USE CASES

Use cases

Catalog of decision shapes you can paste into software — not chatbots. Each card lists who it’s for, steps, expected effect, and sources. Metrics marked author-reported or cookbook-reported. Calibrate on your data.

Start Guides Hands-on recipes

55 catalog cards
Support · Agents · Retrieval · Extraction · Computer use · Ops · Automation
Metrics: author-reported / cookbook-reported
Calibrate thresholds on your data

Agents ✦ 90 · outline

Agent harness supervisor

Who: Teams wrapping Codex/Claude Code/Pi-style agents. Steps: On turn/tool boundary ask parallel questions — skill Choice, action allow/ask/deny, stuck/slop/done Nouls → harness steers or blocks. Expected effect: cheap semantic supervisor around the agent (community cluster).

Get the prompt →

Agents ✦ 75 · outline

Commit / change gate

Who: Dev workflow. Steps: Diff summary → Choice allow/ask + Nouls → pre-commit. Expected effect: Local quality gate.

Get the prompt →

Agents ✦ 73 · outline

Independent worker supervisor

Who: Multi-agent runners. Steps: Worker trace → Nouls complete/tested/off-track/stuck → supervisor acts. Expected effect: Architecture experiment — label as such.

Get the prompt →

Agents ✦ 72 · outline

NL → typed function args

Who: Tooling / trading demos. Steps: Map NL to Choice function + closed-set args → code calls. Expected effect: Confidence-aware tool binding.

Get the prompt →

Agents ✦ 77 · outline

Skill suggestion (large catalog)

Who: Agent platforms. Steps: Rank skills + Noul needs-skill-at-all → second pass top-3 → inject one name. Expected effect: At most one skill line.

Get the prompt →

Agents ✦ 76 · outline

Staged code review routing

Who: Devtools. Steps: Risk Nouls → file profiles → evidence Choice → severity → route. Expected effect: Structured review without one giant prompt.

Get the prompt →

Agents ✦ 94 · outline

Tool-call / jailbreak guardrails

Who: Agent harness authors. Steps: (1) State = tool name, args summary, user utterance, rules. (2) Prefer one Noul per hazard (destructiveness, exfil, scope creep) — not one vague “is this dangerous?”. (3) Code maps probabilities → allow / ask / block with per-hazard thresholds. Expected effect: fewer false negatives vs a single broad question (firewall/pi-warden authors report this pattern — cite as author-reported).

Get the prompt →

Agents ✦ 74 · outline

Tool-result context sieve

Who: Long sessions. Steps: Each tool result → Noul admit-to-context → keep/drop. Expected effect: Less junk in window (sibling to compaction).

Get the prompt →

Agents ✦ 86 · outline

Verbatim context compaction

Who: Coding-agent / long-session product teams. Steps: (1) Candidate tool-call/result blocks as state. (2) Noul/Choice keep vs delete/truncate. (3) Code splices kept text verbatim (no rewrite summary). Expected effect: shrink context without lossy paraphrase of kept spans (author claim). Caveat: keeping text verbatim ≠ proving omitted history was unnecessary.

Get the prompt →

Automation ✦ 57 · outline

Home automation intent

Who: HA experimenters. Steps: Utterance → Choice device/action → code executes safely. Expected effect: Jev advises, code controls.

Get the prompt →

Automation ✦ 78 · outline

n8n workflow routing

Who: Automation builders on n8n. Steps: Trigger → Jev community node or HTTP system-one → IF/Switch on choice/score/noul/confidence → branch. Expected effect: typed routing inside no-code workflows; low-confidence → human queue. Caveat: community node ≠ TypeSafe product.

Get the prompt →

Computer use ✦ 88 · outline

Browser / computer-use action select

Who: Browser-agent builders. Steps: Deterministic observe (DOM/OCR) → element table → one Jev call Choice operation + target → code executes → small LLM only for free text. Expected effect: lower $/step and latency vs screenshot-to-frontier every turn (author-reported demos, e.g. jev-ultrafast). Caveat: task demos ≠ general reliability; hosted Jev is text-only — vision/OCR is separate.

Get the prompt →

Computer use ✦ 62 · outline

Voice → browser intent

Who: Accessibility / demos. Steps: Partial transcript → Choice intent/target → wait/act/ask. Expected effect: Map speech to bounded actions.

Get the prompt →

Extraction ✦ 85 · outline

Candidate span extract + code normalize

Who: Back-office / document ops. Steps: Regex/roster finds candidates → Choice picks span (+ classify currency/country) → code normalizes verbatim (email lower, E.164, Decimal). Expected effect: no invented digits outside the candidate set. Dedup: catalog card; Hands-on extract-and-validate keeps the recipe.

Get the prompt →

Extraction ✦ 66 · outline

Date parts then code resolve

Who: Back-office. Steps: Ask parts named in doc → resolve/validate in code → confidence review. Expected effect: Arithmetic stays in code.

Get the prompt →

Extraction ✦ 63 · outline

Knowledge-graph entity align

Who: Data / catalog teams. Steps: Candidate pairs → Score merge/unlink/curator. Expected effect: Closed action set, no free threshold hunt.

