← All sessionsHomeSearch
Outskill GenAI Learning Portal·Roadmap·

The Five-Level AI Generalist Roadmap (+ Operating System & Monetization)

Outskill (portal authors) The roadmap is the portal's reconstruction of Outskill's May 2026 source document, with the portal's own dated-snapshot labeling discipline layered on
▶local recording  ·  LMS lesson

Session map

THE FIVE LEVELSMETA-METHODSOPERATING & EARNINGFive levelsproof-gated capabilityL1 · PRD methodspec, then shipL2 · RAG + MCPfacts + approved actionsL3 · Multimodalformat selection + rightsL4 · Agentsnarrow scope, real metricsL5 · Vibe codingown what you shipEvaluation methoddated snapshots, own evalsOperating systemT-shape + weekly loopMonetization8 paths, proof first
The five levelsMeta-methodsOperating & earning
click a node — its card pops up (drag it anywhere, × to close)
Concept

The map reads left to right — the five levels flow into meta-methods, then into operating & earning. Click any node to open that idea here; every timestamp jumps into the recording.

The short version

  1. The five-level capability sequence Paul's whole Outskill objective hangs on: L1 Foundations (PRD method), L2 Context & Connections (RAG + MCP), L3 Multimodal Creation, L4 Agents & Automation, L5 Vibe Coding — ~16–23 months total, 'sequential by capability, not by calendar; progress is based on proof, not elapsed months.'
  2. Every level ships three named projects with definitions of done, a proof badge, a month-by-month syllabus, and a project-brief template — the roadmap is a curriculum, not a listicle.
  3. The evaluation method outranks any ranking: name the task and consequence of error, define constraints, run two candidates on the same evaluation set, record primary/fallback/date-tested. 'The source's one-line recommendation is dated. The evaluation method is durable.'
  4. Bonus Guide 1 compresses it into an operating system: T-shaped skill, a six-step weekly cadence (learn one concept → build one proof → real use → measure → document → teach), and a scorecard of 'problems solved, not videos watched.'
  5. Bonus Guide 2 is the monetization playbook: eight paths, an outcome-first offer template, a proof-first 90-day sequence, explicit ethical boundaries, and the seven killers — 'sell the solved problem rather than the name of an AI tool.'

The concepts

01

The five levels: capability sequence, not calendar

Use AI well → connect it to your world → create across media → delegate to agents → ship software: five levels, each gated by shipped proof.

Level 1 Foundations (2–3 months): solo workflows, PRDs, intentional prompting — three recurring workflows saving 30+ minutes daily is the exit gate. Level 2 Context & Connections (3–4): RAG and MCP produce an assistant that knows your business and takes approved actions. Level 3 Multimodal (3–4): image, video, voice, audio, documents — complete creative outputs. Level 4 Agents (4–6): the model plans, calls tools, observes, iterates within limits. Level 5 Vibe Coding (4–6): production software with AI as the primary coding partner.

The governing rules: sequential by capability, not calendar; levels 2 and 3 may overlap; and advancement happens when the definition-of-done evidence exists, not when the months elapse. Each level's readiness gates are behavioral — L1's includes 'when an output is weak, you inspect the prompt, context, examples, and constraints before blaming the model.'

Worked example · from the session

The L2 exit test is characteristically concrete: a domain assistant answering with sources, three approved action types without copy-paste, and teachable to a non-technical teammate in 15 minutes.

Why it matters

This is the framework the 14-Day Sprint certifies (Level 1–5 certificates) and the organizing skeleton for Paul's entire Outskill scope — every other course in the push slots into one of these levels.

People get this wrong

The roadmap is a reading list to complete.

It's a shipping schedule: 15 named projects with acceptance criteria. Reading is the smallest part of every level.

L1 · Foundations PRD + prompting L2 · Context RAG + MCP L3 · Multimodal create + verify L4 · Agents narrow + measured L5 · Vibe coding own what you ship L2 and L3 may overlap 2–3 mo 3–4 mo 3–4 mo 4–6 mo 4–6 mo 3 shipped projects + proof badge gate Progress is based on proof, not elapsed months
Five levels, each gated by proof — with 2 and 3 allowed to overlap
Progress is based on proof, not elapsed months.
For your projects

This is THE map for the Outskill push: Catalyst ≈ L2–L5 applied, the Sprint certifies these very levels, and your KB itself is a Level-2 artifact (source-grounded assistant over your own corpus).

