Lesson 04 · Jev 101

Use cases: decision slots Jev is built for

These patterns echo TypeSafe’s public use-case map and cookbooks. We don’t invent customer logos or unpublished wins. If you came from chat UIs, the “instead of prompting” lines are a translation — not the product definition.

Official category buckets come from TypeSafe’s use-case map. Rows tagged Educational are teaching hypotheses, not TypeSafe customer stories. Interactive stages for these slots live on Demo stages. Launch-week platforms and community builds — labeled, with URLs — live on Ecosystem. Chat-native habit pairs sit at the bottom and in the What’s the difference?.

Official task shapes

TypeSafe’s use-case map names ten decision shapes. Pick the shape first; the industry row is just a place to put it.

ShapeReach for it whenExamples
ClassificationOne known category should winIntent, topic, department, risk type
DetectionA probability that one property is presentSpam, fraud, urgency, jailbreaks
ScoringThe answer belongs on an ordered rubricSeverity, relevance, quality, frustration
RoutingA category selects the next code pathTool use, escalation, model routing
SearchFind items that match a natural-language querySemantic search, document discovery
RetrievalA workflow needs the most relevant contextRAG context, evidence lookup
RankingItems ordered by semantic relevance or qualitySearch results, recommendations
VerificationAn artifact checked for specific failure modesCitations, policy, tool-call errors
ML feature extractionA classical model needs semantic signalsPurchase intent, churn signals
Structured extractionKnown fields recovered from unstructured inputCandidate attributes, order fields

Official industry brainstorm

These bullets are TypeSafe’s map, shortened. They are idea prompts — not a customer roster.

Official map

Search and retrieval

  • Score query-to-candidate relevance
  • Rerank with pairwise comparisons
  • Select context for downstream AI
Official map

Scientific discovery

  • Screen papers against inclusion criteria
  • Label transcripts with predefined themes
  • Check whether citations support claims
Official map

Model routing

  • Choose which LLM gets the prompt
  • Classify intent and domain
  • Estimate difficulty and risk
Official map

LLM guardrails

  • Jailbreak / prompt injection
  • Policy and sensitive-data checks
  • Tool-call and response-quality failures
Official map

Semantic code linting

  • Team conventions as questions
  • Run in CI
  • Flag violations for review
Official map

Feature extraction

  • Probabilistic features from text
  • Combine with structured data
  • Autoresearch against held-out ground truth
Official map

Recruiting

  • Score evidence for required competencies
  • Match candidates to roles
  • Escalate uncertain cases
Official map

Lead generation

  • ICP fit
  • Buyer relevance and purchase intent
  • Prioritize and route leads
Official map

Customer support

  • Issue / product / intent
  • Urgency, frustration, churn, refund
  • Verify replies against policy
Official map

Insurance claims

  • FNOL classification
  • Complexity, missing info, fraud indicators
  • Straight-through vs specialist
Official map

Financial crime

  • Suspicious narratives and KYC
  • Entity match across messy names
  • Prioritize alerts
Official map

Legal and compliance

  • Contracts and filings
  • Missing clauses / prohibited claims
  • Escalate high-risk finds
Official map

E-commerce marketplaces

  • Normalize listings
  • Prohibited / counterfeit / review abuse
  • Rank and route uncertain listings
Official map

Moderation and trust

  • Company-specific criteria
  • Toxicity, spam, fraud, PII, opt-out
  • Allow / warn / review / block
Official map

Advertising

  • Brand safety
  • Regulatory claims
  • Ad-to-landing-page alignment
Official map

Gaming

  • Player reports and chat
  • Abuse / toxicity
  • Frustration, engagement, churn
Official map

Risk assessment

  • Incident and vendor notes → risk indicators
  • Classify risk types
  • Features for broader models
Official map

Demand forecasting

  • Intent and urgency from notes and reviews
  • Supply and competitive themes
  • Features beside time series
Official map

Graphs

  • Classify relationships and entity types
  • Detect contradictions
  • Hierarchical classification

Worked slots (this 101’s wording)

Background automation / smart if-statements

Classify, route, score, and gate inside product paths where a human should not sit in every loop.

