GUIDES
120 reusable mental models for AI.
Use a Guide when you need to understand one Concept now; use a Course when you want an ordered next step. Browse directly or filter by domain, topic, course, and difficulty.
120 Guides·Want an ordered path? View Courses →
START FROM YOUR PROBLEM
You do not need the vocabulary before you can diagnose the symptom.
Choose the failure mode that looks most like your current problem. AhaFrame maps it to existing mental models, Practice, and a deeper Course without creating a second learning graph.
My RAG retrieves the wrong documentThe right source exists, but retrieval keeps selecting weaker, stale, or irrelevant evidence.
My agent repeats an action after timeoutA missing response is being treated as confirmed failure, so retries can duplicate irreversible side effects.
My agent has too much permissionThe model can reach tools or side effects that should be constrained, reviewed, or denied by runtime policy.
My model upgrade looks better overall but breaks important casesAverage quality improved, but a critical slice or workflow regressed and the release decision is now ambiguous.
My long-running agent loses state or cannot recover safelyTemporary context, durable state, checkpoints, and expiry are blurred together, making resume behavior unsafe.
My multi-agent system adds complexity without reliabilityDelegation and parallelism added coordination cost, but did not make responsibility, verification, or outcomes clearer.
Showing 120 of 120
Understand AI
34 Guides
Models, Tokens & Generation·FOUNDATION·6 min·Practice availableProbabilistic model behaviorTreat model output as a distribution of plausible continuations, not a deterministic lookup from question to answer.Used in 2 courses · AI Foundations · LLM Application EngineeringRead Guide Models, Tokens & Generation·FOUNDATION·6 min·Practice availableModel claim vs runtime factA model can describe what it thinks happened; only the runtime and external systems can establish what actually happened.Used in 2 courses · AI Foundations · Data Analysis with AIRead Guide Models, Tokens & Generation·FOUNDATION·5 min·Practice availableTokenizationModels process token sequences, not human-visible characters or words, so cost, limits and behavior depend on how text is segmented.Used in 1 courses · AI FoundationsRead Guide Models, Tokens & Generation·FOUNDATION·6 min·Practice availableContext windowThe context window is a finite working-input budget shared by instructions, conversation, retrieved evidence and tool results.Used in 1 courses · AI FoundationsRead Guide Context & Representation·FOUNDATION·5 min·Practice availableRelevance over raw volumeMore input can reduce decision quality when low-value material competes with the evidence that actually matters.Used in 2 courses · AI Foundations · RAG & Knowledge SystemsRead Guide Inference & Model Selection·INTERMEDIATE·6 min·Practice availableModel selection is a workload trade-offChoose models against the workload and its constraints rather than assuming the largest or newest model is automatically the best choice.Used in 3 courses · AI Foundations · LLM Application EngineeringRead Guide AI Safety & Governance·FOUNDATION·6 min·Practice availableCopyright, licensing and provenanceAI-assisted creation needs provenance records that distinguish what came from sources, what was transformed and what may be reused or published.Used in 3 courses · AI Foundations · Create a Course or Knowledge Product with AIRead Guide Training & Post-training·ADVANCED·7 min·Practice availableSynthetic data and distillationSynthetic data can scale examples and transfer behavior, but it also transfers the generator's blind spots unless quality and diversity are controlled.Used in 2 courses · AI Foundations · Model EngineeringRead Guide Training & Post-training·INTERMEDIATE·7 min·Practice availableSupervised fine-tuningSupervised fine-tuning adapts model behavior using curated input–target examples, but it should solve a demonstrated behavior gap rather than replace application design.Used in 2 courses · AI Foundations · Model EngineeringRead Guide Embeddings & Multimodal·FOUNDATION·6 min·Practice availableVector similarity is a retrieval signalVector similarity estimates closeness in an embedding space; it is useful for candidate retrieval but should not be confused with truth, authority or task relevance.Used in 2 courses · AI Foundations · RAG & Knowledge SystemsRead Guide Reasoning & Planning·INTERMEDIATE·6 min·Practice availableVerification over introspectionWhen correctness matters, check the world or an independent artifact instead of asking the model whether its own reasoning was