This idea elsewherebuilds onL1's prompting layerreappears inthe same staircase, Catalyst-sizedreappears inL5 as Catalyst teaches it
Go deeper

In one line: L1 Foundations (PRD, prompting) → L2 Context & Connections (RAG+MCP) → L3 Multimodal Creation → L4 Agents & Automation → L5 Vibe Coding. ~16–23 months; proof-gated advancement; L2/L3 may overlap; each level has 3 projects, a proof badge, a monthly syllabus, and a brief template.

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

What advances you a level — and what explicitly doesn't?

The definition-of-done evidence: shipped projects and proof badges. Elapsed months explicitly don't — 'progress is based on proof, not elapsed months.'

Which two levels may overlap and why?

2 and 3 — connecting knowledge and creating across media reinforce each other once Level 1 is comfortable.

02

Level 1 — the PRD method and five prompting principles

Problem → Requirements → Deliver: write the spec before prompting, because when you skip it the model must guess what you mean.

The core L1 skill is writing a concise PRD in under ten minutes before any serious prompt. The five prompting principles operationalize it: show don't tell (two examples beat adjectives), constrain the format (exact table schema), assign a specific role (expertise + audience + jurisdiction, not 'lawyer'), use deeper reasoning for non-trivial work, and save your wins into a personal prompt library so 'your best conversations become durable workflows.'

The anti-pattern is named: don't treat a chat model like a keyword search engine. And the three L1 projects — Weekly Digest, Interview Prep Machine, Doc Collapser — each convert a recurring chore into a specified, saved, measured workflow.

Worked example · from the session

The Weekly Digest proof standard: four consecutive weekly briefs and measured review time — not one good demo.

Why it matters

The PRD habit is the single skill the roadmap says matters most at Level 5 too — 'the discipline from Level 1 matters most here.' It's the compounding asset.

People get this wrong

Foundations means learning what AI is.

It means shipping three measured workflows. The roadmap's L1 is a production quota, not a theory unit.

This idea elsewherereappears inthe same layers, PRD-shapedreappears inPRD before build, Catalyst flavor
Go deeper

In one line: PRD loop: Problem → Requirements → Deliver, spec in <10 minutes. Five principles: examples over adjectives, exact format constraints, specific roles, extended reasoning for non-trivial work, saved prompt library. Exit: 3 recurring workflows saving 30+ min/day. Projects: Weekly Digest, Interview Prep Machine, Doc Collapser.

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

An output disappoints. What does an L1-complete practitioner inspect, in order?

The prompt, the context, the examples, and the constraints — before blaming the model.

Why 'save your wins'?

A successful prompt is a capital asset: stored, reused and improved, it converts one good conversation into a durable workflow.

03

Level 2 — RAG vs long context, and MCP as the protocol that won

RAG supplies the facts; MCP supplies controlled actions — and the first decision is whether you need retrieval at all.

The decision rule is refreshingly anti-hype: try long context first when the complete source comfortably fits and can be verified; use RAG when the collection is too large, changes often, or needs precise retrieval controls. The tool ladder runs managed workspaces (Projects, NotebookLM, Custom GPTs) → orchestration (LlamaIndex/LangChain) → vector stores (Pinecone/Weaviate/pgvector) → rerankers — add components only to fix observed failures.

MCP is 'the protocol that won' — USB-C for AI-to-tool connections, with prebuilt servers for mail, calendar, files, GitHub, Slack, CRMs and databases. The safety doctrine: 'the goal is not to connect everything. Start with one trusted tool, define its permissions, test failure modes' — and in the syllabus, read-only before drafts, drafts before writes, with scopes, audit logs, expiration and rollback recorded.

Worked example · from the session

The 'CRM With a Brain' project: read-only connection first, 'which deals are slipping and why' answered with record evidence, and a draft-action capability added only after query accuracy passes.

