Access & methodology
InsightDesk’s first gate is understanding actual YouTube content. A successful metadata check is not a successful analysis.
The latest authorized real trial succeeded at token counting, then its generation request was rejected with HTTP 400 because our output schema was too complex. Its exact error response was retained encrypted in private durable storage before parsing. No generated content was returned to review. This is our request-schema compatibility failure, not a reason to reconnect the key or add funds. Read the actual trial result.
What a check does today
- Validates a link to a single video and removes unrelated URL parameters.
- Requests the canonical URL from YouTube’s public oEmbed endpoint, with an eight-second timeout and no redirects.
- Displays only the returned title and channel as metadata. It does not render returned HTML, load a player, fetch thumbnails or scrape the watch page.
- Reports input-specific URL and metadata outcomes separately from general analysis availability. No spoken facts, timestamps, images or summaries are invented.
- Measures server check time. AI requests are zero and AI cost is $0; hosting cost and future per-video cost are unknown.
The proposed audiovisual route
Google’s Gemini video-understanding documentation supports passing a public YouTube URL directly as video input. This route is intended to process audio and visual content, not just metadata. It avoids our downloading third-party streams or using unofficial transcript services.
The YouTube-URL feature is documented as a preview. Private and unlisted videos are not supported by this route. Static processing samples frames at one frame per second by default; fast UI changes, small code and chart labels can be missed. Long-video reliability, no-subtitle coverage, visual precision and cost still require actual tests. Public availability alone does not establish permission for every use.
Verification before the next test
- Provider: Google Gemini Developer API, through a Google AI Studio project eligible in the account’s region.
- Credential: a server-only GEMINI_API_KEY with permission to call the Generative Language API. No YouTube account login, YouTube download permission or Gmail scopes are needed for the documented public-URL route. Never paste the key into the video field or email it to us; use secure environment-key settings.
- Confirm acceptance of Google’s applicable API terms and the permitted use of the selected public sources. If the route does not permit a target, report it as inaccessible rather than bypassing the restriction.
- The connected project is paid-tier, confirmed by the account owner; the single paid public-video trial replaced the earlier unpaid-only permission. Internal developer testing is separate from making an API client available to users; access itself must be from an available region. Google’s current API terms require Paid Services for API clients made available to users in the EEA, Switzerland or UK. Unpaid-service inputs and outputs may be used for improvement and human review, with regional exceptions. Use only permitted public, non-sensitive sources; do not submit confidential or personal material.
- No subscription, billing activation or top-up has been made by InsightDesk. Model availability, token metering, project quotas and current provider prices must be checked before activation. Permission covered one bounded paid operator trial on the specific public video, not other paid videos or a public generation endpoint. The first run failed at preflight. The founder then explicitly authorized one corrected attempt, with a $1 new-attempt ceiling and $5 cumulative ceiling including the earlier reserve. That attempt returned generation output but failed local evidence validation. The founder then explicitly authorized one retained-response real trial. That trial made one successful count request and one generation request rejected for schema complexity, with no content output. All three permissions are now consumed; another paid attempt requires new approval and reviewed cost bounds.
The supplied key does not silently enable analysis or paid requests in this release. Activation requires a bounded, access-controlled provider test and verification of actual results. This public diagnostic does not keep accounts or private reports.
Routes not treated as proof
| Route | Why it does not pass the gate |
|---|---|
| YouTube oEmbed / video metadata | Identifies a source, not what was said or shown. |
| YouTube Data API captions | Caption download requires permission to edit the video and OAuth. It is not general third-party transcript access and does not provide visual evidence. |
| Watch-page scraping / stream download | Not used. YouTube’s terms restrict automated access, downloading and circumvention unless expressly authorized or permitted. Technical availability is not permission. |
| Creator-authorized media upload | Could evaluate a processing pipeline, but would not prove automatic access to third-party YouTube videos. There is no upload test in this release. |
Acceptance sample
Annalisa supplied three channels and one specific video. The channel pages identify them as Chase AI (@Chase-H-AI); Adrian Twarog (@AdrianTwarog); Duncan Rogoff | Learn Claude Code (@duncanrogoff). The video titles and channel names below were verified using YouTube oEmbed. Player duration is a manual observation, not content acquired by the application.
