Sales Unicorn · Canonical Curriculum · Sep 15, 2026
Two tracks, one system. Every topic: what it teaches, the scored exercise that gates it, and the playbook it ships. Leaders learn to govern, audit, and maintain; operators learn to build end-to-end.
Track One
For the person who owns the contract and the team. No table-building. Watchable in two evenings.
Clay is not an email sender, not a rep self-serve database, and not plug-and-play — it's the programmable data layer under your CRM. The correct mental model (the Use Case Pyramid: data foundation → outbound → inbound → custom) so every later decision starts from reality, not the sales deck.
Data Credits buy information; Actions meter operations; formulas, filters, imports, and lookups of data you hold are free — and you only pay for successful results. A leader who knows this ledger can read any invoice, question any burn, and stop fearing the wrong things.
There is no per-user credit cap in Clay; spend is governed by how tables are built and the protocols your team follows. The 8 habits as team policy: budgets, spike alerts, the pre-flight sign-off — and how one untrained seat burned 95% of a 500K-credit balance while the rest expired.
You can't QA what you don't understand — which is how a vendor once left 150K junk companies in a client's CRM. Open any vendor-built table and, in 15 minutes, check what matters: run conditions on paid columns, waterfall order, model tiers, write-back settings, naming.
The moment a vendor offboards, their tables keep feeding your CRM daily — and nobody on your team can touch them. The maintenance protocol: documentation standards to demand from any vendor, the ownership map, weekly/monthly checks, and how to safely pause, edit, or retire a live table.
Most waste is the expensive tool doing a free tool's job: a formula beats an AI column, a lookup beats an enrichment, a provider beats an agent for contact data, a BYO-key agent beats a provider for public research. The decision tree — including when the answer is "not Clay at all," and when an agency is still right (managed from strength).
Implementations without a named owner become shelfware within six months. Sequence the adoption: pilot first (data quality → broken processes → net-new plays), owners named, the metrics that prove ROI, and which operator topics map to which of your team's roles.
Track Two
For the people who build. Every topic: 8–12 min video → scored exercise → shipped playbook. Milestones gate progression — you pass or you don't advance.
Tables, enrichments, and integrations — what they actually are, where Clay sits between your data sources and your CRM, and the three myths that make teams misuse it.
Data Credits vs Actions, the free-vs-paid ledger, "you only pay for results," and what the March 2026 pricing change means for how you build. Knowing what's free is most of credit discipline.
Nest every table, number workbooks chronologically, name descriptively, keep a team folder. Hygiene isn't cosmetic — it's what stops someone re-running the wrong table at 3am.
Momentum beats mastery early. Fifteen builds you can ship in your first week, each under an hour and a small credit budget: event-list enrichment, CRM gap-fill, competitor-customer scrape, job-change alerts on champions, TAM slice, hiring-signal lists, meeting-brief generator, dedupe pass — tagged by function, effort, and cost.
Sources (CRM import, CSV, Find Companies/People, Google Maps) and the anatomy of a well-ordered table: sources → formulas → enrichments → exports.
Two rows for one person cost twice and produce one usable record. Normalize the domain early — it's your join key everywhere — and normalize titles, because inconsistent inputs poison every downstream step.
Waterfalls stop at the first hit — so ORDER decides the bill. Cheapest plausible provider leads; the expensive specialist fires only on the remainder; one validator per data type; failures cost nothing.
Auto-update off while building · run conditions on every paid column · filter before enriching · look it up before buying it · get the prompt right before it runs · order waterfalls by cost · dedupe first · watch the syncs. The habits that prevent the $700-first-two-weeks story.
Overspend is never one costly column — it's a reasonable-looking table times fifty thousand rows. Estimate rows × credits before running; test on 10–20 rows always.
Formulas, conditionals, and lookups are free and instant. Deterministic signals (keywords, name suffixes, domain TLDs) classify most rows before any AI runs — the single biggest cost saver in real builds.
Four tiers: nano/mini for simple classification, standard for structured extraction, premium for reasoning, web-research agents only for facts not in your data. Using a premium model for yes/no classification is a documented 80× overrun.
Every prompt: System role → Inputs as variables → User prompt with strategy, rules, confidence checks → exact Output format with a JSON schema. The iron law: return EMPTY, never "N/A" — "not found" strings break every downstream filter, sort, and export.
When a task needs web access AND heavy processing, split it: a cheap research column gathers, a gated smart column processes. Source-hierarchy prompts (filings → regulators → investor pages → press) with confidence levels that null low-confidence answers instead of guessing.