Why it matters

L2 is where the KB itself lives — and the twenty-known-answers evaluation habit (including one question whose correct answer is 'not present') is the quality bar this repo's future retrieval layer should adopt.

People get this wrong

RAG is the serious way to give AI your documents.

RAG is the fallback for corpora too big or churny for context. The serious part is the evaluation set and citation discipline, whichever transport you use.

This idea elsewherebuilds onRAG fundamentalsbuilds onMCP, the Catalyst versionreappears ingrounded assistant, applied
Go deeper

In one line: Decision rule: long context first when the verified corpus fits; RAG for scale/churn/precision. Component ladder added only against observed failures. MCP = open tool protocol; adoption doctrine: one trusted tool, defined permissions, tested failure modes, read-only → draft → write. Exit: cited domain assistant + 3 approved action types + 15-minute teammate handoff.

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

When is building a vector-store pipeline the WRONG first move?

When the whole verified corpus fits in context — long context first is the stated rule; retrieval components exist to fix observed failures, not to signal seriousness.

What belongs in the evaluation set for a company assistant?

Twenty known-answer questions including at least one whose correct behavior is saying the answer is not in the sources.

04

Level 3 — multimodal creation as format selection plus verification

Multimodal work is not trying every generator — it's selecting the right input and output format for the job, then verifying the result.

The level's shape: turn a whiteboard photo into a PRD, a meeting into actions and follow-ups, a product into a launch package, a hostile PDF table into structured data. The tool tables (video leaderboard with Kling/Seedance/Veo/Runway, voice with ElevenLabs/Whisper/Suno, documents with frontier models and OCR pipelines) are explicitly dated snapshots — 'treat the ranking and prices as selection guidance, not permanent facts.'

The Sora shutdown story earns its place as doctrine: the source uses a product's death to teach choosing capabilities over product dependence. Document rule of thumb: direct model analysis first for one-offs; build a parsing pipeline only when scale, repetition, or accuracy justify it.

Worked example · from the session

Whiteboard-to-PRD's definition of done is epistemic: 'the whiteboard owner confirms that the PRD preserves the intended meaning and does not invent missing requirements.'

Why it matters

The three projects are immediately practical, and the completion criteria (continuity, consent, rights, cost, disclosure, accessibility) are the professional wrapper most multimodal tutorials skip.

People get this wrong

Level 3 is about mastering the best image and video models.

It's about format selection and verification — the models are interchangeable cameras, and half the completion criteria are rights, consent and accessibility.

This idea elsewherebuilds onthe director craft under this levelreappears inthe beginner rehearsal
Go deeper

In one line: L3 = cross-format transformation with verification: whiteboard→PRD, meeting→everything, launch package. Tool rankings are dated snapshots; choose capabilities over products (Sora lesson). Documents: direct analysis first, pipelines only when justified. Completion includes consent, rights, cost, disclosure, accessibility.

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

What did the Sora story teach, beyond one product's fate?

Depend on capabilities, not products — a generalist who learned 'video generation' survives any vendor's shutdown; one who learned 'Sora' didn't.

05

Level 4 — agents: scope ruthlessly, measure operationally

An agent is a workflow where the model decides what to do next — and the most valuable agents in the source's case study were not the smartest, but the most narrowly scoped.

Definition first: plan → call approved tools → observe → iterate until done or a limit is reached — distinct from fixed automation where every branch is predetermined. The dynamic-workflows pattern scales it: a planner splits one bounded goal (audit a codebase for SQL injection) across parallel specialized subagents, each with only the tools and context it needs, results streamed back, merged and ranked.

The four design principles are the level's spine: scope ruthlessly ('personal assistant' is too broad; 'weekly expense-report categorizer' is testable); give tools, not long procedural descriptions; plan failure modes (checkpoints, cost ceiling, step limit, rollback); and measure operational readiness — resolution time, error rate, customer impact, supervision effort — 'benchmarks alone do not prove that an agent is useful.' The syllabus sequence is deterministic first → bounded model judgment → agent loop only where the next step genuinely varies.

