Deep research
Deep research on any topic, grounded in YouTube captions the user actually holds via asktube. Use when someone wants a researched answer built from real video sources — "research X", "what's the current thinking on X", "go deep on X", "build me a learning roadmap for X" — or wants to extend, deepen or reshape an earlier run in deep-research/. Mines the user's asktube library, tunes YouTube discovery from what it finds, curates a corpus, reads the transcripts, and writes the deliverable in whatever shape was asked for.From its SKILL.md
npx -y skills add asktube/skills --skill deep-researchAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its file declares
Copied from the file, not written here
The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
11.2 KB, as published. Nobody here has run it
deep-research
Deep research over YouTube, grounded in transcripts rather than recall. Discovery runs in the browser, against YouTube itself; the corpus, its metadata and its captions live in asktube, which fetches the captions on its own backend. Every claim in the deliverable traces to a video id and a real timestamp.
The request
Raw request: $ARGUMENTS
(If your agent doesn't substitute $ARGUMENTS, the raw request is whatever the user just asked for, verbatim.)
Parse it into four slots, restate them back in ≤4 lines, then start. Don't wait for confirmation.
| Slot | What it captures | If unspecified |
|---|---|---|
| TOPIC | the subject, who it's for, the decision or understanding it must enable | required — if the request names no topic, ask; this is the only blocking gap |
| PROCESS | how to research: depth, recency, language, sub-topics to force in or out, source bias (official vs practitioner), corpus size, into app playlist <name> to reuse one, "my library only" | phase defaults |
| OUTPUT | what shape the deliverable takes: "a 4-week study plan", "a one-pager", "a checklist I can paste into Linear", "just the shortlist" | a deep-research report |
| RESUME | "resume at research", "re-render as a one-pager", "add 10 more videos on X", a bare slug | fresh run |
Directives beat defaults. Directives never beat the invariants or the output contract.
English-first. asktube has no multi-language support yet, so discovery, vetting and the corpus
all default to English-language videos. A non-English source enters only when a brief question has
no English coverage at all — and then it is named as such in corpus.md and in coverage, not
quietly cited. A PROCESS directive can widen this; nothing else should.
Invariants
These hold no matter what PROCESS or OUTPUT say.
- Never fabricate. Stats, video ids, quotes, channel names, timestamps — all come from tools.
- Timestamps come from real caption
startvalues, converted to[mm:ss]. Check the conversion: seconds are 00–59. - Cite only what you read. A video may be cited only if its captions reached
doneand you read them. A video added but never captured is a gap, not a source. - Every non-obvious claim carries
(videoId, [mm:ss]). - Corroborate before quantifying. A specific number, benchmark result, price or date enters the deliverable only if it comes from a primary source (the author, the institution, or the vendor) or two independently-produced sources agree. Otherwise it stays in
notes/, named as single-sourced in the coverage section. Grounding proves a claim was said, not that it is true — an AI-narrated summary with a real timestamp is still an AI-narrated summary, and on YouTube that is now the likeliest way a wrong number reaches the page. - State coverage honestly — what the corpus doesn't cover, which videos were dropped and why, which recency and language policies applied and what each excluded. "Found nothing credible on X" is a real, publishable result.
- Before declaring a gap, rule out a tooling artifact. An empty harvest, a stalled capture queue, an English-only filter over a topic covered mostly elsewhere, and a genuinely thin topic all look identical. Re-run with different vocabulary and check
captions_statusbefore writing "little content exists on X" — and where the filter is what thinned it, say the corpus was language-filtered rather than reporting a thin topic. A small result set is not an empty one — on a niche topic, four good hits per query is the signal, not a failure. - Never write an empty file, section or directory. Create folders on first write, not up front.
- Sequence anything whose result you'd have to discard. A call is only independent of another if you would keep its output regardless of how the other resolves.
captions_prioritizeis not independent of the enrichment poll even though the two look unrelated: queue a video whosetitleis still null and it silently returnsaccepted: [], so the call succeeds and captures nothing. Batching pressure — from the harness or from your own instinct to save a round trip — is not a reason to collapse a real dependency. - Autonomous. At most one
AskUserQuestion, in Explore, before the brief is written — and only for a slot$ARGUMENTSleft open whose answer would change the run. After that there are no approval checkpoints: finish and report. The one exception is a precondition only the user can satisfy: a failed preflight. That halt comes before any phase, does not count against the one-question budget, and is not a question — name the missing requirement, give the setup steps, and stop. Do not downgrade to a library-only run.