Official theme Choice + Score + Noul

Support ticket triage

Instead of prompting ChatGPT to decide which queue owns a ticket, how urgent it is, and whether the customer is furious… Send ticket state plus Choice(queue), Score(urgency), and Noul(frustration or needs_human). Route in code.

Official theme Noul

Refund eligibility gate

Instead of prompting ChatGPT to decide whether we can auto-refund from a policy plus the message… Ask a Noul on eligibility (and optionally Score abuse risk). Auto-act only above your thresholds.

Official theme Noul composite

Spam / phishing quarantine

Instead of prompting ChatGPT to decide if this message is a phish in one chat reply… Compose several Nouls — credentials asked, sender mismatch, reward bait — then combine in code.

Official theme Choice + Score

Content moderation

Instead of prompting ChatGPT to moderate a community post by chatting with a policy… Choice and Score on toxicity / spam / fraud with severity plus confidence. Escalate the middle band.

Official theme Choice + Score

Insurance FNOL triage

Instead of prompting ChatGPT to read a first-notice-of-loss narrative and assign a desk… Choice on complexity, Score on missing information and fraud indicators, then route investigators.

Official theme Score

KYC / AML alert prioritization

Instead of prompting ChatGPT to rank alerts with a long chat summary… Score suspicious characteristics against your rubric; send high scores to investigators.

Official theme Noul + Choice

Contract clause detection

Instead of prompting ChatGPT to ask ChatGPT whether a contract is “fine”… Noul or Choice per missing clause or prohibited claim. Keep the legal memo for an LLM or a lawyer.

Official theme Choice + Noul

Marketplace listing QA

Instead of prompting ChatGPT to decide if a listing is prohibited or counterfeit in prose… Choice on category issues plus Nouls for counterfeit signals; normalize attributes in code.

Official theme Choice + Score

Brand safety for ads

Instead of prompting ChatGPT to ask a chat model if this creative is “on brand”… Choice / Score on suitability and prohibited claims. Threshold before the ad serves.

Official theme workflow

Invoice pay / hold / return

Instead of prompting ChatGPT to have a chat model narrate what to do with an invoice… Workflow-eval scenario: typed action over invoice state. Code executes pay, hold, or return.

Official theme workflow

Security incident contain / close

Instead of prompting ChatGPT to ask ChatGPT whether to page the on-call… Workflow-eval scenario: decide contain versus close from incident state, then let runbooks fire.

Official theme Choice + rules

Expense claim approval

Instead of prompting ChatGPT to approve meals, travel, and equipment from a chat essay… Choice against policy criteria; keep numeric checks in code. TypeSafe evals use this as a toy example.

Official theme Noul

Opt-out / DNC detection

Instead of prompting ChatGPT to have SDR chat decide whether a prospect said stop… Noul on the message. Honor opt-out in code — do not generate a follow-up when P(yes) is high.

Educational Score

Churn risk from ticket text

Instead of prompting ChatGPT to guess churn from a support thread in chat… Score risk from the message plus account state; alert CSMs above a threshold. Educational mapping.

Verify other AI

Judge prompts, drafts, citations, traces, and tool calls — the company highlights verify / guardrail patterns.

Official theme Noul + Score

LLM output guardrails

Instead of prompting ChatGPT to ask ChatGPT whether its own answer is fine… Take the draft or agent trace as state. Score policy, jailbreak risk, or “needs human.”

Official theme Choice

Citation support check

Instead of prompting ChatGPT to ask GPT to re-read the source and bless the claim… Choice: does this passage support the claim? Cookbook pattern.