Worked example · from the session

The Virgin Voyages case: agent count was not the success metric — narrow scope and operational measurement were. The inbox-triage project keeps label/draft/review mode 'until repeated tests justify broader permissions.'

Why it matters

This level's month-by-month sequence (deterministic → judgment → loop) is the same discipline the portal's n8n content teaches — stated here as the general theory of delegation.

People get this wrong

Better models will make agent design unnecessary.

The case study's winning agents won on scope and measurement, not intelligence. Design is the moat; the model is a component.

The most valuable agents were not simply the smartest; they were the most narrowly scoped.
This idea elsewherebuilds onthe employee rung, formalizedreappears inparallel subagents, Catalyst buildreappears inthe same boundary
Go deeper

In one line: Agent = model-directed loop (plan/tool/observe/iterate) within limits. Dynamic workflows: planner + parallel scoped subagents + merged results. Four principles: ruthless scope, minimal tools, designed failure modes, operational metrics over benchmarks. Build order: deterministic → bounded judgment → genuine agent loop. Exit: monitored narrow agent with owner, metrics, ceiling, rollback, recovery drill.

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

What separates an agent from an automation, and when is the agent unjustified?

In an automation every branch is predetermined; an agent chooses its next step. If the next step doesn't genuinely vary, the agent is unjustified complexity — use rules.

Name the four operational metrics that outrank benchmarks.

Time to resolution, error rate, customer impact, and supervision effort.

06

Level 5 — vibe coding as product ownership

Describe, generate, review, test, ship — and never ship what you cannot explain.

The five principles: PRD before code (L1's discipline matters most here); watch the burn rate (stop non-converging agentic runs); commit every working state ('AI-authored code is fast to create and equally fast to break'); treat security as YOUR responsibility (auth, permissions, secrets, inputs, dependencies); and don't ship what you can't explain — 'use AI as a fast junior developer, not as an unreviewed black box.'

The stack table chooses by who you are (non-technical founder → Lovable/Bolt; designer → v0; developer → Cursor/Claude Code/Windsurf), and the Composer spotlight distills to a durable rule: match the model to the stage of work — strong reasoning for architecture, fast models for in-editor implementation. The three projects graduate responsibility: weekend SaaS with a real user, the internal tool IT won't build (with an owner), and a reviewed PR on a real repository.

Worked example · from the session

The SaaS definition of done: one real user completes the core workflow, payments in test mode 'until commercial, legal, and support readiness are confirmed' — shipping discipline over launch theater.

Why it matters

'The tools will change quickly, but the principles of specification, testing, version control, security, and maintainability remain stable' — the level is engineering culture transplanted to non-engineers.

People get this wrong

Vibe coding means the AI owns the code.

You own the spec, the tests, the security, the commits, and the explanation. The AI owns the typing.

Use AI as a fast junior developer, not as an unreviewed black box.
This idea elsewherebuilds onthe same doctrine at level scalereappears inversion control every working state
Go deeper

In one line: Five principles: PRD first, burn-rate watch, commit every working state, security is the builder's, explainability before shipping. Stack chosen by learner profile; models matched to work stage (reasoning for architecture, fast for implementation). Projects: weekend SaaS w/ real user, owned internal tool, reviewed real-repo PR.

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

An agentic coding session is burning credits without converging. The roadmap's instruction?

Stop it — set a budget and halt runs not making measurable progress. Burn-rate awareness is a named principle, not an afterthought.

What's the explainability bar before shipping?

You can describe what each important function does. If you can't, it's an unreviewed black box wearing your name.

07

The evaluation method: dated snapshots, personal evals, confirmed-vs-rumoured

Every ranking in the roadmap wears a date stamp — the durable skill is running your own evaluation and refusing to let rumours dress up as facts.

The roadmap's most sophisticated layer is epistemic hygiene. All model releases, benchmarks, prices and product claims are 'a historical source snapshot… visually tagged Source snapshot — verify current status.' The replacement is the six-step personal evaluation: name the task and consequence of error; define input types, context, latency, privacy, budget; choose two current candidates; run the same evaluation set; measure accuracy, instruction-following, speed, cost; record primary, fallback, and date tested.

