Parker Prompts' Jev tutorial is more useful than its “lead generation forever” title suggests. It shows four small applications built with Claude Code: a support inbox, a creative prompt checker, a podcast clip finder, and a job-research grid. Jev does not build the apps or write the emails. It makes narrow, typed decisions inside the apps; Claude Code supplies the interface and integration code.
Jev is a fit when software needs to make the same bounded choice many times: choose a queue, score a 30-second transcript segment, or flag a missing prompt detail. The application still owns the data, thresholds, retries, permissions, and final action. Parker's Job Radar ranks public company profiles against a resume; it does not discover verified open jobs or qualified sales leads.
Watch Parker's Tutorial
Credit: The video is published by Instantly; the supplied Jev Prompt Pack is branded Parker Prompts and documents the demo apps, thresholds, and costs. The five copyable prompts below are shortened adaptations of that 11-page PDF, with additional safeguards. For its complete wording and notes, use the original pack linked from the video description. The shortened Google Drive URL supplied with this request could not be verified as a complete destination.
One Model for Decisions, One Agent for Building
TypeSafe's documentation defines three Jev question types: Choice selects an option, Score applies a rubric, and Noul returns a probability for a yes/no statement. One request may contain several separate questions over the same state. The model returns typed results and probabilities rather than long prose. Vercel's Jev guide describes the same boundary: use Jev to classify, route, and score; keep policy and side effects in application code.
That split matters. In Parker's examples, Claude Code writes the local app once. Jev repeatedly evaluates each ticket, prompt, transcript window, or company profile. Code then applies thresholds or weighted formulas. A refund, a published clip, a customer message, or a job application still needs its own authorization and review. A probability is a model output, not a measured guarantee that the classification is right.
Set Up the Smallest Possible Test
- Check current access. Read TypeSafe's docs and Vercel's integration guide. Parker's pack describes direct TypeSafe sign-ups as paused when he recorded; availability can change.
- Choose one route. Use direct TypeSafe access if available, or Vercel AI Gateway. The Gateway model ID is
typesafe-ai/jev; its TypeSafe-compatible client route useshttps://ai-gateway.vercel.sh/typesafe. Verify current SDK examples before coding. - Control spend and secrets. Store the API key server-side in a local
.envignored by Git. Review Gateway budgets and current model pricing; a budget is a soft cap and is not itself prepaid credit. - Make one fictional test call. Ask Jev to route a made-up failed payment integration to billing, technical, or sales. Inspect all probabilities, latency, and the provider's billed usage before running a batch.
The official TypeSafe agent skill can help a coding agent find current patterns. Review its source and installation scope before adding it. The pack suggests a shared Gateway key and notes a 30-requests-per-minute cap in Parker's own account; check your plan's live limits and do not run every demo at once.
What the Four Apps Actually Demonstrate
Support inbox
Jev's job and observed run
For each of 200 fictional coffee-brand tickets, choose a team, score urgency, and flag refund intent. Low-confidence or “other” cases go to human review.
The pack reports about half a cent at its stated list price. On Parker's sample, 171 of 176 clear emails reached the intended column, while 18 of 24 vague ones reached a person. These are creator-reported results, not an independent test.
Production boundary
Set a threshold using hand-labeled real examples. Do not auto-refund or send replies based on classification alone.
Prompt checker
Jev's job and observed run
Detect image or video prompt type and answer precise yes/no checks for camera, light, setting, action, format, and other relevant details.
Parker's PDF says roughly one-sixtieth of a cent per check in his run. The “ready” meter is calculated by his app's weights, not proof the generated visual will be good.
Production boundary
Debounce typing, make custom checks literal, and ask a person to judge the resulting image or video.
Short-video finder
Jev's job and observed run
Score timestamped transcript chunks on hook, standalone clarity, payoff, and energy; code combines the four answers into a rank.
The pack describes 212 chunks from an almost two-hour podcast, under a cent of Jev list-price input in that run. Obtaining the transcript, processing time, and editing are separate costs.
Production boundary
Transcripts miss visual context and tone. Check rights and watch the selected segment before publishing.
Job Radar
Jev's job and observed run
First rank short public YC company descriptions against a resume; then recheck the strongest 100 on skills, plausible need, and location.
The PDF records 4,322 active profiles and about four cents at its list-price calculation. That count is a snapshot and the data source is a community-maintained YC dataset, not a verified vacancy feed.
Production boundary
A strong fit does not mean the company is hiring. Check careers pages and keep a private resume under explicit data-use consent.
Cost reading: Parker's pack uses a stated Jev list price of $0.042 per million input tokens and separately recommends showing Gateway-billed cost. Live Gateway pricing may differ. Comparisons with other models also depend on prompts, token accounting, concurrency, and quality, so do not treat the video's ratio as a universal benchmark.
Five Copy-Ready Prompts From the Workflow
These are concise, safety-aware adaptations of Parker's five PDF prompts, not a verbatim republication of his pack. The original has fuller UI specifications, sample data distributions, test notes, and API details. Start with setup; use one build prompt at a time. Each prompt works in a general coding assistant, though building the apps requires access to a development environment and a Jev provider.
1. Set up and test Jev
A guided first call without exposing your key.
2. Build a support triage prototype
Classify fictional tickets with a human-review lane.
3. Check image and video prompts
Turn a broad quality score into explicit checks.
4. Find candidate short-video moments
Rank transcript chunks, then inspect the footage.
5. Rank companies for job research
A company-fit shortlist, not a verified vacancy list.
Find a Business Worth Building Around It
For a paid offer, start with a buyer's recurring decision and a measured error rate. The prompt below uses the video's method without assuming Jev is necessary or already installed.
Find the decision bottleneck
Test a narrow service against a rules-only baseline.
Sources and Limits
Original work: Parker Prompts' tutorial and his supplied 11-page Jev Prompt Pack. The supplied Google Drive address was shortened, so the original video description is the reliable route to Parker's download. Primary documentation: TypeSafe introduction, TypeSafe skill, Vercel integrations, budget documentation, and current pricing. All demo counts, costs, and quality figures are attributed to Parker; no app was independently rerun for this article.