Output contract
OUTPUT decides shape, format, length and file names. It does not relax these properties:
- Answers the brief — every question in the brief's Frame answered, or explicitly listed as unanswered with the reason.
- Grounded — every non-obvious claim carries its video id and timestamp.
- Each cited source carries a watch verdict —
watch in full/skim <range>/skip — the notes are enough. - A coverage section — corpus size, recency and language policies and what each excluded, drops with reasons, sub-topics nothing credible covered.
- Portable Markdown unless OUTPUT says otherwise: relative links,
- [ ]for anything checkable, no tool-specific syntax. - No layer restates another. If a document would mostly re-render one you already wrote, merge them instead.
When the requested shape is a roadmap or plan, it additionally must be dependency-ordered, name a single first move, and give each step a "done =" a reader can evaluate without asking you.
Output location
deep-research/<SLUG>/ in the current working directory, SLUG = short kebab-case of TOPIC. Nothing is scaffolded ahead of use:
deep-research/<SLUG>/
├── brief.md append-only — Frame (Explore) → Query plan (Explore) → Ledger (Research) → Amendments
├── corpus.md which videos, why each, what was captured, what was dropped, the apl- id
├── notes/ the evidence layer — your own structure, created on first write
├── report.md the deliverable (or whatever name the OUTPUT shape implies)
└── retro.md only when a mechanic misbehaved — this is how this skill gets fixed
Preflight — before the first phase, before any other tool call
Two MCP servers are required and neither ships with this skill. Check both up front. A failure here halts the run under Invariant 10 and does not spend the one question it allows. Per-client probe names and setup live in reference/tooling.md.
1. asktube — required, no fallback. Load the asktube tools, then call
me_get({ include: ["access", "usage"] }). It is read-only and cheap, and it distinguishes four
states a bare tool-existence check cannot:
| What you see | What it means | What to do |
|---|---|---|
| the tools don't exist | not configured | halt with the message below |
| the call fails to authenticate | the config is stale or the token expired | halt — send them back to the dashboard for a fresh config |
it returns, access.mcp is false | the account can't use MCP | halt and name that specifically |
| it returns clean | good | keep the usage library counters — they set expectations for Explore step 1 |
asktube MCP is not connected.
1. Sign up at https://asktube.xyz
2. Copy the MCP configuration from your dashboard (Settings → MCP)
3. Add it to this client and restart
Stopping — this skill has no library-only mode.
2. A browser — resolve BROWSER once, in this order. Discovery needs one; capture never does.
- Chrome DevTools MCP connected → probe
list_pages→BROWSER = devtools. This is the default in every client, Claude Code included. - Claude Code only:
claude-in-chromeconnected → probetabs_context_mcp→BROWSER = cic. There is no fallback anywhere else — outside Claude Code, no DevTools means no run. - Neither → halt, naming the
chrome-devtools-mcpsetup inreference/tooling.md.
The probe is not a formality. --autoConnect attaches to an already-running Chrome, so a server
that is configured with nothing to attach to fails right here — which is where you want it, rather
than at Explore step 5 where a dead browser is indistinguishable from an empty topic.
BROWSER is resolved once and never re-derived. The phases branch on it.
Phases
Read the phase file and follow it, then continue to the next. Each is written to be entered cold.
⚠️ A phase is not entered until its file has been read in this session. Carrying execution
momentum out of one phase and straight into the next one's tool calls is the most common way the
second phase file gets skipped entirely — the run then improvises a process that already exists in
writing. If you are about to capture captions and have not read research.md, stop and read it.
- Explore → phases/explore.md — mine the asktube library, clarify what's still open, write the brief, tune the YouTube search from what mining found, vet, curate into an app playlist.
- Research → phases/research.md — capture the captions, read them, write the deliverable.
The asktube wire contract and caption-capture protocol live in reference/asktube.md; the phases invoke it rather than restating it.
Tool names vary by client and by what the preflight resolved. Before entering a phase, read reference/tooling.md: it maps BROWSER = devtools (the default) and BROWSER = cic (the Claude Code fallback) onto the tools the phases name, and says how each client loads them. The browser is required for YouTube discovery and never for caption capture.
Live Chrome is one shared browser: browsing is serial — never fan sub-agents onto it. Sub-agents are for stateless work (reading captions, ranking captured text).
On RESUME, enter at the named phase and skip the earlier one.