The news discipline generalizes it: any information panel must separate Confirmed (linked official announcement), Reported/rumoured (unconfirmed media claim), and Watch (anticipated) — 'never collapse rumours, previews, and general availability into one timeline.' And the cost-estimation habit: input + output tokens + retries + tool calls + hosting + storage + human review.

Worked example · from the session

'The source's one-line recommendation is dated. The evaluation method is durable' — stated verbatim about model selection, applicable to every table in the document.

Why it matters

This is the portal teaching the KB's own decay-class philosophy back to it — and the confirmed/rumoured/watch split is a ready spec for how Weekly AI Updates snapshots should be rendered.

People get this wrong

Keeping up with AI means reading the release news faster.

It means owning an evaluation set and re-running it on a schedule — recalibration by measurement, not by headline.

The source's one-line recommendation is dated. The evaluation method is durable.
For your projects

This concept is the tools-registry currency check (OI-026) written as pedagogy — asOf/changed/alternatives/verdict maps exactly onto the six-step evaluation.

This idea elsewherebuilds onjob-to-be-done, made proceduralreappears inwhy personal evals beat vibes
Go deeper

In one line: Dated-snapshot labeling for all perishables; six-step personal evaluation (task+consequence → constraints → two candidates → same eval set → measure → record primary/fallback/date); confirmed vs reported vs watch news separation; full-loop cost estimation including retries and human review.

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

What three buckets must a news panel keep separate?

Confirmed (official announcement linked), reported/rumoured (unconfirmed), and watch (anticipated) — collapsing them into one timeline is the named failure.

What does a completed model evaluation record?

Primary choice, fallback, and the date tested — because the answer expires and the dated record tells you when to re-run.

08

The generalist operating system: T-shape, weekly cadence, scorecard

Broad across six capabilities, deep in one vertical, and a six-step weekly loop that ends in teaching — measured by problems solved, not videos watched.

Bonus Guide 1 compresses the roadmap into practice. The T-shape: broad across prompting, context, creative AI, automation, agents, product — deep in one role-matched vertical (ops → automation reliability; product → requirements and prototypes; analyst → document grounding and verification). The weekly cadence: learn ONE concept, build one small working proof, use it on a real task, measure the result, document the method, share or teach it.

The scorecard inverts vanity metrics: problems solved not videos watched, working systems not saved prompts, verified outputs not impressive demos, reuse per week, time saved, real users, failure rate. The first-30-days plan sequences the portal's own activities: decision brief, source-grounded project, published artifact, draft-only automation.

Worked example · from the session

The portfolio evidence template: problem, who experienced it, old process and measured cost, AI-assisted process, tools, human approval point, tests, measured result, known limitations, next improvement — one page that IS the proof.

Why it matters

The habits and traps ('tweaking prompts indefinitely instead of shipping'; 'treating AI as the skill instead of workflow design, domain judgment, and implementation') are the course's most durable two paragraphs.

People get this wrong

Generalist means knowing something about every AI tool.

It means the T: broad capability plus one deep vertical — with weekly shipped proof as the pulse.

Go deeper

In one line: T-shape (broad six capabilities + one deep vertical); weekly loop: learn one → build one → use real → measure → document → teach; scorecard of solved problems, working systems, verified outputs, reuse, savings, users, failure rate; 30-day starter sequence; ship-weekly/recalibrate/teach habits vs the three traps.

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

What are the three traps?

Endless prompt-tweaking instead of shipping; trusting generic benchmarks over task-specific evaluation; treating 'AI' as the skill instead of workflow design, domain judgment, and implementation.

What closes every week?

A published proof: problem, brief, build, tests, measured result, known limitation, next iteration.

09

Ethical monetization: eight paths, outcome-first offers, proof before scale

Sell a measurable outcome to a narrow audience, build proof before promising scale — and never sell the name of an AI tool.

Eight paths: freelance specialist, productized service, creator/brand, vibe-coded micro-SaaS, custom assistants, content/video production, cohorts/coaching, and in-house AI champion — the last flagged as often lowest-risk because domain context already exists. Path selection maps to life situation, and stacking is disciplined: one primary path for 90 days; add the next only when the first has a documented acquisition and delivery loop.