correct.Used in 3 courses · AI Foundations · Data Analysis with AIRead Guide Limits, Uncertainty & Failure·INTERMEDIATE·6 min·Practice availableUncertainty calibrationUseful uncertainty means confidence tracks observed error risk closely enough to change system behavior, escalation and communication.Used in 2 courses · AI Foundations · Data Analysis with AIRead Guide Training & Post-training·FOUNDATION·7 min·Practice availablePretraining objective and learned behaviorPretraining teaches a model to predict patterns in large datasets; useful capabilities emerge from that objective, but product guarantees do not.Used in 2 courses · AI Foundations · Model EngineeringRead Guide Models, Tokens & Generation·FOUNDATION·6 min·Practice availableModel capability envelopeA model has a workload-dependent envelope of tasks it can perform acceptably; outside that tested envelope, confidence should fall before ambition rises.Used in 2 courses · AI Foundations · LLM Application EngineeringRead Guide Models, Tokens & Generation·FOUNDATION·7 min·Practice availableLatency and throughput basicsAI serving performance is a queueing and workload problem: response time, concurrency and token generation compete for finite compute capacity.Used in 2 courses · AI Foundations · LLM Application EngineeringRead Guide Training & Post-training·INTERMEDIATE·7 min·Practice availablePreference post-trainingPreference post-training shapes which behaviors a model tends to choose among plausible continuations, but it does not turn preference signals into hard application guarantees.Used in 2 courses · AI Foundations · Model EngineeringRead Guide Instructions & Prompting·FOUNDATION·6 min·Practice availablePrompt decompositionBreak a complex request into explicit reasoning or production stages when one prompt mixes different responsibilities, evidence needs or verification methods.Used in 2 courses · AI Foundations · Research with AIRead Guide AI Safety & Governance·FOUNDATION·7 min·Practice availablePrivacy and data boundariesAI systems need explicit rules for what data may enter models, logs, retrieval indexes, memory and third-party tools, because each boundary creates a different exposure and retention risk.Used in 2 courses · AI Foundations · Build an AI Knowledge BaseRead Guide Inference & Model Selection·INTERMEDIATE·7 min·Practice availableOpen and closed model trade-offsOpen-weight and hosted closed models offer different control, operational burden, data boundaries and capability economics; the right choice depends on the system constraint.Used in 2 courses · AI Foundations · Model EngineeringRead Guide Models, Tokens & Generation·FOUNDATION·6 min·Practice availableNext-token generationLanguage models produce sequences one token decision at a time; fluent long-form behavior emerges from repeated local predictions conditioned on the growing context.Used in 1 courses · AI FoundationsRead Guide Models, Tokens & Generation·FOUNDATION·6 min·Practice availableSampling controlsTemperature, top-p and related decoding controls reshape how the runtime samples from model probabilities; they change variability, not the model's underlying knowledge or guarantees.Used in 1 courses · AI FoundationsRead Guide Embeddings & Multimodal·FOUNDATION·7 min·Practice availableEmbeddings and semantic spaceEmbeddings encode inputs as vectors whose geometry captures model-specific patterns of similarity, enabling retrieval and clustering without turning distance into meaning, truth or authority.Used in 1 courses · AI FoundationsRead Guide Limits, Uncertainty & Failure·FOUNDATION·7 min·Practice availableHallucination as unsupported generationA hallucination is generated content that sounds usable but lacks adequate support from the system's evidence or constraints; the engineering fix is evidence and verification, not asking the model to feel more certain.Used in 1 courses · AI FoundationsRead Guide Context & Representation·FOUNDATION·6 min·Practice availableContext versus memoryContext is what the model can use in the current call; memory is an application-managed mechanism for preserving state across calls and selectively reintroducing it when relevant.Used in 1 courses · AI FoundationsRead Guide Context & Representation·FOUNDATION·7 min·Practice availableAttention budgetA large context window does not create unlimited useful attention: every extra token competes for model focus, latency and cost, so context must be budgeted around task-relevant evidence.Used in 1 courses · AI FoundationsRead Guide Instructions & Prompting·FOUNDATION·7 min·Practice availableInstruction authorityReliable prompting depends on knowing which instruction source is