The Useful Pattern Behind All Eight Demos
Jev is most useful when software already knows the possible answers and needs a fast judgment about which one fits. Across Matthew Berman's eight examples, the recurring jobs are ranking, filtering, selecting, and scoring. Jev does not write the web page, remove the advertisement, edit the video, or invent the emoji. The application prepares candidates and executes the result.
That boundary makes the demos more useful, not less. Many products contain thousands of tiny decisions that are too semantic for a simple rule and too frequent for a slow reasoning model: which inbox item matters, which sentence matches a search, which component belongs in a dashboard, or which transcript passage deserves review.
Watch Matthew Berman's Eight Jev Use Cases
Credit and disclosure: this guide analyzes 8 Jev Use Cases That Feel Like Cheating, published by Matthew Berman on 24 September 2026. The supplied transcript shaped the walkthrough; TypeSafe documentation and public project pages were checked separately. The Zapier section in the original video is sponsored.
How Jev Works
TypeSafe's documentation describes a decision model that receives state plus typed questions and returns structured answers. Three primitives cover most of the examples:
| Primitive | Returns | Example |
|---|---|---|
| Choice | One option plus probabilities and confidence | Choose the best UI component or email category |
| Score | A position on an ordered rubric | Rate priority, visual quality, or clip strength |
| Noul | A zero-to-one estimate for a statement | Estimate whether a message is spam |
TypeSafe recommends atomic, narrowly scoped judgments. If a decision depends on several factors, ask separate questions and combine the answers in code. The questions run independently against the same state, so a developer can evaluate category, urgency, spam probability, and reply likelihood without requesting a long chain of prose reasoning.
TypeSafe currently lists Jev at $42 per billion input tokens, equivalent to $0.042 per million, with free output tokens. Treat current pricing and vendor benchmark claims as product information that can change. Jev is also available through Vercel AI Gateway as typesafe-ai/jev.
The Eight-Use-Case Reality Check
| Demo | Jev decides | The application still does | Main risk |
|---|---|---|---|
| AI slop score | Scores 35 design judgments | Captures the page, describes it with another model, and writes the verdict | Confusing a style score with proof of authorship |
| Unclutter | Classifies candidate page elements as clutter | Finds DOM candidates, hides elements, and saves site rules | Hiding content without blocking tracking or consent flows |
| Email priority | Scores category, urgency, spam, and reply need | Fetches messages, applies labels, and triggers actions | Archiving or answering on an uncalibrated score |
| Semantic find | Ranks page passages against a query | Extracts text and highlights the selected sentence | Missing meaning that depends on layout or images |
| Real-time UI | Selects components, properties, and placement options | Supplies a catalog, validates a spec, and renders it | Presenting selection as unrestricted code generation |
| Video clipping | Scores transcript segments for relevance or impact | Transcribes, sets boundaries, cuts, captions, and exports | A strong quote with bad context or visual timing |
| Color palettes | Ranks candidate colors against a description | Supplies the palette and renders the result | Ignoring contrast, accessibility, or brand constraints |
| Emoji selection | Chooses from a bounded emoji set | Supplies candidates and inserts the character | Tone, cultural, and platform-rendering mistakes |
1-2. AI Slop Scoring and Browser Cleanup
The AI Slop Detector measures style, not provenance
The Made with Jev detector runs a multi-part pipeline. A headless browser captures the site, DeepSeek describes the screenshot, Jev evaluates 35 design judgments, and another model turns the results into a readable verdict. The public Anthropic report shown in the video scores the site at 26%.
The tool's own explanation is the important caveat: a high score does not prove that AI built the site. It identifies visual patterns associated with generic choices or missing design direction. That makes it a design critique aid, not an authorship detector.
Unclutter is reversible DOM classification
Kitze's open-source Unclutter extension extracts a bounded list of page elements, asks Jev which ones are clutter, and hides high-confidence matches. It stores reusable rules by page template, so previously learned layouts can be cleaned locally without another model request.
The repository also documents the limits. Hidden ads may still load and track, the extension does not click cookie-consent buttons, uncertain elements stay visible, and extracted snippets may contain sensitive text even though raw HTML, form values, and full URLs are not sent. The right product behavior is reversible hiding with an undo path, not silent deletion.
3-4. Email Priority and Smarter Find-in-Page
The email demo reports that Jev classified 100 messages in under half a second. The useful architecture is to score several dimensions independently: category, priority, spam likelihood, and whether a human reply is needed. Those numbers can reorder an inbox without authorizing irreversible action.
Start with recommendations, labels, and review queues. Build a human-reviewed test set, measure precision for the high-priority bucket, and choose different thresholds for different consequences. A false positive that adds a star is cheap; a false positive that sends a reply or deletes a message is not.