Official theme workflow

Agent trace review urgency

Instead of prompting ChatGPT to skim a long agent log in a chat window… Workflow-eval observability: score whether a trace needs a human now.

Educational Noul

Podcast or report claim check

Instead of prompting ChatGPT to paste a transcript and ask “any problems?”… Noul per atomic claim against the transcript. Educational mapping, not a customer story.

Official theme Noul

Semantic code lint in CI

Instead of prompting ChatGPT to ask ChatGPT to comment on every PR… Nouls against conventions and guidelines. Fail the check on high-confidence violations.

Educational Choice

Shell command safety class

Instead of prompting ChatGPT to let an agent decide if a command is dangerous in prose… Choice: read_only / reversible / irreversible. Educational mapping for agent harnesses.

Make harnesses smarter

Routing, retrieval, tool picking, and feature extraction around LLMs and classical ML.

Official theme Score + Choice

Model router

Instead of prompting ChatGPT to send every request to a frontier chat model… Score difficulty or risk, then Choice the model. Cheap path versus frontier path lives in code.

Official theme Choice + Noul

Tool / skill picker

Instead of prompting ChatGPT to let an LLM narrate which tool to call… Choice over tool or skill names plus a Noul for “use a tool at all.”

Official theme Noul / Score

RAG passage filter

Instead of prompting ChatGPT to stuff every retrieved chunk into the prompt… Noul or Score relevance and injection risk before the LLM sees the passage.

Official theme Noul / Score

Semantic rerank

Instead of prompting ChatGPT to ask a chat model to reorder 200 results as a list… Score query–candidate relevance after BM25; sort in code.

Official theme Score

Resume screening

Instead of prompting ChatGPT to write a ChatGPT essay about each candidate… Score fit against job criteria; escalate uncertain cases to a human.

Official theme Score + Noul

Lead / ICP scoring

Instead of prompting ChatGPT to ask ChatGPT if this account is worth a call… Score industry fit and intent; Noul buyer relevance. Queue the SDR list in code.

Official theme Score

Hiring competency evidence

Instead of prompting ChatGPT to grade a work sample with a chat rubric in prose… Score against a predefined competency rubric. Keep the written feedback for an LLM.

Official theme Choice

Function-calling without LLM prose

Instead of prompting ChatGPT to hope tool-call JSON from a chat model parses… Map natural language onto typed function arguments with Choice over allowed values.

Official theme Score + Noul

Feature engineering for classical ML

Instead of prompting ChatGPT to hand-label text features for CatBoost overnight… Autoresearch cookbook: Score and Noul features from text, then train a classical model.

Official theme Score + Noul

Forecast enrichment signals

Instead of prompting ChatGPT to summarize tickets and hope a forecast model understands the essay… Extract intent and urgency features as Score / Noul, then feed numbers to the forecast.

Educational Score

PR risk matrix

Instead of prompting ChatGPT to ask ChatGPT how scary this diff is… Score security and complexity per file. Educational mapping for review routing.

Educational mixed

Inbox sponsor inquiry sort

Instead of prompting ChatGPT to triage a form inbox with one mega-prompt… Noul + Choice + Score on the form fields. Educational mapping.

Map-reduce over large corpora

Many independent questions over chunks, then aggregate in code — instead of one giant summary judgment.

Official theme fan-out

Map-reduce judgments over lots of text

Instead of prompting ChatGPT to prompt ChatGPT for one giant summary judgment… Decompose into many independent questions over chunks; aggregate in code (company theme).

Official theme Choice

Hierarchical patent / product class

Instead of prompting ChatGPT to ask a chat model to invent a taxonomy path… Beam search over Choice probabilities down a predefined tree.

Official theme Choice

Date-part extraction

Instead of prompting ChatGPT to ask which date is earlier in natural language… Choice over months and days, then assemble and compare in code. Jaggedness: do not do the math in the model.

