@chatman-media/sales

LLM-powered sales funnel engine for conversational bots. Persona composition, funnel stage routing, A/B testing with ELO ratings, self-play evaluation, and a coach LLM that iterates on failing styles.

Built from the production sales layer of sales-guru — a Telegram recruitment bot that runs 24/7 qualifying inbound candidates for foreign work contracts.

What's inside

| Module | What it does | |--------|-------------| | Types | Style, FunnelStage, Hook, StageConfig — full Zod schema for a sales persona | | Prompt | composeSystemPrompt — builds multi-section system prompts: persona, tone, framework (AIDA/PAS/SPIN/NEPQ/Belfort), hooks, skills, stage guidance, KB context | | Stage router | nextStage — sub-ms Unicode-aware regex router for Cyrillic + English | | Stage classifier | classifyStage — LLM classifier with regex fallback, {stage, confidence, source} | | ELO | eloUpdate / eloUpdatePair — standard ELO math, K=32, symmetric pairwise | | A/B router | pickVariant — SHA-256 deterministic assignment, same user always gets same variant | | Skills | 25 persuasion techniques (Cialdini × 7, Voss × 5, NLP × 3, classical sales × 5, custom × 5) | | Built-in styles | 4 production-tested personas: marina-prime, cold-direct-pas, empathetic-nepq, flirty-belfort | | Self-play | runSelfPlayMatch — full RAG pipeline vs LLM-driven candidate; per-turn skill grading; reflect guard | | Pairwise | runPairwiseMatch — A vs B against same persona; comparative judge; symmetric ELO update | | Coach | proposeStyleEdits — reads losing transcripts, proposes concrete JSON edits to tone/hooks/guidance/few-shot | | Shadow eval | runShadowEval — Wilson 95% LB on B's win rate → keep / rollback / inconclusive | | Skill recommender | rankSkillRecommendations — Wilson LB ranking; draws count as 0.5 wins |

Install

bun add @chatman-media/sales     # Bun
npm install @chatman-media/sales # npm / pnpm / yarn

Peer dependency: @chatman-media/rag for ChatClient, EmbeddingClient, IKbStore, and answerWithRag.

Quick start — compose a prompt

import { composeSystemPrompt, getStyleOrThrow } from "@chatman-media/sales";

const style = getStyleOrThrow("marina-prime-v1");

const prompt = composeSystemPrompt(style, "qualify", kbContext, {
  userFacts: { city: "Москва", age: "24" },
  skills: attachedSkills,
});

Stage routing

import { nextStage, classifyStage } from "@chatman-media/sales";

// Fast regex (zero cost):
const stage = nextStage({ turnNumber: 3, currentStage: "qualify", lastUserMessage: msg });

// LLM with regex fallback:
const result = await classifyStage({ chat, userMessage: msg, currentStage: stage, turnNumber: 3 });
console.log(result.stage, result.confidence, result.source); // "pitch" 0.87 "llm"

A/B testing with ELO

import { pickVariant, eloUpdate } from "@chatman-media/sales";

// Deterministic assignment — same user always gets same style:
const styleSlug = pickVariant(
  { slug: "summer-2025", variants: [
    { styleSlug: "marina-prime-v1", weight: 50 },
    { styleSlug: "empathetic-nepq-v1", weight: 50 },
  ]},
  userId,
);

// Update ELO after a match:
const newRating = eloUpdate(currentRating, "won"); // K=32, baseline=1500

Self-play evaluation

import { runSelfPlayMatch, CANDIDATE_PERSONAS } from "@chatman-media/sales";

const result = await runSelfPlayMatch(deps, {
  style: myStyle,
  styleId: 42,
  persona: CANDIDATE_PERSONAS.find(p => p.slug === "skeptic-anya")!,
  maxTurns: 20,
});

console.log(result.outcome);          // "won" | "lost" | "draw"
console.log(result.transcript);       // full back-and-forth
console.log(result.skillsAttributed); // which skills the bot actually used
console.log(result.fabricationsCaught); // reflect guard catches

Coach LLM

import { proposeStyleEdits, applyEditsToStyle } from "@chatman-media/sales";

const proposal = await proposeStyleEdits({ style, matchesRepo, chat });
console.log(proposal.summary);   // "Bot too formal with price-sensitive personas"
console.log(proposal.edits);     // { voice_tone: "...", hooks_add: [...] }
console.log(proposal.rationale); // per-edit explanation

// Pure merge — returns new Style, nothing persisted:
const improved = applyEditsToStyle(style, proposal.edits);

Storage interfaces

All DB-heavy modules accept injected interfaces — no ORM dependency:

import type { ISelfPlayMatchesRepo, ISkillsRepo, IStyleRatingsRepo } from "@chatman-media/sales";

// Implement for your DB (Postgres, SQLite, in-memory):
class MyMatchesRepo implements ISelfPlayMatchesRepo {
  async insert(match) { /* ... */ }
  async byId(id) { /* ... */ }
  async list(opts) { /* ... */ }
}

Built-in styles

| Slug | Persona | Framework | Voice | |------|---------|-----------|-------| | marina-prime-v1 | Марина, PrimeConnect | NEPQ | Human recruiter, Telegram-native, warm | | cold-direct-pas-v1 | Менеджер | PAS | Direct, no fluff, fast pitch | | empathetic-nepq-v1 | Наталья | NEPQ | Warm, anxiety-aware, unhurried | | flirty-belfort-v1 | Виктория | Straight Line | Assertive, playful, confident |

Persuasion skills

25 atomic techniques from Cialdini, Voss, NLP, and classical sales — each with a promptFragment injected into the system prompt and an applicableStages filter:

import { SKILL_CATALOGUE, SKILL_BY_SLUG } from "@chatman-media/sales";

const mirroring = SKILL_BY_SLUG.get("mirroring");
// { slug: "mirroring", family: "voss", promptFragment: "...", applicableStages: ["qualify", "objection"] }

Architecture

@chatman-media/sales
├── types.ts          Style / FunnelStage / Hook schemas (zod)
├── prompt.ts         System prompt composition
├── stage-router.ts   Regex funnel stage router
├── stage-classifier.ts  LLM stage classifier
├── elo.ts            ELO rating engine
├── ab-router.ts      Deterministic A/B picker
├── skills/
│   └── catalogue.ts  25 persuasion techniques
├── styles/
│   └── *.ts          4 built-in personas
├── self-play/
│   ├── personas.ts   8 candidate archetypes
│   ├── judge.ts      LLM match judge
│   ├── orchestrator.ts  Full match loop
│   └── pairwise.ts   Head-to-head comparison
├── coach.ts          Style iteration from losses
├── shadow-eval.ts    Wilson-LB A/B shadow runner
├── skill-recommendations.ts  Wilson LB skill ranker
└── store.ts          Storage interfaces (IKbStore, ISelfPlayMatchesRepo, …)

Depends on @chatman-media/rag for ChatClient, EmbeddingClient, IKbStore, answerWithRag, and gradeSkills.

License

MIT — Alexander Kireev / chatman-media