The connective tissue is proof-first: the 90-day sequence (niche + three demo projects → published case studies → small paid pilot → standardize and price the outcome), the outcome-first offer template with explicit exclusions, and the proof-before-scale checklist (working demo, measured baseline, acceptance tests, human-review plan, known limitations, maintenance owner). Ethics are hard lines: no fake case studies, no unlicensed likenesses, disclosure of synthetic media, no selling legal/medical/financial certainty, human approval retained for consequential actions, everything in writing. All source revenue figures are 'examples, not guarantees.'

Worked example · from the session

The seven killers list is the diagnosis chart: learning privately, selling tool knowledge, positioning for everyone, underpricing without scope control, waiting to feel ready, ignoring distribution, scaling before documenting delivery.

Why it matters

This is the portal's version of Catalyst's entire commercial arc (sessions 1–8) in playbook form — cross-linking them turns the KB into one coherent monetization corpus.

People get this wrong

Monetizing AI skills starts with picking the hottest niche.

It starts with one solved, measured problem and honest proof — the niche is whoever verifiably has that problem.

Implement early, document results, and sell the solved problem rather than the name of an AI tool.
This idea elsewherereappears infirst clients, Catalyst arcreappears inproductization logicreappears inoutcome-first offers
Go deeper

In one line: Eight paths chosen by situation, stacked only after a documented loop; outcome-first offer (audience + measured pain + target state + system + inclusions + exclusions); 90-day proof-first sequence; ethics: no fabricated proof, rights and consent, disclosure, no certainty-selling, human approval, written terms; proof-before-scale checklist; seven killers.

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

What must exist before pitching a retainer?

Working demonstration, measured baseline, acceptance tests, human-review plan, known limitations, and a maintenance owner — proof before scale, all six.

Which path does the source call often lowest-risk and why?

In-house AI champion — the learner already understands the domain and builds with stakeholder context; the evidence then powers a role or compensation case.

Every concept, three clicks deep

The same concepts as a quick reference: the closed row is the glance, open is the study card, and every timestamp jumps into the recording.

01The five levels: capability sequence, not calendarL1 Foundations (PRD, prompting) → L2 Context & Connections (RAG+MCP) → L3 Multimodal Creation → L4 Agents &…›

L1 Foundations (PRD, prompting) → L2 Context & Connections (RAG+MCP) → L3 Multimodal Creation → L4 Agents & Automation → L5 Vibe Coding. ~16–23 months; proof-gated advancement; L2/L3 may overlap; each level has 3 projects, a proof badge, a monthly syllabus, and a brief template.

02Level 1 — the PRD method and five prompting principlesPRD loop: Problem → Requirements → Deliver, spec in <10 minutes.›

PRD loop: Problem → Requirements → Deliver, spec in <10 minutes. Five principles: examples over adjectives, exact format constraints, specific roles, extended reasoning for non-trivial work, saved prompt library. Exit: 3 recurring workflows saving 30+ min/day. Projects: Weekly Digest, Interview Prep Machine, Doc Collapser.

03Level 2 — RAG vs long context, and MCP as the protocol that wonDecision rule: long context first when the verified corpus fits;›

Decision rule: long context first when the verified corpus fits; RAG for scale/churn/precision. Component ladder added only against observed failures. MCP = open tool protocol; adoption doctrine: one trusted tool, defined permissions, tested failure modes, read-only → draft → write. Exit: cited domain assistant + 3 approved action types + 15-minute teammate handoff.

04Level 3 — multimodal creation as format selection plus verificationL3 = cross-format transformation with verification: whiteboard→PRD, meeting→everything, launch package.›

L3 = cross-format transformation with verification: whiteboard→PRD, meeting→everything, launch package. Tool rankings are dated snapshots; choose capabilities over products (Sora lesson). Documents: direct analysis first, pipelines only when justified. Completion includes consent, rights, cost, disclosure, accessibility.

