Patterns: marketing, sales, GTM, content, support and ops

TL;DR Eleven business-side decision shapes from the captured X posts and repos: leads and deals, content and creative, support and approvals, plus one group (metric-driven ad and budget calls) that is a poor fit as posted and needs restructuring. Community tier: official pages win every conflict.

How to read this

One pattern = one narrow judgment. Fit is our verdict against Jev 1.13 jaggedness: known failure modes and System One Models: strong = bounded, semantic, code owns the rest; workable = real, caveat named; poor = leans on arithmetic, dates, time series, generation or huge state. Pattern IDs are permanent; the index across all domains is Decision patterns from the community (with fit verdicts). Two main sources (@AIGuide_, @shannholmberg) are idea lists with no shipped system or measurement behind them: unverified until run on your own labelled examples.

Lead, account and deal decisions

P19 Lead and ICP scoring, fit kept apart from intent

• Decision how well does this account match the ICP; does it want to buy now
• State firmographics, role, inbound form or message, the ICP definition verbatim, the account record you already hold
• Ask separate Scores for fit, intent, urgency, pain — never one blended one; then a Choice over sales | self_serve | nurture | review for the next step
• Code weight the dimensions yourself; keep each answer beside the lead record so a rep can inspect the assessment; change coefficients, not prompts
• Fit strong — textbook composite scoring, and the fit/intent split is the key move
• Map Composite scoring, Score questions
• Seen @yuhasbeentaken, @chddaniel, @shannholmberg (form response + ICP criteria, four next steps), @AIGuide_ (inbound qualification, outbound account prioritisation), @startupideaspod (a design agency's contact form scored 0–1 on "is good lead")

P20 Signal detection and reply prioritisation

• Decision does this post signal a switch, hire, raise or pain point; which comment to answer
• State one post or comment plus minimal author context
• Ask one Noul per signal type in a single request, plus a reply-value Score
• Code rank by probability, dedupe by author, hand the top slice to a human
• Fit strong — many tiny independent judgments per record
• Map Speculative fan-out
• Seen @yuhasbeentaken, @skeptrune, @AIGuide_ (re-score accounts on hiring, funding and site changes instead of working a static list)

P28 Next-best action on a deal or account

• Decision after a call, email or CRM update, what should happen next
• State the deal record: stage, last interactions summarised by an LLM, open objections, what has already been sent — not the whole CRM history
• Ask Choice over call | email | wait | send_proof | bring_in_founder | nurture | close_lost, one criterion per option; the live-call variant picks bring_in_ae | technical_help | case_study | surface_pricing | book_next_meeting | nothing
• Code map the answer onto a playbook step code already owns; keep the probabilities on the record; when a wait or follow-up fires is a code rule, not part of the answer (see P33)
• Fit workable — bounded options and a semantic judgment, but the options overlap easily (call vs email vs wait) and Jev reads criteria literally, so write each option as a distinct situation and run it recommendation-only first
• Map Intent routing, Writing instructions and criteria that Jev reads correctly
• Seen @AIGuide_ (next-best action, live-call escalation, "CRM automation as choices rather than if-X-then-Y"), @chddaniel

P29 Churn and retention intervention triage

• Decision does this account need an intervention right now, and which one
• State pre-computed labels only: account-value band, usage change in words ("weekly sessions roughly halved since the last billing period"), support-history summary, NPS band, contract stage, interventions already tried
• Ask Noul per risk signal (unresolved support issue? champion gone?), a severity Score, then Choice over ignore | send_education | csm_outreach | offer_help | escalate
• Code your telemetry computes every delta, ratio and date gap before the call; cap outreach per account per period; log the chosen intervention so thresholds can be retuned
• Fit workable — the usage drop and the contract date are arithmetic and date comparison (failure modes 2 and 3 in Jev 1.13 jaggedness: known failure modes); bucket them in code and what is left is an ordinary routing choice
• Map Intent routing, Jev 1.13 jaggedness: known failure modes
• Seen @AIGuide_ (usage drop → ignore, education, CSM outreach, help, escalate)

Content and creative

P21 Creative scoring, ad teardown and content dedupe

• Decision which hook is worth testing first; does this idea already exist on our site
• State one creative with its format attributes; or one idea paired with one existing page
• Ask structural Nouls (does the first line open a loop?); Noul "does existing cover candidate?"; for teardowns a Choice per defined axis (hook, format, offer, awareness level) plus Noul "does the landing page support the ad's promise?"
• Code loop the pairs in code; store classifications beside the source links; join labels to your performance data and compute rates there
• Fit strong for labels and per-pair dedupe; poor as "here are 50 ideas and my site, pick 10"
• Map Cookbook: Knowledge graph entity alignment
• Seen @RoundtableSpace (430 ads/second, $0.6), @yuhasbeentaken, @shannholmberg (Meta Ad Library research into fixed categories), usenotra/notra