allowed to govern behavior; system, developer, retrieved context and user text are not interchangeable peers.Used in 1 courses · AI FoundationsRead Guide Instructions & Prompting·FOUNDATION·7 min·Practice availableInstruction conflictInstruction conflict occurs when multiple valid-looking sources imply incompatible actions; reliable systems detect the conflict explicitly instead of letting the model improvise a compromise.Used in 1 courses · AI FoundationsRead Guide Instructions & Prompting·FOUNDATION·7 min·Practice availablePrompt, context and runtime boundaryPrompt wording, supplied context and runtime enforcement solve different problems; reliable AI systems separate persuasion, evidence and capability control instead of treating them as one prompt-design layer.Used in 1 courses · AI FoundationsRead Guide Reasoning & Planning·INTERMEDIATE·7 min·Practice availableReasoning budgetReasoning effort is an engineering budget, not a free quality switch: more search, deliberation or model effort can improve hard tasks while increasing latency, cost and sometimes error exposure.Used in 1 courses · AI FoundationsRead Guide Limits, Uncertainty & Failure·FOUNDATION·7 min·Practice availableCapability is not a guaranteeA model demonstrating a capability in examples or benchmarks does not guarantee that the behavior will occur for every production input; reliability requires contracts, evidence and controls around probabilistic capability.Used in 1 courses · AI FoundationsRead Guide Instructions & Prompting·FOUNDATION·7 min·Practice availablePrompt specificityPrompt specificity makes the task, constraints, output contract and success criteria concrete enough that the model does not have to invent important requirements from vague language.Used in 1 courses · AI FoundationsRead Guide Embeddings & Multimodal·FOUNDATION·7 min·Practice availableMultimodal representationMultimodal systems must convert text, images, audio or video into representations that preserve the information needed for a task; sharing a model does not make every modality equally observable or interchangeable.Used in 1 courses · AI FoundationsRead Guide Limits, Uncertainty & Failure·INTERMEDIATE·7 min·Practice availableDistribution shiftDistribution shift occurs when production inputs, users, tools or environments differ materially from the data and cases used to evaluate a model or AI system, making past quality estimates less reliable.Used in 1 courses · AI FoundationsRead Guide AI Safety & Governance·INTERMEDIATE·7 min·Practice availableBias and fairnessBias and fairness engineering asks whether an AI system produces systematically different errors, opportunities or burdens across relevant groups and contexts, and whether those differences violate the product's intended policy.Used in 1 courses · AI FoundationsRead Guide
Build AI
67 Guides
LLM Application Engineering·FOUNDATION·6 min·Practice availableStructured output is an application contractA JSON shape is useful only when the application validates both syntax and the meaning required by downstream code.Used in 1 courses · LLM Application EngineeringRead Guide Context Engineering·FOUNDATION·6 min·Practice availableFinite context budgetContext engineering is allocation: decide which instructions, evidence, memory and tool state deserve the next token of budget.Used in 1 courses · LLM Application EngineeringRead Guide Context Engineering·INTERMEDIATE·7 min·Practice availableContext managementContext management continuously decides what to load, refresh, compact and remove as a task evolves.Used in 4 courses · LLM Application Engineering · Research with AIRead Guide RAG & Retrieval·INTERMEDIATE·6 min·Practice availableDense, sparse and hybrid retrievalDifferent retrieval signals catch different evidence; hybrid retrieval combines semantic similarity with exact lexical evidence.Used in 1 courses · RAG & Knowledge SystemsRead Guide RAG & Retrieval·INTERMEDIATE·6 min·Practice availableFreshness, authority and source priorityThe most similar passage is not necessarily the source the system should trust.Used in 5 courses · AI Customer Support · Build an AI Knowledge BaseRead Guide Memory Systems·INTERMEDIATE·6 min·Practice availableWorking memory vs durable memoryKeep temporary reasoning context separate from durable state that must survive compaction, retries and new sessions.Used in 1 courses · RAG & Knowledge SystemsRead Guide Agent Runtime·INTERMEDIATE·6 min·Practice availableAct → observe → verify loopAn agent is a control loop that chooses an action, observes the environment and verifies whether the task state actually improved.Used in 3 courses · Agent Engineering · AI Customer