SkillAgentSearch skills...

media-search

Find and play music/videos via youtube_search + browser_navigate

Install / Use

npx skills add taracodlabs/aiden --skill media-search

Installs into whichever agent you are using.

About this skill
📄

SKILL.md

Installable skill definition

Quality Score

86/100

Category

Other

Supported Platforms

Universal

Tags

Our assessment of media-search

media-search scores 86/100 on our quality scale, 187th of 331 Other skills we index.

Its SKILL.md is 17 KB long, well organised into 17 sections with 1 code example: a thorough specification that gives an agent plenty to work with.

It has 843 GitHub stars, a meaningful sign that others use it.

Substance
30/30
Structure
17/20
Description
12/15
Adoption
12/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated about 2 months ago, so media-search is actively maintained.
  • Our last check on 2026-11-03 found the source still online.
  • It is released under AGPL-3.0, a copyleft license: you can use it, but modified versions you distribute must carry the same license.
  • Its trust signals score 100/100, with no cautions. These come from repository metadata, not a code audit — read the skill file before letting an agent act on it.

Safety scan

No issues found

Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful.

AI review by kimi-k2.7-code on 2026-10-06. Automated pattern scan on 2026-10-05. It catches known dangerous patterns, not every risk — read a skill before letting an agent act on it.

media-search compared with similar skills

All 4 of these similar skills score higher than media-search; compare them before choosing.

SkillScoreStarsUpdatedFormat
media-search (this skill)by taracodlabs868432mo agoSKILL.md
x402by browser-use100121.0k4d agoSKILL.md
algorithmic-artby anthropics100177.9k56d agoSKILL.md
pptxby anthropics100177.9k56d agoSKILL.md
prospectingby coreyhaines3110060.1k6d agoSKILL.md

Frequently asked questions

How do I install media-search?
Run npx skills add taracodlabs/aiden --skill media-search. The install tabs above show the steps for each supported agent.
Which AI agents does media-search work with?
It is written for Universal, as a SKILL.md file. Other agents that read the same format can often use it too.
Is media-search safe to use?
Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful. It is AGPL-3.0-licensed and scores 100/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is media-search still maintained?
The repository was last updated about 2 months ago, so media-search is actively maintained.

name: media-search description: Find and play music/videos via youtube_search + browser_navigate category: media version: 1.3.3 origin: aiden license: Apache-2.0 tags: music, video, youtube, play, song, listen, media, autoplay required_tools: [youtube_search, browser_navigate]

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) → browser_navigate (REQUIRED, must use a URL that came back from youtube_search). Keep navigation, interaction and Evidence inside the owned Browser Session. Navigation alone does not prove playback.

<execute_action> The user's request authorizes the requested media action, subject to the runtime's approval and browser policy. Use the actual available controlled browser tools. External open_url launching is unavailable inside durable Jobs and is not a recovery path.

After navigation, inspect the player. If playback has not started, use the observed play control through an admitted browser action. Verify observable player state (for example paused=false and advancing currentTime, when supported) before claiming playback. A screenshot or loaded watch URL alone does not establish that audio is playing.

Do not bypass approval, login, CAPTCHA, browser policy, or autoplay restrictions. If a supported action or verification is unavailable, report that exact limitation. Never claim success from an attempted action. </execute_action>

<selection_rule> How strict you should be about picking a candidate from youtube_search depends on how specific the user's query is.

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. </selection_rule>

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. browser_navigate — 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 inspect playback through supported browser tools and retain action Evidence before reporting. Searches alone do not fulfill a playback request. The required search/navigation sequence does not replace runtime approval or verification. Copy real result URLs; never invent video IDs.

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 browser_navigate.
  • ❌ Stopping after the searches and asking the user to confirm. The user's "play me a song" request is the consent — proceed to browser_navigate.
  • ❌ 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 browser_navigate 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:

| User said | Pass-1 query | |---|---| | "play me a popular song" | Billboard Hot 100 #1 this week | | "play me something upbeat" | top upbeat pop songs 2026 | | "play me a sad song" | top ballads 2026 spotify | | "play me kendrick lamar" | kendrick lamar most popular song 2025 | | "play me a hindi song" | bollywood top hindi songs this week | | "play me jazz" | jazz standards essential listening | | "play Despacito" | (skip pass 1 — title already given) | | "play Bohemian Rhapsody" | (skip pass 1 — title already given) |

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

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 you may use for the chosen video. Preserve existing URL provenance checks; do not invent or reconstruct a watch URL.

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 — Navigate, interact and verify

Call browser_navigate with the chosen result's URL: copy the url field verbatim. Inspect the resulting page with supported browser tools and operate the actual play control if needed. Reuse the same owned session and tab; do not launch duplicate tabs just because the player is slow.

Observe playback state after the action. If playback cannot be verified, report navigation as navigation and playback as unverified. Do not invent a successful play receipt or assume autoplay.

Step 4 — Report the observed result

Report the chosen title, URL, observed action and any limitation. Say "playing" only when playback was observed; otherwise say "opened" and explain what remains blocked or unverified. Keep Evidence linked to the actual action.

For fuzzy intent, briefly explain the selection. Do not present selection, navigation, or a user-facing browser window as proof that playback succeeded.

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. browser_navigate 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 browser_navigate 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-

Truncated for display — read the full file on GitHub.

Related Skills

View on GitHub
GitHub Stars843
CategoryOther
Updated2mo ago
Forks147

Languages

TypeScript

Trust signals

100/100

From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.

No cautions