P30 Draft QA against voice, brand and claim rules

• Decision does this draft pass each rule, and is every claim supported
• State the draft plus the rule text verbatim — voice file, anti-slop rules, approved and rejected examples, the brief, approved claims, required offer details, channel requirements, and the source passages
• Ask one Noul per rule, each written literally ("does the draft state the trial length?", "is claim C supported by sources?"), plus a Score per dimension (voice, clarity, brand fit); give the claim check an uncertain option rather than forcing a verdict
• Code aggregate failures with max, not a mean; return failed drafts to the writing agent with the failed criteria attached; passes move to sign-off, uncertain claims to a person; chunk long drafts — 32k covers state plus the longest question
• Fit strong per rule; poor as a single "is this good?" question. Never ask Jev to rewrite the draft
• Map Cookbook: Double-checking citations, Cookbook: Guardrails for LLMs, Writing instructions and criteria that Jev reads correctly
• Seen @shannholmberg (voice/brand check on drafts; pre-sign-off check of offer details, claim support and channel requirements)

P31 Brief, idea and news-response triage

• Decision is this creative brief ready for production; is this incoming idea worth developing; does this story deserve a response
• State one item plus the campaign goal, audience, offer, brand rules and production limits; for news, the story with its source and publication time
• Ask a Score per named criterion (hook clarity, brand fit, offer clarity, fit with the campaign goal, similarity to work in production); Noul "does this fit an active campaign?"; Choice over open_brief | link_duplicate | save | editor, or research | skip | review for news
• Code rank and route on the scores yourself; attach the weak criteria to anything returned to the briefing agent. "Already published?" is P21's per-pair dedupe — loop candidate × existing page in code, never paste the content inventory into one state
• Fit strong per criterion and per pair. Publication time is a date: bucket it in code before asking
• Map Composite scoring, Cookbook: Knowledge graph entity alignment
• Seen @shannholmberg (brief review, idea triage, news response), monteduro/killmyidea (8 questions scored 0–4 in one request; weighted average, KILL/FIX/SHIP thresholds and an understandability gate all in code)

P32 SEO page review and internal-link checks

• Decision does this page answer its target query, what is missing, and is this internal link worth adding
• State one page or draft with its target query and business priority; for links, the source page, destination page and proposed anchor text
• Ask Score "how completely does this page answer query?", one Noul per subtopic you care about, and Noul "does the destination help the reader take the next step?"
• Code the missing subtopics are your false answers, not a list you ask Jev for; sort the update queue in code; approved links enter the implementation queue with both URLs
• Fit workable per page and per link. Flag: "current search results" comparisons smuggle in ranking positions — numeric inputs jev-1.13 handles badly. Ask pairwise semantic questions ("does the competing page cover X that ours omits?") and keep positions and any ranking estimate in code
• Map Cookbook: Re-ranking, Jev 1.13 jaggedness: known failure modes
• Seen @shannholmberg (SEO prioritisation, internal links; the quoted 2026-09-18 post adds top-10 SERP comparison — the part to restructure)

Support, ops and approvals

P22 Inbound triage fan-out: tickets and email

• Decision owning team, urgency, category, spam, refund intent, frustration, churn risk, whether a reply is needed
• State one message with the minimum context — plan, account age, recent history; never the archive
• Ask Choice for department or next action, Score for priority and for reply value, Noul per intent, all in one request per item
• Code worker pools for volume; your policy combines the answers. Same shape backfills: run an old inbox or ticket archive to surface missed leads. Which results a person sees is P06 in Patterns: agent internals, context and coding agents, not a new question
• Fit strong — the canonical documented shape, and no writing is required
• Map Speculative fan-out, Intent routing
• Seen @chddaniel, @akshay_pachaar, @0xMovez (cites 500 emails for ~3.5 cents), @startupideaspod (Vogel: 1,700 real emails → category, priority low | medium | high | important | urgent, spam score, reply score, 18 cents; plus ticket routing to product teams and old-inbox mining)

P23 Policy-bounded approval triage (refunds, incidents, discounts)

