AI User Course

🎯 The proposal DECISION OPEN

Audience — one room, ruthlessly: employed professionals and small-business owners in Spanish-speaking LatAm — Colombia beachhead first, México as the Cohort 2 expansion — specifically the non-coders: operations, marketing, finance/admin, founders, PMs, consultants. They use ChatGPT or Claude in a chat window today and have shipped nothing. Not the audience: Python-dev agent engineers (that Spanish lane is owned), absolute beginners (free courses serve them), English-first buyers (a separate Cohort 2 test).

The four-part wedge — any one part alone loses:

• En vivo, sobre tu trabajo real, hasta que funcione — live, on your real job, until it works. Spanish is table stakes, not the moat; the shipped outcome is.
• The Oct-2026 stack, taught live as it changes — model selection across the GPT-6 family (Astra / Sol / Luna) and the Claude family, Claude Code multi-agent + Skills / MCP / hooks, and evals + safety/authorization as the graded spine. None of that is free, and none of it exists in Spanish elsewhere yet.
• Practitioner proof, verifiable — Carlos shipped three working apps in Oct 2026 (RutinaGrid, Cartas, DocuES) by orchestrating agents, with public repos and failures shown. This answers the category's #1 student complaint head-on: instructors who never shipped anything to production.
• A shipped-outcome contract — you leave with one system running on your real job, a portfolio artefact and a Demo Day. Not 25 hours of video.

• Framework name — still to lock before any landing copy. Every premium funnel sells a named method, not a syllabus. Working proposal for Carlos to approve or replace on Day 1: Dirige · Verifica · Despliega (Direct · Verify · Deploy), built on the eval/safety spine.
• The conversion bridge: a ~$49 paid intro workshop / Setup Clinic (the flagship Safety/Evals lesson, credited toward the cohort). It warms buyers, filters for serious ones, and pre-solves the setup-abandonment problem before Session 1.
• The guarantee: Session-1 full refund + keep the template vault. Cheap at live-cohort scale, and — see the funnel wave — a genuine differentiator in Spanish, where the closest competitor publishes no guarantee at all.
• Framing: lead with opportunity and adaptation, fear as secondary urgency (regional conversation runs ~41% opportunity vs ~29% threat; Colombia is the exception where a stronger fear line can be tested — 42% see replacement as probable within 5 years).

🧾 How it started — the source doc, validated CHECKED

On Oct 26 Carlos pasted two versions of an AI-only curriculum from another agent — a 4-phase "World's Best AI User" framework and an older Weeks 3–10 version, the same material repackaged. Before any course talk, Mavi checked it against live sources. That validation is why the course below looks different from the document it came from.

• Real: Jensen Huang's core advice — from Jan 2025, on Cleo Abram's Huge Conversations: "the first thing I would do is learn AI," meaning ChatGPT, Gemini Pro and Grok, with prompting as the art of asking good questions. Also real, from GTC Berlin (Oct 21, 2026): GPT-6 Astra (Sep 3, 1.05M context, $10/$50 per 1M tokens), RTX Spark (~1 petaFLOP, 128GB), the DGX Station (20 petaFLOPS, ~748GB) and the Agile ONE robot (Q1 2027). Qwen 3.5 is real (Feb 2026).
• Fabricated — all removed, never to be republished: the headline Jensen quote ("completely obsessed with Cloud Code and Astra" — misspelled, and impossible in Jan 2025: neither product existed for him to name); Llama 3.5 (no such Meta model — the line went 3.2 / 3.3 to Llama 4); "Claude Code didn't exist in Jan 2026" (it launched Feb 2025); the decorative stats (AI "wrong 15–30%", a fine-tune that should "beat Astra by 15%", a "$50k/month" architecture burn) — all unsourced inventions presented as facts.
• The rule that came out of it: the course is rebuilt from primary sources and sold as inspired by Jensen's real advice — never as his playbook. In the corrected curriculum, students measure their own system's error rate in Week 6 instead of quoting a made-up one.