Get the prompt →

Extraction ✦ 64 · outline

Markdown structure recovery

Who: Ingestion. Steps: Stitch wrapped lines → classify blocks with companion questions. Expected effect: Recover structure without full regen.

Get the prompt →

Extraction ✦ 65 · outline

SDE mini→verify cascade

Who: Extraction pipelines. Steps: Cheap extract → Jev verify fields → escalate reasoning model. Expected effect: Cost/quality tradeoff (cookbook-reported).

Get the prompt →

Guardrails ✦ 82 · outline

“Done” claim verification

Who: Coding agents. Steps: Agent claims done → Noul evidenced-by-tests/diff → allow stop or continue. Expected effect: Stop false completion.

Get the prompt →

Guardrails ✦ 83 · outline

Irreversible action gate

Who: Agent harness. Steps: Detect destructive/irreversible tool → higher threshold or ask user. Expected effect: Fewer irreversible mistakes.

Get the prompt →

Guardrails ✦ 85 · outline

Jailbreak / injection input screen

Who: LLM app owners. Steps: State=user message → Noul jailbreak/injection + Score harm → pass/review/block. Expected effect: Fail closed on high hazard.

Get the prompt →

Guardrails ✦ 84 · outline

LLM output policy check

Who: Trust & safety. Steps: State=model output + policies → Nouls → rewrite/block. Expected effect: Catch policy breaks before user sees.

Get the prompt →

Guardrails ✦ 78 · outline

MCP / tool semantic firewall

Who: MCP client authors. Steps: Pre-tool Nouls per hazard → ALLOW/APPROVE/BLOCK in code. Expected effect: Multi-hazard beats one vague question (author-reported).

Get the prompt →

Guardrails ✦ 56 · outline

Moderation with uncertain outcome

Who: Trust & safety. Steps: Choice labels include uncertain → compare auto vs human. Expected effect: Self-consistency pattern.

Get the prompt →

Guardrails ✦ 79 · outline

Secret / PII exfil gate

Who: Security-minded harness. Steps: Tool args + destination → Noul exfil/PII → block/ask. Expected effect: Catch data leaving machine.

Get the prompt →

Guardrails ✦ 80 · outline

Slop / low-quality patch detect

Who: Code review bots. Steps: Diff/patch state → Noul/Score slop signals → ask rewrite. Expected effect: Name the pattern back to agent.

Get the prompt →

Guardrails ✦ 81 · outline

Stuck / retry-loop detect

Who: Coding agents. Steps: Recent actions state → Noul same-strategy-repeating → inject steer. Expected effect: Break loops early.

Get the prompt →

Media ✦ 60 · outline

Caption sponsor skip

Who: Media tooling. Steps: Caption segments → sponsor probability → seek UI. Expected effect: Author-reported demo vs SponsorBlock sample.

Get the prompt →

Media ✦ 61 · outline

Page clutter hide rules

Who: Browser extension users. Steps: Page regions → Jev clutter judgments → save reversible hide rules. Expected effect: Cleaner reading (Unclutter).

Get the prompt →

Media ✦ 59 · outline

Social draft vibe check

Who: Creators. Steps: Draft (+ optional vision desc) → Scores → gate post. Expected effect: Pre-post quality gate.

Get the prompt →

Ops ✦ 58 · outline

First-visit page comprehension

Who: SaaS marketers. Steps: Fetch MD → bounded questions → present scores. Expected effect: Deterministic fetch + Jev judgments.

Get the prompt →

Ops ✦ 80 · outline

Inbox / log triage

Who: Solo ops / on-call adjacent tooling. Steps: Batch messages or collapsed logs → Choice category + Score/Noul urgency/noise → code maps to suppress / watch / notify / page (no auto-exec). Expected effect: cut noise before expensive analysis or human attention.

Get the prompt →

Patterns ✦ 55 · outline

Composite weighted Score

Who: Product scoring. Steps: Atomic Scores/Nouls → weighted sum in code. Expected effect: Tunable product score without re-prompt.

Get the prompt →

Patterns ✦ 53 · outline

ML features from Jev questions

Who: ML eng. Steps: Free text → many Noul/Score → vectors for classical model. Expected effect: Semantic features + Autoresearch loop.

Get the prompt →

Patterns ✦ 54 · outline

Noul band → human review

Who: Risk workflows. Steps: Keep noul visible; mid-band → human. Expected effect: Calibrated review queue.

Get the prompt →

Platform ✦ 90 · outline

Route to cheap vs expensive LLM

Who: Platform / AI eng owning multi-model stacks. Steps: (1) State = user request (+ risk tags). (2) Choice difficulty/risk or Score complexity. (3) Code picks small LLM vs frontier vs abstain. Expected effect: spend frontier tokens only when needed.

Get the prompt →

Realtime ✦ 52 · outline

Game/sim action per tick

Who: Interactive demos. Steps: Symbolic state (not pixels) → Choice action → code steps env. Expected effect: Real-time advise loop; safety in code. Demo / interactive experiment — keep safety and win conditions in code, not in free-form generation.