The Needle semantic find demo uses the same ranking pattern on page text. Instead of matching the exact typed phrase, Jev scores passages for meaning and the extension highlights the strongest sentence already present. It does not generate an answer, and it does not need an embedding index for the current page.
5-6. Real-Time Interfaces and Instant Video Clipping
The experimental json-render integration is a clean example of constrained generation. The developer supplies atomic components, allowed properties, state bindings, and actions. Jev decides which pieces fit; ordinary code assembles and validates the specification; the renderer displays it. Jev does not invent missing prose or execute actions.
This pattern can make dashboards adapt quickly, but the application must own accessibility, layout constraints, data validity, and permissions. The json-render documentation labels the integration experimental, so pin versions and expect API changes.
The Clipfast demo applies Jev to long-form video. The likely pipeline is transcript first, candidate segments second, Jev ranking third, and media rendering last. Matthew reports that a video longer than 90 minutes was processed into candidate clips in under two seconds, but the public demo does not document the full pipeline or provide a quality benchmark.
Judge a clipping system by precision among the top candidates, missed highlights, duplicate rate, boundary corrections, editor review time, and performance after publishing. For a deeper implementation pattern, see our Jev auto-clipping workflow.
7-8. Color Palettes and Emoji Selection
Color and emoji selection are small decisions with immediate feedback. An application can provide a finite palette or emoji set, then ask Jev which candidates best match a phrase, mood, object, or intent. The result can feel generative because it responds to natural language in real time, but the model is ranking a prepared catalog.
These are good first experiments because mistakes are visible and usually reversible. Production versions still need deterministic rules: contrast checks for text and controls, brand-locked colors, skin-tone and platform-rendering tests, locale review, and an option to return no suitable result.
The Sponsored Zapier Example
The original video contains a sponsored Zapier segment. Matthew shows Jev as a Zapier action inside a workflow connected to other apps and gives a calendar example: inspect a meeting request, estimate whether it is important, and accept or decline based on a threshold.
The safer first version stops before the final action. Let Jev score the meeting and explain the routing fields to the workflow, then ask the user to approve acceptance or rejection. After a labeled trial proves the thresholds, low-risk cases can be automated while conflicts, external guests, ambiguous requests, and high-value meetings remain reviewable.
Where Jev Fits and Where It Does Not
| Good fit | Poor fit |
|---|---|
| Many repeated, narrow decisions | Open-ended writing or conversation |
| A fixed candidate set or explicit rubric | A task requiring facts absent from the supplied state |
| Latency matters inside a live interface | Long multi-step planning or deep research |
| Errors are reversible or reviewable | Irreversible actions with no approval gate |
| You can build a labeled evaluation set | No ground truth, monitoring, or fallback path |
Vercel advises calibrating Jev's probabilities and confidence on labeled examples. That matters more than a striking demo. A score is useful only when your team knows what error rate it produces on your own data and what the system should do when confidence is low.
A Seven-Step Build Checklist
- Define the state: send only the evidence needed for the decision.
- Constrain the output: use explicit choices, an ordered rubric, or a Noul statement.
- Split complex judgments: score urgency, relevance, risk, and quality separately.
- Create labeled examples: include ordinary, ambiguous, adversarial, and out-of-scope cases.
- Calibrate thresholds: tie each confidence band to a review, fallback, or reversible action.
- Log the decision: preserve input version, question version, answer, confidence, action, and reviewer outcome.
- Measure the whole system: track accepted-result accuracy, latency, cost, human review time, and recovery from failure.
Video Chapters
| Time | Topic | Time | Topic |
|---|---|---|---|
| 00:00 | Introduction | 06:31 | Prioritizing emails |
| 00:30 | What is Jev? | 07:03 | Smarter find-in-page |
| 03:29 | Detecting AI slop | 07:49 | Building web pages in real time |
| 04:14 | Blocking ads and clutter | 08:34 | Instantly clipping videos |
| 05:17 | Sponsored Zapier segment | 09:16 | Generating color palettes |
| 09:52 | Finding the right emoji |
Sources and Original Demos
- Matthew Berman: 8 Jev Use Cases That Feel Like Cheating and Matthew on X
- TypeSafe AI, Jev documentation, and Vercel AI Gateway access
- AI Slop Detector demo and live detector
- Unclutter demo and source repository
- email priority demo and semantic find demo
- real-time UI demo and json-render Jev documentation
- Clipfast demo, color-palette demo, and emoji demo
YouTube lists the video's publication date as 24 September 2026. This article was reviewed on 25 September 2026. Model access, prices, integrations, and experimental APIs can change.