👔 Wave 0 — CEO: market & competition DONE

• Free just killed paid Phase 1. Anthropic launched 10 free courses on Coursera in Oct 2026 (Claude Code 101, Subagents, Agent Skills, AI Fluency for Small Businesses) plus free Claude Academy with certificates. Nobody can charge for "what is Claude Code" anymore — only for doing it live, on the student's own job, until it ships. Those free courses are now assigned as pre-work instead of competed with.
• Udemy is a $10 universe. Incumbents at 100k–400k students (one Claude Code + Cowork masterclass at 100,840; Ed Donner's tracks at ~354k–403k), $9.99–$12.99 street prices, and Udemy keeps ~63% of organic sales — roughly $3.70 per student to the instructor. A Year-2 catalog play, not a launch.
• Cohorts are where the money is validated: Maven's "Build an AI Agent" at $1,095, Claude Code for Founders at $1,597 (Cohorts 3+4 already scheduled), AI PM Accelerator at $1,500. Every winner shares: a named shipped outcome, tool costs disclosed on the sales page, and a first-session / Week-2 refund guarantee.
• Spanish: real, not empty. Santiago Hernández owns Spanish dev-deep on Udemy (~200k students catalog-wide; his Claude Code course ~15,397 at $9.99). The live proof of this wedge is Nexion IA Academy (Spain): a €50 intro into a €550 live Claude Code cohort — small group, work on your real case, updated every edition, positioned explicitly against Udemy.
• Student complaints are a fixable checklist: setup skipped (beginners strand in Session 1), beginner/advanced whiplash in one course, outdated within weeks, no checkpoints, hidden tool costs, instructors who never shipped.

🛠 Wave 0 — CTO: stack, topics & sourcing DONE

• Corrected Oct-2026 stack: Claude Code (primary) + Codex CLI on GPT-6.1 Sol (comparison) in Phase 1; Astra is a Phase 3 model, not a Phase 1 tool. With every frontier model now at ~1M context, the lesson is cost per accepted task — Sol at $2/$10 vs Astra at $10/$50.
• Cursor demoted to optional editor, never required — product churn (SpaceX acquisition, new X Subscription pricing Oct 12, legacy Pro closed) makes it the #1 production risk in the stack. The course must not be load-bearing on it.
• Evals become the spine of the whole course — the only content that survives the next model release, and the clearest differentiation from the $12.99 incumbents.
• Memory rebuilt: four layers ending in real frameworks (Mem0, Zep/Graphiti, Letta) plus memory safety — not the 2023 pattern of dumping embeddings into Pinecone. Retrieval defaults to pgvector with hybrid search + rerank + eval; the SEC-filings lab was fixed because Series B and layoffs are not structured fields in EDGAR (that becomes a lesson on queries your data cannot answer).
• Dead tech cut: AutoGen and Swarm are dead; the three-harness Week 5 stunt is cut to one framework deep-dive; "RAG 2.0" as a term is cut; QLoRA fine-tuning runs on Qwen 3.5 (real sizes) with Llama 4 as contrast — Llama 3.5 does not exist.
• Sourcing rule: the best free material (Anthropic Academy, HF Agents Course, Neo4j GraphAcademy, Microsoft's MIT-licensed AI Agents for Beginners) is link-and-assign only — no reuse licence. Prices never get hard-coded into video; they live in a dated "stack sheet" re-pulled from vendor pages the week of recording.

