How skim works
Skim is an AI tool by Credible AI that analyzes YouTube videos and news articles so you can get to the substance in minutes. Skim finds the key points, categorizes what is claimed and by whom, and plays the parts of a video that matter on YouTube's official player, so the creator keeps every view and ad. It runs on the web at skim.plus and as the Skim extension for Chrome.
skim is an AI consumption layer for YouTube videos and news articles. When you skim something, the AI identifies the parts that matter most, tags each statement, and surfaces what's claimed by whom — so you can get to the substance without losing the source.
What skim does in plain English
For YouTube videos, skim finds the segments where the substantive content lives — the actual interview answers, the analytical claims, the punchlines — and plays just those segments on YouTube's official player. Same listener-minutes, less filler, the creator gets full credit because every view, ad, and watch-time signal stays on YouTube.
For news articles, skim identifies the article's core takeaways (3 max), flags the bias direction, attributes claims to specific sources, and points out potential conflicts of interest. You read the takeaways in seconds, then decide whether the full article is worth your time.
How skim categorizes claims
Every meaningful statement skim analyzes gets categorized into one of three types:
- Claimed Fact — a statement asserting something as true. Skim notes who's making the claim and what sources support or contradict it.
- Opinion — a statement expressing a view or perspective without asserting empirical truth.
- Claim — a statement that's neither pure fact nor pure opinion; usually a mixed assertion that requires more context to evaluate.
The classification doesn't replace your judgment — it gives you a starting frame. The AI's reasoning behind each classification is shown next to every claim.
What "credibility" means in skim
skim assigns credibility scores (0-100%) across each piece of content. The score reflects:
- How well-sourced the content's claims are
- Whether the sources are independent or share an agenda
- The track record of the platform (channel or outlet) — though this is one input, not the whole story
- How transparent the content is about uncertainty
A high score doesn't mean "true." It means "well-sourced and transparent about its claims." A low score doesn't mean "false." It means "the AI didn't find strong supporting sources for the claims being made."
This is AI-generated analysis, not a verdict. Treat it as a structured starting frame, not a final answer.
What "bias" means in skim
skim labels content with a bias direction (e.g., "Accountability Focused", "Technological Optimism", "Skeptical of Institutional Claims"). The labels are descriptive, not political — we don't tag content as "left-wing" or "right-wing."
The label tells you the perspective lens through which the content is written, so you can:
- Read it knowing the angle going in
- Compare against content with different lenses on the same topic
- Decide for yourself whether the angle distorts the claims
The label is not a quality judgment. Many of the highest-quality publications have a clear bias direction. Bias and credibility are independent dimensions in skim.
Source attribution
Every claim skim surfaces is attributed to a specific source — the speaker on a video, the journalist or expert quoted in an article, the organization making a statement. We list the source's role and affiliation when known.
We don't publish numerical reputation scores against individual sources or named third-party brands as machine-readable data. We expose source attribution + role + affiliation; we don't expose numerical reputation scores.
How skim slices (Short / Medium / Relaxed) work
Every analyzed YouTube video on skim is available in three depth slices. Above the keypoint list there's a three-button control: Short, Medium, Relaxed. The page opens to Short by default; you can switch slices instantly mid-watch.
The depth slice you pick is a duration-budget filter applied at read time to the same underlying key-point set. The backend does not pre-compute three separate summaries. It computes one full key-point set per video, then filters that set per request based on the slice you asked for. All three slices cache as separate entries client-side, so swapping is instant.
The slice durations:
- Short — highest-impact moments. Roughly 20-30% of the original video duration for normal videos; roughly 20% for long videos (90 minutes and up — typically podcasts) so absolute skim time stays manageable.
- Medium — Short slice plus supporting context. Roughly 40-50% of the video duration; scaled down for long videos.
- Relaxed — comprehensive breakdown. Roughly 70% of the video duration. Every key moment skim flagged as substantive, with the full for-and-against analysis, source attribution, and credibility framing.
Long videos use lower percentages so a 2-hour podcast at "Short" lands at around 24 minutes — the cognitive contract is the same regardless of source length: pick the depth, get a skim that fits the time you have.
The slice filtering algorithm. When you pick a slice, skim sorts all key points chronologically, then applies a duration-budget filter:
- If the sum of all key-point durations fits within the slice's budget, return all key points (fast path).
- If not, iteratively remove the key point with the smallest temporal gap from its neighbors (the one whose removal disturbs the timeline least).
- If gap-removal can't reach the budget, switch to a greedy bin-packing pass that prefers retaining more (smaller) key points over fewer (larger) ones.
- If no subset fits, fall back to the single best-fitting key point.
The algorithm is deterministic — the same video at the same slice always produces the same key-point subset. The credibility, bias, originality, and depth scores are slice-independent. Slice choice only changes how many key points you see and which time ranges play, not the credibility framing.
Selective playback re-derives time ranges from whichever slice you picked. Hit Play All in Short and you watch only the highest-impact moments back-to-back on YouTube's official IFrame Player. Switch to Relaxed and Play All replays a deeper sequence. Creators get the watch time, ads play, and like/subscribe prompts stay intact regardless of which slice you choose.
Why three slices, not a slider. Three labeled depths are easier to remember and easier to share than a slider position. You can tell a friend "watch this on Medium" and they know what you mean. A slider position is meaningless without context.
Patent-pending. The selective-playback claim covers slice-based depth control as a user-elected duration-budget filter applied at read time to a fixed key-point set.
For the dedicated explainer with example durations and FAQs, see /features/skim-slices.
What skim isn't
- Not the truth police. We don't tell you what's true. We decompose what was said and surface who said it with what supporting evidence.
- Not a replacement for the original source. skim's analysis points back to the YouTube video or news article. Read the full source for context.
- Not unbiased. Our methodology is documented; our model has training-data biases like any AI; our human guardrails reflect editorial choices. We're transparent about all of this.
- Not finished. This methodology evolves as we ship new analytical capabilities and respond to user feedback.
How AI is involved
skim uses Gemini and Claude — frontier large language models — to perform the claim extraction, classification, and credibility analysis. The models are prompted with a structured analytical framework (the same one described above) and produce the per-piece outputs.
We don't use the AI to determine truth. We use it to identify and structure what's being claimed, by whom, with what evidence, with what consistency to other sources skim has analyzed. The structure is the value; the truth determination stays with you.
Patent + IP
skim's selective playback technology — the combination of automated keypoint identification, timestamp-anchored claim decomposition, and persistent shareable analysis URLs — is patent-pending. The selective playback architecture using YouTube's official IFrame Player API is what makes skim creator-aligned: views, ads, watch-time all stay on YouTube.