Layer 01: Model Reality
Understand the hard constraints first: tokens, capability boundaries, and output contracts.
Roadmap
Roadmap
The roadmap is no longer just a lesson list. It follows the underlying LLM knowledge network: model reality first, system design second, reliability and delivery after that.
Knowledge Network
The network underneath the roadmap
Lessons are the surface entry points. Underneath them is a graph of nodes, prerequisites, and delivery paths. Reading the knowledge network first makes the order far easier to understand.
Layer 02: System Design
Design context, retrieval, and tool use as explicit system layers instead of piling text into prompts.
Layer 03: Reliability
Make the workflow measurable, debuggable, safe enough, and efficient enough to survive repeated use.
Layer 04: Delivery
Turn the workflow into a product with identity, access, entitlement, and launch standards.
01
Stage 01 · Model Reality
Understand tokens, capability boundaries, and output contracts before you try to design more complex systems.
Token Budgeting for Serious AI Work
Token budgeting is not a billing detail. It is the first hard constraint on any AI system.
18 minFree
Reviewed lesson
Study this lessonDo Not Mistake Fluency for Truth: Capability Boundaries and Uncertainty Management
People who can really steer AI know when the model should answer, when it should clarify, when it needs retrieval or tools, and when it should stop.
24 minFree
Reviewed lesson
Study this lessonFrom Writing Prompts to Defining Contracts: Prompting and Output Contracts
Useful AI systems do not rely on model improvisation. They rely on clear task framing, structured output, and results that can actually be validated.
26 minFree
Reviewed lesson
Study this lesson02
Stage 02 · System Design
Learn to design context, retrieval, and tool workflows instead of stacking bigger prompts.
Context Architecture Instead of Giant Prompts
Context management is not about stuffing more text into a prompt. It is about designing how information enters and leaves the model.
22 minPremium
Reviewed lesson
Study this lessonRetrieval Is Not Just More Context: Retrieval and Grounding in Practice
Reliable AI systems do not pretend the model already knows every fact. They decide when evidence is required, how it is retrieved, and how the answer stays tied to sources and freshness.
28 minPremium
Reviewed lesson
Study this lessonStop Treating Agents Like Magic: Tool Use and Workflow Design
A reliable agent does not act because the model sounds confident. It routes through clarification, evidence, action boundaries, recovery order, and reusable operator skills.
30 minPremium
Reviewed lesson
Study this lessonGrounding Dies When Docs Rot: Source Freshness and Document Governance
Retrieval is not enough. A grounded system still fails when it retrieves stale policy, mixed document versions, or evidence with no owner, timestamp, or expiry rule.
29 minPremium
Reviewed lesson
Study this lesson03
Stage 03 · Reliability
Add evals, observability, guardrails, and cost control so the system can improve without drifting.
Designing Eval Loops That Actually Improve the System
Without eval loops, an AI product is mostly random trial and error.
20 minPremium
Reviewed lesson
Study this lessonStop Saying 'Looks Better': Rubric-Based Evaluation and Grading
If you cannot score quality in dimensions, you cannot improve it responsibly. Rubrics turn vague taste into reviewable evidence and repair priorities.
31 minPremium
Reviewed lesson
Study this lessonStop Guessing the Prompt: Observability and Debugging for AI Workflows
Mature AI systems do not debug by intuition alone. They use traces, failure labels, and replayable evidence so problems can be located and fixed instead of guessed at.
30 minPremium
Reviewed lesson
Study this lessonGuardrails Are Not a Slogan: Prompt Injection, Authority Boundaries, and Risk Control
Reliable systems do not trust a single line like 'ignore malicious input'. They define who can issue instructions, what content is untrusted, and which actions require confirmation.
32 minPremium
Reviewed lesson
Study this lessonDo Not Stare Only at Model Price: Latency and Cost Control for Real AI Products
The most common production failure is not that the model is too weak. It is that the workflow is too slow, too expensive, and full of avoidable waste. Mature systems treat latency and cost as product constraints from day one.
30 minPremium
Reviewed lesson
Study this lessonModel Routing and Unsupported Answer Policy
Serious teams do not send every request to the same model and they do not force every request into an answer. They route by task value, evidence need, latency budget, and the right to abstain.
32 minPremium
Reviewed lesson
Study this lessonWhen the System Must Stop: Human Escalation and Review Queues
Reliable systems are not the ones that answer everything. They are the ones that know when to stop, escalate, and preserve the evidence a human needs to review the case fast.
30 minPremium
Reviewed lesson
Study this lessonNext Expansion
Guided build lessons
The roadmap does not stop at concept lessons. Guided builds turn real tool setup, verification, and troubleshooting into checklists so learners can actually ship something.
Delivery · 55 min
LiveCreem Billing End-to-End Practice
Run checkout, webhook, customer portal, and in-app entitlement as one chain.
Open tutorialDelivery · 45-60 min
LiveOpenClaw from Zero to Running
Get OpenClaw running step by step with real validation instead of guessing.
Open tutorialDelivery · 50 min
LiveSupabase Auth in Production Practice
Build the full chain from database and auth to live page state.
Open tutorialNext Expansion
Assessment lessons
Assessment lessons force the learner to audit their own workflow for guardrails, latency, and cost, then leave with reports, boundary maps, and optimization priorities instead of passive understanding.
Delivery · 35 min
LiveFreshness Governance Audit for Retrieval Workflows
Audit one retrieval workflow for freshness classes, ownership, metadata, and stale-content handling before it quietly ships old truth as current truth.
Open tutorialDelivery · 40 min
LiveGuardrail Audit in Practice: Injection, Confirmation, and Containment
Audit one real workflow and turn vague safety concerns into a trust-boundary map, confirmation matrix, and containment plan.
Open tutorialDelivery · 35 min
LiveHuman Review Queue Lab for Safe Escalation Paths
Audit one workflow into a real escalation path with hard stops, queue ownership, SLA, and a handoff packet so risky or unsupported cases stop cleanly.
Open tutorialDelivery · 40 min
LiveLatency and Cost Audit in Practice
Audit a real workflow for request waste, context bloat, caching, async opportunities, and budget tradeoffs before changing models.
Open tutorialDelivery · 35 min
LiveOutput Contract Workshop for Verifiable Interfaces
Turn one fuzzy AI step into a contract with explicit schema, failure states, and downstream acceptance checks.
Open tutorialDelivery · 40 min
LiveRetrieval and Grounding Audit in Practice
Audit one evidence-dependent workflow for retrieval scope, freshness, provenance, and unsupported-answer handling.
Open tutorialDelivery · 35 min
LiveRouting Policy Audit for Model Choice and Unsupported Answers
Audit one workflow for task classes, model-path choices, fallback thresholds, and explicit unsupported-answer behavior before it reaches users.
Open tutorialDelivery · 35 min
LiveRubric Grading Lab for Reviewable AI Quality
Turn one workflow into a scoreable review system with dimensions, anchors, hard-stop rules, and grader instructions another reviewer can reuse.
Open tutorialDelivery · 35 min
LiveWorkflow Routing Lab for Tool Boundaries and Operator Skills
Audit one tool-using workflow for routing order, confirmation gates, recovery steps, and the operator logic that should become a reusable skill.
Open tutorial