SupportRead Guide Agent Runtime·INTERMEDIATE·6 min·Practice availableBounded autonomyGive an agent enough freedom to complete useful work, but place hard boundaries around cost, time, permissions and irreversible effects.Used in 2 courses · Agent Engineering · AI Customer SupportRead Guide Agent Runtime·INTERMEDIATE·7 min·Practice availableWorkflow decompositionBreak a complex goal into stages with explicit inputs, outputs and verification before adding more agent intelligence.Used in 4 courses · Agent Engineering · Multi-Agent & OrchestrationRead Guide Evaluation & Reliability·INTERMEDIATE·6 min·Practice availableTimeout ambiguityA timeout proves that a timely response was not observed; it does not prove that the remote operation failed.Used in 1 courses · Production AI ReliabilityRead Guide Evaluation & Reliability·INTERMEDIATE·6 min·Practice availableIdempotency boundaryRepeated attempts should represent the same business intent instead of silently creating a new side effect each time.Used in 1 courses · Production AI ReliabilityRead Guide Observability·INTERMEDIATE·6 min·Practice availableObservability as a diagnosis interfaceLogs and traces are valuable when they let an operator reconstruct why an AI outcome happened across model, retrieval, tool and runtime layers.Used in 1 courses · Production AI ReliabilityRead Guide Evaluation & Reliability·INTERMEDIATE·7 min·Practice availableEvaluation environments and verifier designAn AI evaluation is only as credible as the environment, task contract and verifier that determine what counts as success.Used in 6 courses · Agent Engineering · Build an AI Knowledge BaseRead Guide Tools & Function Calling·INTERMEDIATE·6 min·Practice availableHuman review boundariesHuman review is useful when it owns a concrete decision at a consequential boundary, not when a person merely glances at AI output.Used in 5 courses · Agent Engineering · AI Customer SupportRead Guide Vibe Coding & Agentic Coding·FOUNDATION·6 min·Practice availableSpecify before you generateAI-assisted implementation becomes controllable when the intended behavior and acceptance evidence are defined before code is generated.Used in 2 courses · Create a Course or Knowledge Product with AI · Vibe Coding & Agentic Software EngineeringRead Guide Evaluation & Reliability·INTERMEDIATE·6 min·Practice availableConfidence, variance and sample-size humilityAI evaluation results are samples from variable systems, so confidence requires repeated observations, uncertainty estimates and enough data for the decision.Used in 2 courses · Data Analysis with AI · Production AI ReliabilityRead Guide Tools & Function Calling·INTERMEDIATE·6 min·Practice availableValidate tool results before using themA successful tool call only means a response arrived; the runtime still has to validate its shape, semantics and relationship to the intended action.Used in 2 courses · Agent Engineering · AI Customer SupportRead Guide Production Architecture·INTERMEDIATE·6 min·Practice availableVersioned dependencies and change boundariesAI behavior depends on models, prompts, tools, schemas and retrieval sources that change independently, so production systems need explicit version boundaries.Used in 2 courses · Production AI Reliability · Vibe Coding & Agentic Software EngineeringRead Guide Security & Human Control·INTERMEDIATE·6 min·Practice availableAuditability as durable decision evidenceAn auditable AI system preserves enough durable evidence to reconstruct what happened, which rules applied and who or what authorized the consequential action.Used in 2 courses · AI Customer Support · Production AI ReliabilityRead Guide RAG & Retrieval·INTERMEDIATE·7 min·Practice availableRetrieval is a pipeline, not one search callReliable RAG separates candidate generation, filtering, reranking, context assembly and provenance so each failure can be measured and repaired.Used in 3 courses · Build an AI Knowledge Base · RAG & Knowledge SystemsRead Guide Tools & Function Calling·INTERMEDIATE·7 min·Practice availableTool contracts for agentsA tool contract defines what an agent may request, what the runtime guarantees, what can fail and which side effects require stronger controls.Used in 2 courses · Agent Engineering · AI Customer SupportRead Guide RAG & Retrieval·INTERMEDIATE·6 min·Practice availableEvaluate retrieval before generationRetrieval evaluation asks whether the system surfaced the necessary evidence, at usable rank and under the right authority constraints, before judging the generated answer.Used in 2 courses · Build an AI Knowledge Base · RAG & Knowledge SystemsRead Guide Vibe Coding & Agentic Coding·ADVANCED·6 min·Practice