How to Get the Most Out of Jev

Why this is in the vault

Third Every piece on TypeSafe's Jev in nine days — this one is a hands-on usage guide (install steps, prompt templates, a refinement loop) rather than a review, and it supplies the adoption evidence and concrete workflow recipe the prior two notes flagged as still missing.

The core argument

Jev — TypeSafe's classification-only "System One" model — has gone viral since its September 15 release: demos of it playing video games, re-ranking search results, and inferring intent from speech-plus-gesture (Jack Cheng's tldraw demo, 1M+ X views) are everywhere, and TypeSafe's release spawned an X leaderboard tracking who burns the most Jev tokens. Every's staff writer Laura Entis packages three internal use cases into a "steal this workflow" recipe: (1) install the official TypeSafe skill from github.com/typesafe-ai/skills so a coding agent knows how to structure Jev questions; (2) use a coding agent to interrogate a fuzzy judgment ("is this urgent," "does this sound like AI") into a plain-English working definition by walking through examples and edge cases; (3) break that definition into discrete, independently-scored checks — Jack Cheng did this for an email filter (real person? someone waiting on him? costs him something if ignored?) run across 90 days of his inbox; (4) validate by running checks on a sample the human has already judged, without feeding those judgments to Jev, comparing misses, and revising. Douglas Brundage's news-triage Grok bot went from reviewing <20 posts per run to 500+ after bolting on Jev. Pricing: 4.2 cents per million input tokens, output free — vs. Gemini 3.7 Flash ($0.75/$3.75), Haiku 4.5 ($1/$5), Sol ($4/$20), Astra and Fable 5.1 ($10/$50 each).

Mapping against Ray Data Co

This directly answers the open item both prior Jev notes left hanging. The 2026-09-15 vibe-check note flagged Jev as "the kind of cheap judge that could make a real benchmark-driven selector affordable" but had no concrete build recipe; the 2026-09-20 Innermost Loop note flagged Jev as "a tool already in RDCO's own stack (typesafe:typesafe-ai skill)" worth "a beat of due diligence" rather than passive filing. This issue supplies both the recipe and the due-diligence signal: TypeSafe ships an official skill (github.com/typesafe-ai/skills, separate from and possibly more current than RDCO's installed typesafe:typesafe-ai skill — worth diffing) plus a four-step pattern (define → decompose into scored checks → validate against a held-out human-judged sample → revise) that maps almost exactly onto how RDCO's fresh-eyes critic family (verify-vault-write, verify-dispatch, station-critic) already works, except those critics run once, post-hoc, on a full-reasoning model. The concrete next step this justifies: prototype a Jev pre-filter for one narrow, high-volume check inside an existing critic (e.g., a single binary sub-check inside verify-vault-write's rubric) using Entis's exact refinement loop, and measure it against the 4.2-cents-per-million-tokens number here rather than assuming the economics. Separately, the adoption signal itself (viral demos, a public token-burn leaderboard, Douglas's 20-to-500-posts before/after) is a real escalation from the 09-20 note's one-line curation mention — this is no longer "a vendor named in a roundup," it's a vendor with a public usage surge and an official agent-skill distribution channel, which raises the priority of the due-diligence beat those two prior notes opened.

Related

• [[2026-09-15-every-typesafe-jev-vibe-check]] — Mike Taylor's original vibe-check of Jev; this issue is Every fielding a full usage guide nine days later, with three of Taylor's own use cases folded in
• [[2026-09-20-innermost-loop-claude-rd-share-jev-embedded-evaluators]] — flagged Jev as a vendor already in RDCO's stack worth due diligence; this note is that due-diligence beat's first real data
• [[2026-09-10-every-evals-for-everyone]] — same "turn judgment into a repeatable, scored check" thesis, applied here to Jev specifically rather than personal benchmarks generally
• [[feedback_fresh_eyes_subagent_for_own_artifacts]] — the standing rationale for RDCO's post-hoc, full-reasoning critic passes; this issue's define-decompose-validate loop is a candidate pattern for a cheap pre-filter ahead of that pass
• [[feedback_delegation_model_effort_pairing]] — RDCO's model/effort selection is heuristic today; this issue's per-model pricing table (4.2¢ Jev vs. $10 Fable/Astra input) is a concrete data point for that gap

Curation section

• How We Write Now: Mike Taylor's 12-step AI writing process (companion piece to his prior "13 beliefs about writing with AI" post) — not about Jev, a separate Every series entry.
• Write-along event: Every writers (Mike, Jack, staff writer Katie Parrott) livestreaming AI-assisted writing sessions the following day at 1pm ET on X/YouTube.
• Data point: per-million-token pricing comparison table (Jev 4.2¢ vs. Gemini 3.7 Flash, Haiku 4.5, Sol, Astra, Fable 5.1) — folded into the core argument above rather than treated as a throwaway line, per the schema's absorption-failure guidance, but flagged here since it's a distinct, reusable data point.
• Every All Access / Builder Pack promo: standard newsletter-footer self-promo for Every's own subscription bundle (Sparkle, Cora, Spiral, Monologue) — unrelated to Jev, house self-promotion, not a third-party sponsorship.