⚙️ Wave 0 — COO: launch, platform & the live cohort DONE

• Platform: Hotmart for Cohort 1 (~9.9% + $0.50/sale, no monthly fee). The deciding factor is payments, not fees: PSE, Nequi, Efecty and Bancolombia at checkout, installments, COP payouts to a Colombia account, full Spanish checkout, built-in affiliates. Maven / Teachable / Circle pay via Stripe, which doesn't support Colombia-registered businesses — Maven payout eligibility for Carlos is unverified and needs written confirmation before any Cohort 2 move.
• 14-day pre-sale: waitlist + 10 problem interviews first, build-in-public threads with the app demos as proof, the flagship Safety/Evals lesson live in Spanish, an English run tracked separately, cart closes Day 14. Flagged honestly: this compresses Maven's normal 6-week launch into 2 weeks.
• Live-curriculum-change protocol (Carlos's specific ask — AI changes weekly, so the curriculum will change mid-cohort): three tiers promised on the sales page — Stable Core (~70%, principles, frozen per cohort), Guided Current (~20%, tool walkthroughs, updated only between modules), Live Lab (~10%, experimental, never assessed). Plus: semantic versioning + a dated public changelog, no changes within 72h of a live session, no breaking change mid-module, students always assessed against the requirements published at cohort start, a weekly "What changed" post, a migration note within one business day if a taught tool dies, and 6 months of updated recordings for graduates. Explicitly not promised: "always latest" or lifetime updates.
• Carlos's time: a 25-student, 6-week cohort runs ~12–14 hrs/week solo, ~7–9 hrs/week with AI ops (agents draft community answers, delta notes and capstone pre-checks; Carlos approves everything student-facing). Structural caps: 30 students hard, no 1:1s in the base price, peer-review trios before Carlos feedback, 1-business-day support SLA, no weekend support.

💰 Wave 0 — CFO: pricing & student costs DONE

• What a student pays on top of tuition (6-week cohort = 2 billing months): minimum viable ≈ $100 total — Claude Pro $20/mo (Claude Code included) + ChatGPT Plus $20/mo (Codex/Astra access) + a ~$20 Sol-first API budget cap, with Cursor, Vercel, Pinecone, Weaviate and Neo4j Aura all usable on $0 tiers. A single-model lane drops it to a ~$50 floor. Full stack ≈ $781–$821 (Claude Max, ChatGPT Pro, paid tiers). RTX Spark hardware ($2,599+) is explicitly optional, never required. The ~$100 figure goes on the sales page, above the price — disclosed costs are what the winning cohorts do, and surprise is the refund driver.
• $299 is the only viable founding price. At Hotmart's fees, with 5% refunds and ~67 delivery hours: $99 needs 42 students to break even at $50/h — impossible under the 30 cap; $149 needs 28, right at the cap with no slack; $299 breaks even at 14 students (27 even at $100/h).
• Cohort 1 scenarios at $299: pessimistic 10 students = $1,920 net (~$35/h) · base 22 = $5,057 (~$75/h) · optimistic 30 = $7,298 (~$97/h). In every row the recordings become the self-paced asset afterwards. Later ladder: Cohort 2 $499, then $799; teams at $499/seat (min 10) or a $7,500 private cohort — B2B pricing is extrapolated from global sources, flagged as such.
• Paid ads are dead at this price: Meta's e-learning cost per purchase ($26.80–$58.90) eats the margin. Cohort 1 sells organically and through partners, or it doesn't sell.

🌊 Validation waves — every 20 min, 22:36 → 00:15 5 OF 6 RAN

Each wave took the next unverified angle and was allowed to falsify the plan. Five landed; Angle 6 (named live-cohort change exemplars) never ran, so the change protocol rests on the COO research alone — noted, not hidden.