availableParallel agent workParallelism helps when subproblems are genuinely independent or provide independent verification; otherwise coordination cost can exceed the speedup.Used in 2 courses · Multi-Agent & Orchestration · Vibe Coding & Agentic Software EngineeringRead Guide Tools & Function Calling·INTERMEDIATE·6 min·Practice availableLeast privilege for AI capabilitiesGive an AI system only the permissions required for the current task, for the shortest practical scope, with stronger gates around irreversible effects.Used in 2 courses · Agent Engineering · AI Customer SupportRead Guide Agent Runtime·INTERMEDIATE·6 min·Practice availableVerification loopsReliable agents repeatedly compare intended state with observed state, then use the gap to decide whether to continue, repair, escalate or stop.Used in 1 courses · Agent EngineeringRead Guide Security & Human Control·INTERMEDIATE·6 min·Practice availableTrust boundaries in AI systemsA trust boundary marks where data, instructions or actions cross between components with different authority, ownership or reliability assumptions.Used in 1 courses · Production AI ReliabilityRead Guide Evaluation & Reliability·INTERMEDIATE·6 min·Practice availableTraceability from input to outcomeTraceability connects the user's request, selected evidence, model decisions, tool actions and final outcome so failures can be reconstructed across layers.Used in 1 courses · Production AI ReliabilityRead Guide Evaluation & Reliability·INTERMEDIATE·7 min·Practice availableSlices and regression suitesAggregate quality is not enough: release evaluation needs stable slices that represent important users, tasks and failure modes, plus regression tests that prevent solved problems from returning.Used in 1 courses · Production AI ReliabilityRead Guide Memory Systems·INTERMEDIATE·7 min·Practice availableMemory lifecycleDurable AI memory needs admission, provenance, update, retrieval and deletion rules; persistence without a lifecycle only makes stale mistakes harder to remove.Used in 2 courses · Build an AI Knowledge Base · RAG & Knowledge SystemsRead Guide Multi-Agent & Orchestration·INTERMEDIATE·7 min·Practice availableOrchestration topology trade-offsLoops, graphs, supervisors and parallel workers solve different coordination problems; topology should be chosen from dependency structure rather than architectural fashion.Used in 1 courses · Multi-Agent & OrchestrationRead Guide Memory Systems·INTERMEDIATE·6 min·Practice availableMemory expiry and invalidationDurable memory should stop influencing decisions when its evidence becomes stale, its scope ends or stronger information supersedes it.Used in 2 courses · Build an AI Knowledge Base · RAG & Knowledge SystemsRead Guide Agent Runtime·INTERMEDIATE·7 min·Practice availableCheckpoint and resumeLong-running agents become recoverable when they persist explicit checkpoints that separate completed durable work from transient model state and define how execution can resume safely.Used in 1 courses · Agent EngineeringRead Guide Agent Runtime·ADVANCED·7 min·Practice availableAsynchronous long-running workTasks that outlive one request need explicit job identity, durable state, cancellation and result retrieval; keeping an HTTP or model call open is not a long-running architecture.Used in 1 courses · Agent EngineeringRead Guide Tools & Function Calling·INTERMEDIATE·6 min·Practice availableReversible actionsAgent actions are safer when the system prefers operations that can be previewed, undone or compensated before granting irreversible authority.Used in 1 courses · Agent EngineeringRead Guide Context Engineering·INTERMEDIATE·7 min·Practice availableContext compactionContext compaction preserves the smallest task-relevant state needed for future decisions instead of repeatedly carrying the full interaction history forward.Used in 2 courses · LLM Application Engineering · Write a Book with AIRead Guide LLM Application Engineering·INTERMEDIATE·7 min·Practice availableStreaming and backpressureStreaming improves perceived latency only when producers, consumers and downstream parsers agree on flow control; otherwise fast token production can overwhelm slower clients or application work.Used in 1 courses · LLM Application EngineeringRead Guide Evaluation & Reliability·INTERMEDIATE·7 min·Practice availableRetry amplificationA retry intended to improve reliability can multiply load and duplicate side effects when several layers retry the same ambiguous failure without shared budgets or idempotency.Used in 1 courses · Production AI ReliabilityRead Guide Security & Human Control·INTERMEDIATE·8 min·Practice availablePrompt