Official theme Noul + Choice

Structure recovery / autoformat

Instead of prompting ChatGPT to ask an LLM to rewrite a messy blob as JSON… Stitch candidates, then Choice block types and Noul “is this a heading?” Official extraction theme.

Official theme Score

Entity alignment / merge decide

Instead of prompting ChatGPT to ask ChatGPT if two records are the same person… Score merge / leave / send-to-curator. Humans take the uncertain band.

Official theme Choice

Knowledge-graph relation classify

Instead of prompting ChatGPT to extract free-form relationships as sentences… Typed relation Choice over a closed relation set.

Real-time UX / games

System One–shaped calls where the UX needs roughly 100 ms-class decisions. Company demos and claims; do not overclaim game results.

Official theme Choice

Game / real-time control

Instead of prompting ChatGPT to prompt ChatGPT on every frame (too slow, too pricey)… Action Choice on structured game state. Company demos exist; do not invent win rates.

Official theme Choice

Wikiracing / high-cardinality links

Instead of prompting ChatGPT to ask a chat model to invent the next article title… Choice among many existing links without inventing new ones. Company demo theme; cardinality up to 255 (launch blog).

Official theme Choice

Smart-home intent → device action

Instead of prompting ChatGPT to chat with the house until something happens… Evaluate the user request as a Choice over allowed device actions (official demo theme).

If you want Jev inside an agent loop

Official how-to-build: System One is for AI-powered software, not agents. It does not generate code or choose its own next action. Code owns the workflow; Jev answers narrow, structured questions.

  • Questions in one request are independent. One answer is not hidden context for another. Need a later option set? Make a second call.
  • Keep permissions, argument validation, and side effects in application code. A typed Choice is not an ACL.
  • Vercel’s agent-control page says the same in platform words: Jev cannot execute a tool; Gateway evaluate needs the AI SDK — compatibility endpoints do not support it.Vercel ↗PRESS
  • eve’s approval: auto() helper asks Jev clear vscaution. Failed or invalid reviews pause for a human.eve KB ↗PRESS

If you want Jev to extract or align

Official habit: Jev does not invent a span and does not do calendar or ABV math. Code finds or proposes candidates; Jev picks; code copies or computes.

  • Find-then-pick: pre-parsed cookbook — regex over-finds, Choice options are those spans plusnone, code copies verbatim.
  • Dates: date cookbook — seven Choices read the parts; code builds thedate. Confidence is the weakest used part.
  • Alignment: entity alignment — Score levels are the outcomes; nearest level wins; no threshold to fit.

If you want Jev to rank a corpus

Official habit: cheap search first. Then one judgment per pair. Do not dump the shortlist into a single state.

  • Parallel questions: cookbook — all questions about one document ride together. Batching does not change the answers.
  • Rerank: cookbook — one Noul per query–candidate pair; sort by that noul.

If you want Jev next to an LLM

Official habit: Jev judges; the LLM writes. Screen both directions. Filter retrieved passages before they become prompt. Check citations after the LLM attaches them.

  • Guardrails: cookbook — Nouls + harm Score on the prompt and on the reply. Code routes pass / review / block / support.
  • RAG filter: cookbook — four Nouls per passage; first matching rule wins.
  • Citations: cookbook — string-match first; fabricated has no model confidence.
  • Find: cookbook — Choice ranks a line; exists Noul says whether it answers.

If you want Jev to pick UI

Official theme: Jev does not generate HTML, labels, or copy. Community recipes that call this “generative UI” still give Jev a prepared catalog and let code render the choice.

Anti-patterns (keep ChatGPT)

  • Write the customer email
  • Refactor this module
  • Brainstorm product names
  • Explain quantum computing like I’m 5
  • Draft a multi-step plan from a blank page
  • Generate image, audio, or video understanding

All 36 habit pairs

Pedagogical mappings from the corpus. Not A/B results.Educational = writer hypothesis for Luke-style teaching.

