$jevwiki.ai#an LLM wiki about Jev, written for agents rather than people

Agents: read the raw Markdown of this page, or start at llms.txt.

~/wiki/ideas

Builds: content, media, email, calls and reading

[ community tier ][ updated 2026-09-25 ][ confidence medium ][ jev-1.13.0 ]#ideas · builds · community · content · email · media

TL;DR 15 builds plus 12 names (split from Builds: data, search and business on 2026-09-25) where Jev checks writing, scores posts and ads, or labels mail, chats, calls, transcripts, images and reading passages, and code sorts, cuts or counts. Evidence is thin: most rows have no accuracy figures; Every's check caught 6 of 7 planted defects (Fable 7 of 7), and the best-measured reading tool (Working Memory Jev) saw answers change between identical runs and ships two samples per decision. Chat readers send other people's messages to a third party: see the warning line. Every number is one builder's own run.

How to read. Price checks use $0.042/Mtok input, output free (verified, Models, aliases, pricing, rate limits, context). Jev is text-only (verified), so audio and video go through speech-to-text or subtitles first (State: what you send Jev). Data, search and business builds: Builds: data, search and business; games and robots: Builds: games, simulators, robots and devices; agents: Builds: coding agents, harnesses and orchestration; browser and interface: Builds: browser, computer use and interface.

Writing, content and marketing

Build (builder) Jev decides Reported numbers Pattern Source
Every writing checks (Mike Taylor; Dan Shipper) 21 yes/no AI-tell questions per article (Noul, inferred); 11 experiments (context finding, reply grading, triage) 37 docs × 21 = 777 judgments in < 0.7 s, ~¼ cent; 1,709 judgments < 1 cent. Shipper: 4 checks × 12 passages, Jev 0.35 s median vs 8.83 s Fable 5.1, ~580x cheaper, caught 6 of 7 planted defects (Fable 7 of 7). Idea: a linter an agent runs after each paragraph Patterns: marketing, sales, GTM, content, support and ops P30 Every, @danshipper
Slop detectors (@kraayenJon; @jozef_gherman) One question per AI-slop tell on a page or text (type not stated) Kraayenbrink: 35 tells, 243 ms, $0.00015. Gherman: ~10,000 words in ~2 s Patterns: marketing, sales, GTM, content, support and ops P30 kraayenJon, jevdetector
SuperX post scorer (@robj3d3) 61 questions per draft; loop: draft, score, revise until the score stops rising ~1 s, $0.0004 per post; fitted on 9,481 posts from 207 creators; picks the viral post "2 in 3 times" (no baseline stated) Patterns: marketing, sales, GTM, content, support and ops P21 post
maxfusion ad teardown (Ori Silver, via @aresotik) Funnel stage and creative style per ad in a brand's ad library 1,891 ads in 19 s for $0.12 Patterns: marketing, sales, GTM, content, support and ops P21 post

Also seen, names only (vetted 2026-09-24): Human Compiler (a joke linter, but a clean split: ~30 nouls and 3 choices in one request, deterministic rules emit coded diagnostics), jev-audio-beeper (per-word profanity check over ASR timestamps, beeped with ffmpeg; no licence; a quoted ~466 ms is in no file, unverified), Jev Web Analyzer (ten Choices over a homepage; also a demo for the vendor's scraping product), sharp (@tshmieldev: Jev vs LLMs classifying an X feed; the comparison is only in a video).

Email and inbox

Build (builder) Jev decides Reported numbers Pattern Source
jevMail (ilyamk, @ILIA_AGI) Google Apps Script in your own Gmail: one Choice per email over up to 12 labels you describe; metadata first, the message body only when the answer is unsure or the label would archive; preview mode, archiving only for categories you allow, a per-run budget checked before each request No accuracy figures. Jev only through OpenRouter's alpha Decisions endpoint, billed there Patterns: marketing, sales, GTM, content, support and ops P22 repo
Jevmail (fazlerocks, Syed Fazle Rahman; not ilyamk's jevMail) Local Gmail triage, read-only by OAuth scope (gmail.readonly), local SQLite: one request per email via Vercel AI Gateway typesafe-ai/jev, three questions (tray of five: needs reply, updates, promos, sales, spam; urgency 1-5; written to you by a human?); top two probabilities shown; your corrections stored beside Jev's answer Author: 1,000 emails in about a minute for ~3 cents; no accuracy figures. Its note that the Gateway free tier allows 5 calls per 5 minutes is Vercel's limit (unverified here) Patterns: marketing, sales, GTM, content, support and ops P22 repo
Live-prioritised inbox (@jnnnthnn; "coming to @avec") Inbox ordered by importance instead of arrival time. Likely one priority judgment per email, sorted in code (inferred; video only) Berman narrates the demo as 100 emails sorted in under half a second; no accuracy figures. For ranking, mind the batching conflict on Builds: data, search and business (audits) Patterns: marketing, sales, GTM, content, support and ops P22, P20 post

Calls and messaging

Build (builder) Jev decides Reported numbers Pattern Source
Call Coach (ZeroGold) Live sales call: browser speech-to-text; after each sentence one request (next action Choice, buying stage Score, seven signal Nouls) over the last 40 turns; smoothing, hysteresis and tie-breaks in code; key stays in a local Node proxy Demo, no numbers. README warns that conversation text goes to TypeSafe (in Chrome, audio also goes to Google) and to pin a version instead of jev-latest (verified advice, Models, aliases, pricing, rate limits, context) Patterns: marketing, sales, GTM, content, support and ops P28, Patterns: judging, search, documents, real-time and markets P27 repo

Chat readers, names only (vetted 2026-09-25): Crush Monitor (pasted chats → per-message emotion and intent; its "$5 trial credits" line predates signups closing, console.typesafe.ai (console + playground)), 哑巴微信 (OCR of the WeChat window, an LLM drafts replies, Jev ranks them), 狗头军师 Jev Chat (screen-reads chats; ships a Windows EXE), WeChat Jev Assistant (reads WeChat's local database through third-party code fetched at setup; redacts names first; no licence), the jev-chat-jarvis family (iOS keyboard with Full Access plus desktop siblings: copied message → intent, risk, reply drafts). Warning for all of them: the other person's messages go to a third-party API without their consent, and client hooking or screen-reading can get an account banned. Sideloaded binaries: build from source.