- Claude Just Changed: The 6 Big Updates You Need to Know — Duncan Rogoff | Learn Claude Code. First acceptance source. Player showed 7:49; corrected generation returned but local evidence validation failed; no verified report. Subtitle type is unverified.
- This Experimental Feature Is Claude Code's BIGGEST Upgrade of 2026 — Chase AI. Selected from Chase AI’s channel. Player showed 8:01; visual content and subtitle type still require verification.
- Import Figma Web Designs in 1-click using Bolt.new — Adrian Twarog. Selected from Adrian’s channel. Player showed 5:52. Title suggests a UI demonstration; actual procedure and visuals are not yet evaluated.
- Adobe XD to Wordpress — Adrian Twarog. Player showed 4:45:24 (channel card 4:45:25). A short excerpt would not establish whole-video reliability. Subtitle type is unverified.
Open sample gaps: no verified no-creator-subtitle case yet. A caption button alone does not distinguish creator subtitles from automatic captions, and pre-roll ads can change player controls. The visual-demo candidate needs actual screen checks. The 4h45m tutorial establishes a long-form source, not successful long-form processing. No channel has passed automated audiovisual analysis.
Bounded provider trial
An operator-only runner is prepared in the repository, separate from this public diagnostic. It sends one selected video URL directly to Gemini, first requests a token count, and permits at most one content-generation request per run. It requires a server-only credential, account/region/terms review, a selected compatible model, reviewed token limits and pricing, and explicit trial activation. It makes no automatic retries or source downloads.
The first full-video request is Annalisa’s 7:49 Duncan Rogoff video. Longer requests require a reviewed duration limit; an excerpt remains labelled partial. Provider output is checked for bounded timestamped speech and visual evidence, with claims separated from interpretation. Any valid future response would still need manual source review before acceptance as faithful analysis. Usage estimates are not a verified invoice or a provider-enforced hard spending cap. Credential and model-catalogue checks are not video processing. The earlier preflight returned HTTP 404 in 122 ms. The corrected attempt made one successful token-count request and one completed generation request, taking 9.514 seconds in total. Estimated generation cost from reported usage was $0.0160726; preflight charges and the actual invoice remain unknown. The rejected output was lost, so no independent content-quality check could be completed. The third, separately approved trial failed at provider schema validation, not local evidence review: one count (48,567 tokens), one HTTP 400 generation request, no output usage. Its exact response is durably retained privately. Read the latest attempt, timing, reserves and evidence gap; earlier attempts.
Quality and coverage checks still required
- Run the acceptance sample, starting with Annalisa’s specific video.
- Confirm access on long-form videos, a verified video without creator subtitles, software/code demonstrations and charts.
- Compare extracted facts and timestamps with the original. Separate source claims from interpretation. Preserve substantive caveats, costs and procedures when removing filler and sponsorship.
- Record actual covered ranges and visual sampling limits. “URL accepted” must never be reported as “100% analyzed.”
- Treat every transcript, caption and on-screen instruction as untrusted source data. Do not give the analysis model tools, secrets or private account context. Source text must not initiate actions.
- Measure acquisition and analysis latency, tokens and per-item costs where available; show unknown where the provider cannot expose a reliable measurement.
What is kept
The metadata-check application does not write visitor-submitted URLs or reports to its database, browser storage or a public history. The operator trials’ redacted status and usage for the selected public video are published separately, without credentials, private account details or unvalidated generated text. Rejected output from the earlier corrected run was not retained. The third trial’s bounded, credential-redacted original error is encrypted in private database storage with a separate environment key, for operator-only offline diagnosis; no public route reads it. It is retained while compatibility diagnosis remains open, without an automatic deletion period set. No substantive candidate was returned. The current metadata report lives only in the open tab’s memory. Clear test or reload removes it; a downloaded JSON file stays on your device. The canonical URL is sent to YouTube for the metadata request. Hosting infrastructure can process requests under its own operational policies. See Privacy.
Questions or a known no-subtitle target: insightdesk@brainer.one.