Get the prompt →

Realtime ✦ 51 · outline

Trading desk dry-run decision

Who: Quant experiments. Steps: Market state → Choice long/short/hold → **default dry-run**. Expected effect: Decision without auto-order unless key set. **Not financial advice.** Default to dry-run; never auto-order unless an explicit key/ops policy enables it.

Get the prompt →

Recruiting ✦ 82 · outline

Resume / recruiting screen

Who: Recruiting ops / ATS integrators. Steps: (1) State = resume JSON + role must-haves. (2) Score/Choice per requirement; Noul for red flags. (3) Code ranks / gates advance; low confidence → human. Expected effect: consistent first-pass filter. Caveat: not a hiring decision engine; no protected-class scoring.

Get the prompt →

Retrieval ✦ 71 · outline

BM25 shortlist then Jev rerank

Who: Search eng. Steps: BM25 top-n → per-pair question → reorder. Expected effect: Cookbook-reported top-1/top-10 lifts on CLERC.

Get the prompt →

Retrieval ✦ 88 · outline

Citation / claim support check

Who: RAG / research product teams. Steps: (1) Gather claim + evidence snippets (search optional). (2) Choice: supported / contradicted / mixed / insufficient. (3) Insufficient → do not assert. Expected effect: fewer hallucinated citations in generated answers.

Get the prompt →

Retrieval ✦ 67 · outline

Document batch classify

Who: Ops / compliance. Steps: Batch docs → Choice labels + `other` → store probs. Expected effect: Scale labeling; keep distributions.

Get the prompt →

Retrieval ✦ 68 · outline

Hierarchical classification beam

Who: Taxonomy owners. Steps: Beam over Choice probs down tree; confidence climb. Expected effect: Deep trees without 255-option blowup.

Get the prompt →

Retrieval ✦ 70 · outline

Line-by-line semantic find

Who: Docs search. Steps: Choice over line ids + Noul answer-exists → abstain if none. Expected effect: Avoid least-bad passage when answer absent.

Get the prompt →

Retrieval ✦ 69 · outline

NL predicates after SQL

Who: Data apps. Steps: SQL returns rows → Jev Noul/Choice per row → filter/sort in client. Expected effect: Semantic filter without LLM SQL gen.

Get the prompt →

Retrieval ✦ 93 · outline

RAG passage gate

Who: RAG / search engineers. Steps: Retrieve top-k → per passage Nouls (relevant / usable evidence / contradicts premise / injection) → route() in code → build prompt with accepted vs conflict blocks → LLM answers. Expected effect: drop injections/off-topic; surface premise conflicts; cookbook shows reshuffle vs similarity rank alone (cookbook-reported).

Get the prompt →

Routing ✦ 87 · outline

Agent / Codex model router

Who: Agent harness authors. Steps: State=task → Choice which model/tier → code dispatches. Expected effect: Cheaper default path.

Get the prompt →

Routing ✦ 89 · outline

Always include `other` / exception route

Who: Support & classifiers. Steps: Choice options include `other` → code handles unknown. Expected effect: Avoid forced mislabels.

Get the prompt →

Routing ✦ 91 · outline

Confidence-gated action

Who: Any high-stakes workflow. Steps: Choice/Score answer → if confidence low escalate else act. Expected effect: Act only when sure enough.

Get the prompt →

Routing ✦ 86 · outline

Edge HTTP intent router

Who: Edge/API authors. Steps: Request text → Jev Choice → Hono route. Expected effect: Low-latency edge classify.

Get the prompt →

Routing ✦ 92 · outline

Intent routing (handler split)

Who: Platform eng. Steps: State=message → Choice intent + Score complexity → code routes DB/LLM/human. Expected effect: Right handler without full LLM every time.

Get the prompt →

Routing ✦ 88 · outline

Priority Score separate from category

Who: Support ops. Steps: Choice category + Score priority independently → compose in code. Expected effect: Urgent wrong-queue still escalates.

Get the prompt →

Routing ✦ 90 · outline

Speculative fan-out questions

Who: API designers. Steps: Ask many independent questions one call → code keeps relevant answers. Expected effect: More signal per round-trip (cookbook latency/cost claims = cookbook-reported).

Get the prompt →

Support ✦ 84 · outline

Support reply policy audit

Who: Support quality / trust & safety. Steps: (1) State = draft reply + policy bullets. (2) Noul/Choice for breaches (over-promise, PII ask, wrong policy). (3) Block / rewrite path / escalate on confidence. Expected effect: catch policy misses before send.

Get the prompt →

Support ✦ 95 · outline

Support ticket triage

Who: Support / ops engineers owning queues. Steps: (1) Build state — message, sender, open orders, policy lists. (2) One call: Choice intent + Score complexity (+ optional Nouls). (3) Code confidence-gates → human, or routes to lookup / specialist LLM / human. Expected effect: cheap classify before expensive handlers; uncertain tickets escalate.

Get the prompt →