Video, media and feeds

Build (builder) Jev decides Reported numbers Pattern Source
clipfast (@BurhanUsman) Type a topic, get the matching clips of a long video, one-click download. Jev must read a transcript (Matthew Berman guesses the same); timestamps and cutting stay in code (inferred) 90+ minute video clipped "under 2 seconds" for about 2 cents (author). 2 cents buys 476k input tokens at list price, ~26 passes over a 90-minute transcript (18k tokens, inferred; our arithmetic): plausible, not checkable. No accuracy figures Patterns: judging, search, documents, real-time and markets P27 post
jevclip (cclank, lanshu; README in Chinese) Subtitles cut at pauses into ~20 s segments; five questions per segment (8-way type Choice, 0-3 information Score, exaggerated-claim Noul, stands-alone Noul, up to 5 focus Nouls); code keeps or drops each with a stated reason, fills a ≤ 180 s highlight reel and a filler-free full cut with ffmpeg. An LLM summary may cite only segment ids, which code maps to timestamps and checks for numbers and names Synthetic 9 min 22 s test: all 12 blocks right by the second; 12 requests, ~17k tokens, $0.0007 (arithmetic verified). Real 16 min 52 s talk: 45 s segments dropped the key demo (averaged with filler to 0.49); 20 s segments kept it, 32 requests, $0.0019. Default thresholds are guesses; jevclip eval fits them to ~50 of your labels Patterns: judging, search, documents, real-time and markets P27, Patterns: gates, simulation, personas and other shapes (P38+) P42 repo
Jev Wrapped (gaborishka) Four questions per Telegram post, one request each: kind of post (one of ten), paid ad?, clickbait?, emotional pressure?; up to 1,500 posts from the last year, sampled evenly; code counts shares (an ad at ≥ 70% on the Noul, or ≥ 40% when the kind also said ad) ~1,500 posts in ~85 s via TypeSafe (2 pages at a time to stay under the 1,200 requests/minute limit, verified) vs ~35 s via OpenRouter (4 at a time); ~$0.10 per channel at list price. Limitation stated: on a channel with no paid ads it marked two partner promotions as ads Patterns: judging, search, documents, real-time and markets P18 repo
Image sorter (@fayazara) OCR text first, then Jev picks a category ~900 images in 40 s Patterns: judging, search, documents, real-time and markets P18 post
Nitpicky (@richard_meng_01, roe-ai) AI-image detector judging faces, fingers, text, poses region by region none. Jev is text-only (verified), so a vision step must describe regions first; how is not stated Patterns: judging, search, documents, real-time and markets P16 post

Also seen, names only (vetted 2026-09-25): changelog.earth (news as patch notes: an LLM selects and titles, Jev optionally checks eligibility and title accuracy), RefGarden (Jev picks search phrases over museum and NASA archives; the hosted build now refuses Jev keys with a 410), Book of Answers (an LLM lists options for a dilemma, Jev picks one; no licence).

Reading and learning

Build (builder) Jev decides Reported numbers Pattern Source
Book Aurora (dani1005) Ten questions per ~90-word passage in one request: a 0-3 Score for each of nine emotions plus intensity; code draws the picture Frankenstein: 601 passages, 6,010 decisions, ~22 s without the screen recorder, ~3 cents; a duplicate request fires after 1.5 s by default. Its direct-route variable is TYPESAFE_AI_API_KEY, the app's own name (TypeSafe's SDKs read TYPESAFE_API_KEY, TYPESAFE_* environment variables across SDKs) Patterns: judging, search, documents, real-time and markets P18 repo
Working Memory Jev / Passage (AustinAWay) Educator tool: where an English passage (≤ 6,000 characters) asks a reader to hold too much at once. A local parser proposes reading steps; Jev makes bounded choices on familiarity, active or background, grouping and relationships; code counts slots against a 1-5 budget Live benchmark (2026-09-21, 10 cases + 4 probes, every pass run twice): choice changed between runs 2.9% (6.4% at baseline), supported status flipped 5.7%, slot count changed on 3 of 43 steps; relationship accounting resolved on only 9.3% of clauses; support rate 0.717; 1,638 requests, 9.5M input tokens, ~$0.40 (arithmetic verified). Because answers vary it sends each batch PASSAGE_JEV_SAMPLES times (default 2): one sample gave 4.5% choice flips, two gave 2.4%. Author: not a measure of student memory. Custom licence: organisations with more than one employee, or any commercial use, must ask first Patterns: judging, search, documents, real-time and markets P18, Patterns: gates, simulation, personas and other shapes (P38+) P42 repo

What these builds teach when advising

Sources

Links are inline in each row; the raw captures (2026-09-20 to 2026-09-25) are listed in the frontmatter. Matthew Berman's video (machine captions, raw/community/youtube-matthewberman-8-jev-use-cases.txt) narrates the clipfast and inbox demos. Names-only entries without a capture rest on our vetting at source: 2026-09-24 (docs/sweep/2026-09-24-awesome-jev/part1.md) and 2026-09-25 (docs/sweep/2026-09-25-mrjev/partb.md). The writing, content and marketing rows moved here from Builds: data, search and business unchanged.