• 22:36 — LatAm demand. Validated: demand is large and accelerating (Hotmart AI course sales +700% in two years; Coursera GenAI enrolments in México +356% YoY; 422k Spanish AI-and-work mentions in six months), and $299 sits in an empty pricing hole between $69–$249/yr subscriptions and $600–$1,800 live diplomaturas. Falsified: "Spanish-first is underserved" in broad form — Platzi, Coderhouse, Crehana and Domestika make native supply dense; the wedge narrows to live + non-coder + real work + Oct-2026 stack. Also changed: installments are critical for LatAm conversion, Colombia is the first beachhead, the fear headline is a minority frame regionally (Colombia the exception at 42%), employer reimbursement is weak in México (~11%) so the funnel can't be built on it, and free intro courses get assigned as pre-work.
• 22:58 — Competitor funnel teardowns (Maven $1,597, Nexion €50→€550, Pro University $47.97, page by page). Changes: a guarantee is a free differentiator in Spanish (Nexion publishes none) → Session-1 refund; add the paid ~$49 intro rung (a $49 buyer is a real signal, a free attendee isn't, and it gives the deposit gate a product); name the framework before the landing page exists. Pro University's stacked-bonus $47.97 pattern is the anti-model — stay out of that price universe entirely.
• 23:16 — IP & licensing. Teaching about the tools in your own words and linking vendor material is sellable. But: Anthropic's terms require written permission for brand use implying endorsement, OpenAI bars GPT/model names in product names, and Hotmart puts the IP warranty on the creator. Jensen's name/image in a title, hero or thumbnail adds publicity-rights and FTC false-endorsement exposure on top of the fabricated quote — three independent legal stops on any Jensen or vendor branding. Changed: Cohort 1's brand contains none of it; Module 0 links the real Cleo Abram interview, never clips it; two pre-deposit gates — a SOURCES.md licence page, and one scoped IP/ad-law review of just the sales page.
• 23:36 — Primary-source price check. The ~$100 minimum student cost was confirmed on official pricing pages. Cursor pricing was falsified: legacy Pro/Pro+ closed to new buyers on Oct 12 — new students face the X Subscription ($30 Premium + metered model usage), so the CFO's aggregator-sourced Cursor figures could not be published and Cursor stays optional. The dated stack sheet caught this before a student did.
• 23:58 — Channel economics. Validated: engaged email lists convert at 1–3%, webinar show rates plan at 40% for a Spanish consumer audience, Hotmart affiliate guidance is 20–40%. Falsified: Carlos's own list alone can pass the gate — ~150 names at 1–3% yields only ~2–5 buyers, and the original ~100-registration flagship target was far too low. Changed: JV/affiliate partners at 30% ($89.70/buyer) move from Days 9–12 to Days 1–3 as the primary engine; the flagship registration target rises to ~300; 1:1 calls with warm leads in the final days; any paid test gets a hard kill line of $10 per registration. Read on the gate: under 5 deposits after 100+ real contacts = a trust/distribution failure, not a price problem — cutting price won't fix it.

📚 Corrected curriculum v1 — 6 weeks, 10 modules DRAFT v1

• Week 0 (pre-work) — Why AI, honestly. The real Jensen talk assigned by link; how to spot fabricated AI advice, using the source curriculum itself as the case study; a free vendor intro as the baseline. Artefact: a personal "future-proof" problem statement.
• Week 1 — Operator fluency + coding agents. Context engineering and instruction files; Claude Code primary, Codex comparison, Cursor optional sidebar; Setup Clinic live before Session 1. Artefact: a deployed app/automation from the student's own job + one custom Skill.
• Week 2 — Memory. The four-layer model; a Mem0 vs Graphiti lab on the same history — diff what each remembers after a fact changes mid-history; memory safety (poisoning, right-to-forget). Artefact: an assistant that survives a fact change.
• Week 3 — Retrieval + graphs. pgvector → hybrid + rerank → Neo4j GraphRAG, over the student's own documents (SEC EDGAR as the shared fallback, reworked to what it can actually answer). Artefact: cited Q&A over a real corpus + a 20-question golden set.
• Week 4 — Multi-agent + MCP. Subagents/agent teams in one harness, written handoff contracts, a QA agent whose only tool is the eval suite; MCP security. Artefact: a mini-SaaS / research system with memory, cited retrieval and ≥2 agents.
• Week 5 — Computer use + safety/authorization. Supervised browser agents in an isolated profile (record → written procedure → supervised execution — the "watch me in Notion, replicate it" promise is cut); OWASP Agentic Top 10, permissions-as-code, the Stripe read-only / refund-approval lab, a promptfoo red-team lab. Artefact: a hardened agent + red-team report.
• Week 6 — Model selection, evals, fine-tune, scale + capstone. Routing by cost per accepted task on your own golden set; a Qwen 3.5 QLoRA lab; caching/batch economics and an architecture for 1,000 concurrent users with a bill of materials the student prices themselves; domain capstone + Demo Day. Optional enrichment: a sim-only physical-AI lab (Isaac Lab / Microduck) — with the honest correction that NVIDIA trained cloud-scale; no hardware required, ever.