05Level 4 — agents: scope ruthlessly, measure operationallyAgent = model-directed loop (plan/tool/observe/iterate) within limits.›

Agent = model-directed loop (plan/tool/observe/iterate) within limits. Dynamic workflows: planner + parallel scoped subagents + merged results. Four principles: ruthless scope, minimal tools, designed failure modes, operational metrics over benchmarks. Build order: deterministic → bounded judgment → genuine agent loop. Exit: monitored narrow agent with owner, metrics, ceiling, rollback, recovery drill.

06Level 5 — vibe coding as product ownershipFive principles: PRD first, burn-rate watch, commit every working state, security is the builder's, explain…›

Five principles: PRD first, burn-rate watch, commit every working state, security is the builder's, explainability before shipping. Stack chosen by learner profile; models matched to work stage (reasoning for architecture, fast for implementation). Projects: weekend SaaS w/ real user, owned internal tool, reviewed real-repo PR.

07The evaluation method: dated snapshots, personal evals, confirmed-vs-rumouredDated-snapshot labeling for all perishables;›

Dated-snapshot labeling for all perishables; six-step personal evaluation (task+consequence → constraints → two candidates → same eval set → measure → record primary/fallback/date); confirmed vs reported vs watch news separation; full-loop cost estimation including retries and human review.

08The generalist operating system: T-shape, weekly cadence, scorecardT-shape (broad six capabilities + one deep vertical);›

T-shape (broad six capabilities + one deep vertical); weekly loop: learn one → build one → use real → measure → document → teach; scorecard of solved problems, working systems, verified outputs, reuse, savings, users, failure rate; 30-day starter sequence; ship-weekly/recalibrate/teach habits vs the three traps.

09Ethical monetization: eight paths, outcome-first offers, proof before scaleEight paths chosen by situation, stacked only after a documented loop;›

Eight paths chosen by situation, stacked only after a documented loop; outcome-first offer (audience + measured pain + target state + system + inclusions + exclusions); 90-day proof-first sequence; ethics: no fabricated proof, rights and consent, disclosure, no certainty-selling, human approval, written terms; proof-before-scale checklist; seven killers.

Tools referenced

ToolCoverageMomentContext
ClaudementionedAcross levels: Projects (L2 managed RAG), Claude Code (L4/L5), Cowork and Claude in Chrome (L4 consumer agents), the nine-surface ecosystem snapshot
n8nmentionedL4 automation platform shortlist with Make and Zapier
CursormentionedL5 developer track; Composer 2.5 spotlight as the own-the-inference-stack example (dated benchmarks)
LovablementionedL5 non-technical track with Bolt, v0, Replit; also the micro-SaaS prompt-first stack
LangChain / LangGraphmentionedL2 orchestration and L4 agent frameworks (with LlamaIndex, CrewAI, AutoGen)
Pinecone / Weaviate / pgvectormentionedL2 vector storage tier — to be added only against observed retrieval failures
ElevenLabsmentionedVoice generation across L1 starter stack and L3 audio (with Whisper, Suno, Udio, AssemblyAI)
NotebookLMmentionedL1 starter stack and L2 managed knowledge workspace
OllamamentionedLocal experimentation option (with LM Studio)
VercelmentionedL5 hosting shortlist with Supabase, Firebase, AppSheet
GitHubmentionedL5's real-repository PR project; Copilot on the developer track
Kling / Veo / Runway / Seedance / Midjourney / FluxmentionedL3's dated video/image leaderboard snapshot — 'selection guidance, not permanent facts'

Action items

Resources mentioned

Resources
  • docOutcome-first offer template (Bonus Guide 2)
  • docOutcome-led outreach DM pattern (Bonus Guide 2)
  • docPortfolio evidence template (Bonus Guide 1)
  • docPer-level project-brief templates (L1–L5)
  • docResource libraries per level: official docs, papers (Attention, CoT, GPT-3, LoRA, hallucination, OWASP prompt injection), evaluation sites, MCP references

Extraction notes

This page was built from an auto-generated transcript, which garbles product and people's names. Those were corrected silently in everything above and logged here for transparency. The warnings flag claims that were true on the recording day but change fast.

True on recording day — verify before relying