Structuring YouTube Search Results
This skill does not have web access, API access, or any ability to query YouTube, Archivarix, or any archive service. It cannot discover videos, verify whether a video still exists, or produce view counts/durations/IDs that weren't given to it.
This skill only reformats data the user has already provided (pasted scrape output, a messy list, a CSV, etc.).
Mandatory rule: If the user asks to "find," "search," or "look up" videos without pasting source data, respond that this skill cannot perform live searches and ask the user to paste the raw results they want structured. Never invent video IDs, view counts, durations, upload dates, or Live/Deleted status. If a field is missing or unclear in the source text, output unknown — do not guess or fabricate a plausible-looking value.
Given messy pasted text like:
Turning a Weaving DraftLiveTurning a Weaving DraftActon Creative·h2_ByZnK5cI·Watch on YouTube·Log in to download subtitles·Summarize9:34·806 views
Parse and normalize it into:
Title: Turning a Weaving Draft
Channel: Acton Creative
Video ID: h2_ByZnK5cI
Status: Live
Duration: 9:34
Views: 806
Has Subtitles: unknown (login required to confirm)
Watch URL: https://www.youtube.com/watch?v=h2_ByZnK5cI
Every value above traces directly back to a substring in the input. Nothing was added.
Progress:
- Step 1: Confirm source data was actually provided (paste/file); if not, stop and ask for it
- Step 2: Split the raw blob into per-video chunks
- Step 3: Extract fields from each chunk using consistent pattern rules
- Step 4: Mark any field that can't be confidently extracted as
unknown - Step 5: Apply requested filters (status, has-video, has-subtitles)
- Step 6: Apply requested sort order
- Step 7: Render normalized table/list, citing only given data
Step 1 — Verify input exists. If the user's message contains no pasted search results, do not proceed to "search" — this skill has no fetch capability. Reply: "I can't run live searches. Paste the raw results (e.g., copied Archivarix Tube Search output) and I'll structure them."
Step 2 — Split into chunks. Archivarix-style dumps repeat the title twice (once as a heading, once in a metadata line) and concatenate fields with ·. Split on recurring patterns like a duration pattern (\d+:\d+) followed by a views pattern ([\d,]+ views), which reliably mark the end of one record and the start of the next.
Step 3 — Extract fields per chunk:
title: first text segment before the status word (Live/Deleted) or before the repeated titlechannel: text segment immediately before the·videoID·, typically ending in·video_id: the YouTube ID pattern — exactly 11 chars, alphanumeric plus-/_, isolated between·delimitersstatus: literal wordLiveorDeletedas it appears in source — never inferredduration:\d+:\d+patternviews: number immediately preceding the literal wordviewshas_subtitles: true only if source explicitly says subtitles are downloadable; "Log in to download subtitles" means unknown, not falsewatch_url: construct only ashttps://www.youtube.com/watch?v={video_id}— only valid if video_id was extracted, never fabricatedarchive_url: only if source text contains "View in Archive" or similar explicit link indicator; otherwisenone
Step 4 — Flag uncertainty. If a chunk doesn't cleanly match the expected pattern (e.g., missing duration or views), output the extracted fields plus extraction_notes: field X not found in source.
Step 5 — Filter. Apply only filters the user explicitly requested (status=Live/Deleted, has_video, has_subtitles). Filtering happens on normalized fields, never on guessed values.
Step 6 — Sort. Default: preserve original input order (which is usually "most relevant" from the source tool — this skill has no way to compute relevance itself, so it trusts the source order unless told to re-sort). If asked for newest/oldest/longest/shortest and the needed field is unknown for some rows, place those rows last and say so.
Step 7 — Render. Output a table or list using only normalized fields, each traceable to the input.
Example 1 — Raw blob to normalized table
Input (pasted by user, verbatim from source):
Reading a Weaving DraftLiveReading a Weaving DraftJane Stafford Textiles·qD6bKAhlDuI·Watch on YouTube·Log in to download subtitles·Summarize19:32·96,909 views
Output:
| Title | Channel | Video ID | Status | Duration | Views | Subtitles | Watch URL |
|---|---|---|---|---|---|---|---|
| Reading a Weaving Draft | Jane Stafford Textiles | qD6bKAhlDuI | Live | 19:32 | 96,909 | unknown (login required) | https://www.youtube.com/watch?v=qD6bKAhlDuI |
Example 2 — Filtering request with no matching data
Input: "From this pasted list, show me only the Deleted ones." (list contains 22 entries, all marked Live in source text)
Output:
0 of 22 entries are marked "Deleted" in the source data you provided. All 22 are marked "Live". No deleted videos to show.
(Does not fabricate a deleted entry to satisfy the request.)
Example 3 — User asks for a search with no data provided
Input: "Find deleted YouTube videos about weaving drafts."
Output:
I don't have web/API access to search YouTube or any archive service, so I can't look this up live.
If you have search results already (e.g., copied from Archivarix Tube Search or similar), paste them here and I'll parse, filter, and sort them for you.
Example 4 — Partial/ambiguous chunk
Input:
Weaving the pattern "Sharp toother with a turn" with four tubes. Part 2Live·•●♥๑ஐ♥ღ Newspaper weaving ღ♥ஐ๑♥●•··VfPVForrXYw·Watch on YouTube·Log in to download subtitles·Summarize3:49·2,214 views
Output:
Title: Weaving the pattern "Sharp toother with a turn" with four tubes. Part 2
Channel: •●♥๑ஐ♥ღ Newspaper weaving ღ♥ஐ๑♥●• (unusual characters preserved as-is from source)
Video ID: VfPVForrXYw
Status: Live
Duration: 3:49
Views: 2,214
Subtitles: unknown (login required)
Watch URL: https://www.youtube.com/watch?v=VfPVForrXYw
extraction_notes: channel name contains decorative symbols; preserved verbatim rather than cleaned, to avoid misrepresenting the source
- Always show a direct line from source text to extracted field; if asked, be able to point to exactly what substring produced each value.
- Preserve exact numbers, spellings, and symbols from the source — don't normalize channel names, round view counts, or "fix" apparent typos.
- Treat "Log in to download subtitles" as
unknown, notfalse— the source didn't confirm absence, only that confirmation requires login. - When re-sorting, only use fields that were actually extracted; never sort by a guessed duration or view count.
- If the same video ID appears twice in a dump (duplicate listing), de-duplicate and note it rather than silently dropping or silently doubling.
- Keep "Deleted" entries clearly flagged and surfaced first if the user's goal is clearly "recover lost videos" — but only based on literal status text from source, never inferred from absence of view count or other proxy signals.
- Do not treat this skill as a live search tool — it has no fetch/browse/API capability. If no data is pasted, say so instead of producing a plausible-looking fake list.
- Do not infer "Deleted" status from a video having low views, no subtitles, or short duration — status is only ever read verbatim from the source text.
- Do not invent an
upload_datefield — Archivarix-style dumps in this format don't include one; always markunknown. - Do not fabricate a
View in ArchiveURL — only include it if the literal text "View in Archive" (or equivalent) appears in the source for that entry. - Do not silently "clean up" messy channel names or titles (e.g., stripping emoji/symbols) — this can misidentify the actual source and breaks traceability.
- Do not apply relevance-sorting logic Claude invents itself; without the original ranking algorithm, claiming to compute "relevance" is just fabrication dressed as methodology.