🚦 The pilot gate — exact 14 DAYS

• PASS — ≥10 paid $50 deposits (or ≥8 full-pay $299 seats) and ≥50 waitlist emails and ≥5 completed buyer interviews → build Cohort 1; confirm the start date within 7 days; add a 2-week teaching runway.
• STRONG PASS — ≥20 deposits → build, plus open a second section or waitlist Cohort 2 at $499 immediately; record everything to broadcast quality.
• GREY (5–9) — extend once, 7 days, after interviewing near-buyers on the blocker; change exactly one variable in this order: proof artefacts / flagship lesson → partner distribution → installments. Price cuts last. No course build during the extension.
• KILL / PARK (<5) — do not build. Refund deposits the same day, publish the flagship lesson as free content, publish what was learned, keep the waitlist warm, revisit only with a distribution partner. Sunk cost by design: landing page + outline only.
• Confirmation-week bridge: 10 deposits is the gate, but 14 paid seats is the $50/h viability floor — if confirmation converts to fewer than 14 total paid seats, refund and park rather than deliver a cohort that loses money.
• Before the deposit button goes live (Day 4), two gates must clear: SOURCES.md + credits page with licence checks ticked, and the scoped IP/ad-law review of the final sales page, thumbnail and ad copy.

The 14 days, in summary: Day 1 — contact list (≥150 own names) assembled, JV shortlist built and outreach started, waitlist page + 5-question survey live, framework name locked. Days 1–3 — 10 problem interviews; landing copy mined from their words. Days 2–5 — one build-in-public thread/day + first app demo video. Day 4 — landing page live on Hotmart (only if both pre-deposit gates cleared; otherwise waitlist-only). Day 5 — flagship lesson announced; $49 Setup Clinic offered to registrants. Days 6–7 — flagship prep; mid-window count; if registrations track under ~150, add a partner or a run — never discount. Day 8 — flagship LIVE #1 (Spanish) with the pitch at the end. Day 9 — recording published as the lead magnet; partner mailings land. Day 10 — flagship run #2 (Spanish, evening) + Setup Clinic delivered. Day 11 — flagship run #3 (English; EN deposits tracked separately, never blended into the ES gate). Day 12 — deposit deadline #1; final-48h push begins (40–60% of launch sales land there). Day 13 — 1:1 calls with every warm lead (≥10); blockers logged verbatim. Day 14 — cart close + verdict, executed and communicated the same day.

⚠️ Open risks & unverified — do not publish until checked FLAGGED

• Payout eligibility: Maven / Stripe for a Colombia-resident seller has no primary source — it blocks every non-Hotmart route. Hotmart's exact fee for his category and its Colombia payout hold (~30 days, secondary source) still owe a primary help-centre check.
• Cursor Nov 12 report (OpenAI models leaving Cursor) is a single aggregator source — plan around the scenario, never publish it as fact. The GPT-6.1 Astra cancellation case study is likewise single-digest sourcing: corroborate before any marketing use.
• Competitor volumes: Nexion / Pro University enrolments and revenue are not public (prices and positioning are verified); Udemy student counts are aggregator-reported and two sources differ ~15% on the same course.
• Planning figures used directionally: completion rates, webinar attendee→purchase rates and replay uplift have wide variance in the sources; the channel wave's report also arrived truncated after its first section (its verdict and rates were complete and were used).
• B2B / team pricing ($499/seat, $7,500 private cohort) is extrapolated from global/US sources — no LatAm-specific corporate willingness-to-pay was verified.
• Legal: the IP review's cost is unquoted, and the licensing research is US-anchored — the counsel pass must also cover Colombian consumer and publicity law. A few ChatGPT / Vercel price numerals are corroborated but still owe one primary visual pass before they appear in student instructions.