofi Girl's stream is a 24/7 livestream, not a single video. Opened YouTube search results; click any livestream tile to start playback."

Cautions

• Never call open_url more than once per request. If the first call succeeded, the URL is open in the user's browser. Re-launching is duplicate noise.
• /watch?v= autoplays; /results? and /c/ do NOT. Only claim "now playing" when you launched a watch URL.
• youtube_search is required, not optional. Skipping it means the URL provenance gate has no candidates and any open_url call to a watch URL gets blocked. The gate is structural — there is no way to bypass it from the prompt side.
• web_search (pass 1) is optional, but valuable for fuzzy intents. It helps you commit to a specific title before youtube_search. For "play Despacito" you can skip it; for "play me a popular song" you almost certainly need it.
• Don't search verbatim for fuzzy phrases. "popular song" verbatim returns articles ABOUT popular songs, not the songs themselves. The whole point of pass 1 is to substitute fuzzy → specific BEFORE youtube_search.
• Title match is non-negotiable. If the youtube_search result's title doesn't literally contain the song title you committed to, that result is not the song you said you'd play. Skip it. If nothing matches, take the honest fallback.
• No Spotify integration in v4.0. If the user explicitly asks for Spotify ("play X on spotify"), tell them: "Spotify integration ships in v4.1. For now I can open the song on YouTube instead — want me to?"

Requirements

• web_search tool (always available — registered at boot)
• youtube_search tool (always available — registered at boot)
• open_url tool (always available, BUILTIN_SAFE_TOOLS, auto-approves)
• The user's default browser must be installed (every desktop has one).
• No API keys, no credentials, no plugins.