Sculptor drafts tables and columns from plain English; Sandbox Mode tests prompts and workflows without burning credits — the sandbox Clay "didn't have," ending the pay-to-learn problem.
Your CRM already paid for most of the answers. The free lookup chain — account ID → domain → normalized name — runs before any provider fires; "thin webhook, fat lookup" keeps stale payload data out of builds.
The CRM value wins every merge. Fill blanks only; stamp every Clay-written value with source and date; dedicated fields for discovered data; validate picklists character-for-character; ship writes click-to-run first, automatic once trusted.
Hygiene as a standing system: gap segments that enrich only what's missing, scheduled refresh cadences (contacts quarterly, firmographics monthly), and enrichment-status fields that make the engine observable.
Fire on the TRANSITION into a qualified state — never on every edit — and carry only the record ID; rehydrate everything from the CRM. This single design decision separates a precise system from a credit fire.
Identity resolution (person-first for PLG, company-first for sales-led), de-anonymization waterfalls that turn a Gmail signup into a named buyer at a named company, and free disqualification logic so reps only see real leads.
An ICP-fit score from firmographic weights, an intent score from behavior and signals, combined into A/B/C tiers — explainable so sales trusts it, fresh so it stays true.
Round-robin, territory, and ICP routing with SLAs and escalation; route-row and timed delays as orchestration; Slack alerts with full context; a routed record that lands complete.
Five signal categories, weights and decay windows (job change 60 days, funding 90), and why job change is the highest-converting B2B signal. Start with ONE — teams that launch five monitor none.
The living layer above tables: CRM-connected dynamic segments, gap-only enrichment, continuous enrichment for new records, scheduled refresh for old — and testing on 10–20 records before committing anything.
Build the account universe from ICP filters (or describe it to Sculptor — draft mode is free, committing costs), dedupe on LinkedIn URL, and stage sourced records so they never pollute pipeline metrics.
The champion-mover play end-to-end: monitoring segment → signal fires → new company enriched → owning rep alerted with context → contact created and linked → fed to outreach. The pattern generalizes to every signal.
When truth lives in BigQuery/Snowflake/Databricks, Clay connects directly — product usage, billing, and PQL signals flow into tables and Audiences without CSV shuttling. Service-account setup (least privilege), and the warehouse-vs-CRM source-of-truth decision.
Nothing reaches a sequencer without passing the gate: verified email, first name, signal populated. Then the handoff — Clay Sequencer or Smartlead/Instantly/Outreach — plus the deliverability basics whose absence is a hiring-interview red flag.
A base prompt governs tone, length, structure; personalization uses 1–2 data points (more feels creepy); a review step tracks edit rate (>30% = fix the prompt) and skip rate (>10% = fix targeting); automation is earned gradually.
Reps don't open Clay; Clay comes to them. Through MCP, a rep asks their AI assistant to find, enrich, and research — with per-rep credit budgets — ending the "reps can't use Clay" problem without making every seller a builder.
Functions package reusable logic (your waterfall, your classifier) with testing and observability, callable from any table, audience, or workflow; Workflows chain steps visually so multi-stage automations stop living as delay-hack table chains. Function vs formula vs table — and how a team shares one standard instead of five drifting copies.
Most teams can run a campaign; almost none can say what it returned. The six-part telemetry spec — campaign object, member-status ladder, source/UTM mapping, influence model, routing SLA, scorecard — built so attribution lands at the campaign level and survives every handoff.
The full defect taxonomy from real production audits: uncompiled run conditions, prompt/schema drift, orphaned columns, fields computed but never written, whitespace names, "N/A" strings, wrong model tiers — plus the 60-minute triage for "reply rates dropped 50% overnight."
Graduation. Pick your function's playbook track and build it completely — Inbound (capture → enrich → score → route) · Outbound (signals → list → personalization → sequence) · RevOps (hygiene engine + routing + warehouse/CRM data layer) · Marketing (audiences → segments → ads/nurture sync) · Sales (rep intelligence + expansion signals) — with campaign telemetry on top and a scorecard that answers "what did this cost per opportunity?"
Install the course's Claude Code plugin in one command; drive Clay from the terminal — build tables from plain English, generate house-format prompts, estimate costs, run Clay workflows via CLI and MCP. The on-ramp to the measured $40–45K "coding premium."
← Mission Control · Canonical source: Clay-Course/curriculum/CURRICULUM.md · Sales Unicorn · September 2026