• Decision approve, counter, hold, review, deny or escalate inside a policy you already wrote
• State the request, the policy text verbatim, relevant history, fraud signals, and for a discount the stage and any competitor pressure — as words, with the numbers pre-bucketed
• Ask Noul "is this covered by policy?", Noul per fraud or exception signal, severity Score, Choice over approve | counter | hold | escalate
• Code every currency limit, margin floor, deal-size band and cap is a code comparison, not a question — the numeric half of a discount call belongs to P33
• Fit workable — coverage is semantic, money is arithmetic; run recommendation-only until you have a labelled set
• Map Jev 1.13 jaggedness: known failure modes, Cookbook: Guardrails for LLMs
• Seen @Layton_Gott, @chddaniel, @AIGuide_ (rep asks 20% off → approve, counter, hold, escalate)

Metric-driven decisions (poor fit)

P33 Metric-driven ad, budget and timing decisions

• Decision stop/scale/hold an ad; refresh or replace a creative; where the next ad dollar goes; when the next touch happens; whether margin supports a discount
• State as posted: spend, conversions, CPA, ROAS, frequency, click/conversion history, reporting period, margin, deal size
• Ask as posted: Choice over stop | scale | hold, refresh | replace | leave | review, or a delay bucket
• Fit poor as posted — comparing CPA/ROAS to a target is arithmetic and numeric representation (failure mode 2); "last three periods", "every few minutes", "5 minutes vs 2 hours" are date/time comparison (mode 3); a metric history is a digit-heavy time series with no semantic content (mode 5) (Jev 1.13 jaggedness: known failure modes). Workable once restructured:
• Code computes every metric, threshold and trend, then sends labels plus the semantic parts only. A human owns spend changes. Sketch (inferred, not from a post):

{"model": "jev-latest",
 "state": {"ad": "UGC testimonial v3", "goal": "demo bookings",
  "cpa_vs_target": "above target, 3 periods running", "roas_trend": "falling",
  "frequency": "in fatigue band", "comments_summary": "several 'seen this already'",
  "landing_page_matches_promise": "yes"},
 "questions": {
  "fatigued": {"type": "noul", "instructions": "Do the comments show audience fatigue with this creative?"},
  "action": {"type": "choice", "instructions": "Which action do the labels support?",
   "criteria": {"stop": "goal missed and no fix in sight", "scale": "beating target, no fatigue",
    "hold": "on target", "refresh_creative": "concept works, creative is tired",
    "human_review": "labels conflict"}}}}

Gate action on its confidence; low → human_review. Timing: Jev labels the lead (buying signal | researching | cold), a code table maps label → delay.

• Map Jev 1.13 jaggedness: known failure modes, Cookbook: Date extraction, Confidence-gated routing
• Seen @shannholmberg (stop/scale/hold; refresh/replace/leave/review), @AIGuide_ (response timing, budget reallocation "every few minutes", the numeric half of discount approval; policy half is P23)

Starting point

• Find the expensive queue. Greg Isenberg's filter: a business with a costly queue of incoming information; "put Jev at the front of the queue" — inbox, contact form, tickets (@startupideaspod).
• One task, on examples you already reviewed. Start with a single decision and items your team has already judged; compare Jev's answers with those reviews before the workflow acts (@shannholmberg). Method: Testing and evaluating a Jev workflow.
• Advisory first. Vogel keeps Jev advisory, not on 100% of interactions: high-value or low-confidence items go to a person (Confidence-gated routing).
• Cost check (verified). Vogel's 1,700 emails, ~4M input tokens, 18 cents: 4M × $0.042/M = $0.168, output free (Models, aliases, pricing, rate limits, context). His ~200 ms/query is unverified.

Related

• Decision patterns from the community (with fit verdicts) — the index of all patterns; Consult guide: could Jev help this project? — running the "could Jev help this project?" conversation
• Patterns: agent internals, context and coding agents — agent internals, routing, guardrails (P06 human queue); Patterns: judging, search, documents, real-time and markets — search, ranking, labelling at scale
• Field reports: independent evaluations, critiques, open replicas — independent measurements; Community repos: what people built and how they use Jev — the repos named in Seen
• Use-case map by industry — the official use-case map by industry

Sources

Post and repo links are inline in each pattern's Seen line. The three lists driving this page are @AIGuide_ (GTM, 2026-09-20), @shannholmberg (marketing workflows, 2026-09-21) and Ryan Vogel on the Startup Ideas Podcast via @startupideaspod (2026-09-18). Every captured file, including the earlier posts and monteduro/killmyidea, is in this page's frontmatter sources: under raw/x/ and raw/x-repos/ in the private repo (indexes raw/x/INDEX.json, raw/x-repos/INDEX.json).