#ChatGPT / LLM habitJev / System One pattern
1Prompt: reply yes/no if this ticket is urgent.Ask a Noul on ticket state, then threshold on the probability.Theme
2Prompt: classify billing / tech / other as JSON.Ask a Choice with your label criteria, then branch on the choice plus confidence.Theme
3Prompt: rate severity 1–5 and explain.Ask a Score on your severity levels. If you need an explanation, call an LLM separately.Edu
4Hope JSON.parse works on the reply.Read typed answers. TypeSafe claims there is no prose-parse step.Theme
5Ask ChatGPT “are you sure?”Use the returned confidence or probability distribution as the control signal.Theme
6Write one mega-prompt that decides everything.Ask many atomic questions in one call; compose the policy in code.Theme
7Ask for chain-of-thought, then extract the decision.Skip generation. Ask the decision directly.Theme
8Fine-tune a classifier whenever labels change.Change Choice criteria per request. TypeSafe does not offer customer LoRA.Theme
9Route only with embeddings similarity.Optionally add a semantic Choice or Score rerank on candidates (cookbook theme).Theme
10Let an LLM agent choose the next tool in prose.Ask a Choice over tool names, then gate on confidence.Theme
11Run a second GPT call to check for prompt injection.Ask cheaper Noul / Score guardrail questions (company framing).Theme
12Ask the model to count keywords.Count in code. Use a Noul per item only when the question is semantic.Theme
13Ask which date is earlier.Extract parts as Choice questions, then compare in code.Theme
14Generate a refund email and decide the amount in one reply.Let Jev decide eligibility. Let an LLM write the email.Edu
15Moderate community posts in a chat window.Ask Choice / Score questions with policy criteria and a confidence gate.Theme
16Ask for a JSON confidence field.Use native probabilities. Do not invent a confidence schema in chat.Theme
17Retry the call when the JSON is invalid.TypeSafe claims schema-constrained answers cannot type-error.Theme
18Paste the entire CRM history into the prompt.Filter state. Send only the fields the decision needs.Theme
19One question per API round-trip.Use speculative fan-out: batch questions in one request.Theme
20Ask an LLM to rank 200 links by writing a list.Score or choose among candidates; sort in code.Theme
21Ask the model to invent categories.Predefine the options. Include an “other” label if you need an escape hatch.Theme
22Use chat inside a real-time game loop.Typed decisions at a ~100 ms company-claimed band for System One–shaped work.Theme
23Verify a citation by asking GPT to re-read the passage.Ask a citation-support Choice over the passage (cookbook pattern).Theme
24Route models with hand-written rules only.Ask Jev to Score difficulty or risk, then pick the model in code.Theme
25Screen resumes with a ChatGPT essay.Score against job criteria; escalate when confidence is low.Theme
26Detect churn risk in a free-text reply.Score or Noul on the message plus account state.Edu
27Ask “what should we do?” as an open-ended question.Offer a Choice over the actions your product is allowed to take.Theme
28Parse invoice fields via LLM JSON.Generate candidates with regex or an LLM, then let Choice select.Theme
29LLM-lint every PR with a comment.Run semantic lint Nouls or Scores in CI.Theme
30Hide rage-bait with a GPT classify call.Ask a per-item Noul or Choice in the feed pipeline.Edu
31Summarize first, then label the topic.Label on the full or filtered state with a Choice. Summaries stay with an LLM.Edu
32Ask whether two hex colors are similar.Convert colors in code. Ask Jev only semantic naming questions.Theme
33Write a double-negative prompt: “is it not false that…”Keep wording literal. Rewrite as a positive atomic question.Theme
34Trust a single chat “yes” for a wire transfer.Set a high-stakes threshold and require a human confirm.Theme
35Let an agent while-loop until it feels done.Code owns the loop. Jev only answers the judgments.Theme
36Ask Jev to write the customer reply.Don’t. Use a generative model. Jev is not trained to generate text.Theme