injection defensePrompt injection is untrusted content attempting to influence model behavior; defense comes from capability boundaries, source labeling and runtime enforcement rather than a stronger instruction telling the model to ignore attacks.Used in 1 courses · Production AI ReliabilityRead Guide Observability·INTERMEDIATE·7 min·Practice availableOnline monitoringOffline evaluation predicts release quality; online monitoring checks whether real traffic, dependencies and user behavior continue to match those assumptions after deployment.Used in 1 courses · Production AI ReliabilityRead Guide Production Architecture·ADVANCED·7 min·Practice availableRollout and fallbackAI releases are safer when exposure increases gradually and the system has a tested fallback that preserves critical product behavior when the new model, prompt or dependency fails.Used in 1 courses · Production AI ReliabilityRead Guide RAG & Retrieval·INTERMEDIATE·7 min·Practice availableReranking retrieved candidatesReranking is a second selection stage that spends more computation on a smaller candidate set to order evidence by task relevance after broad retrieval maximizes recall.Used in 1 courses · RAG & Knowledge SystemsRead Guide Multi-Agent & Orchestration·ADVANCED·7 min·Practice availableCoordination overheadMulti-agent systems pay real costs for handoffs, duplicated context, conflict resolution and verification; parallelism is valuable only when independent work saves more than coordination consumes.Used in 1 courses · Multi-Agent & OrchestrationRead Guide Vibe Coding & Agentic Coding·INTERMEDIATE·7 min·Practice availableTest-first AI codingAI coding is safer when executable expectations exist before generation, because tests give the agent and reviewer an external contract that does not depend on the generated code explaining itself.Used in 1 courses · Vibe Coding & Agentic Software EngineeringRead Guide Vibe Coding & Agentic Coding·INTERMEDIATE·7 min·Practice availableAI code reviewReviewing AI-generated code means independently checking behavior, scope, security and maintainability against repository evidence rather than accepting a plausible diff or the agent's own explanation.Used in 1 courses · Vibe Coding & Agentic Software EngineeringRead Guide LLM Application Engineering·INTERMEDIATE·7 min·Practice availableConversation stateConversation state is application-owned structured state about what the interaction is doing now; it should not be confused with raw chat history, model memory or whatever happens to fit in the context window.Used in 1 courses · LLM Application EngineeringRead Guide LLM Application Engineering·INTERMEDIATE·7 min·Practice availableModel routing and fallbackProduction LLM applications should choose models by task, risk and current system conditions, with explicit fallbacks when the preferred route is unavailable, too slow, too expensive or insufficiently reliable.Used in 1 courses · LLM Application EngineeringRead Guide RAG & Retrieval·INTERMEDIATE·7 min·Practice availableEvidence granularityRAG quality depends on retrieving evidence at a granularity that preserves meaning while remaining selective enough for ranking and context budgets; chunk size is therefore a decision, not a magic constant.Used in 1 courses · RAG & Knowledge SystemsRead Guide Memory Systems·INTERMEDIATE·7 min·Practice availableMemory authority conflictDurable memory can be useful and still be wrong for the current decision; when remembered facts conflict with fresher or more authoritative evidence, the system needs an explicit authority and update policy.Used in 1 courses · RAG & Knowledge SystemsRead Guide MCP & Capability Interfaces·INTERMEDIATE·7 min·Practice availableMCP boundariesMCP connects models to external capabilities, but the protocol boundary should not be mistaken for a permission boundary; clients and runtimes still need explicit trust, capability and side-effect controls.Used in 1 courses · Agent EngineeringRead Guide MCP & Capability Interfaces·INTERMEDIATE·7 min·Practice availableMCP capability negotiationCapability negotiation lets an MCP client expose only the features and operations that both sides support and the current task is allowed to use, rather than assuming every available capability is active.Used in 1 courses · Agent EngineeringRead Guide Agent Runtime·INTERMEDIATE·7 min·Practice availableInterrupt and cancel long-running agentsLong-running agents need explicit interrupt and cancellation semantics so users and operators can stop work safely, distinguish 'stop requested' from 'stopped', and recover without corrupting side effects.Used in 1 courses · Agent EngineeringRead Guide Multi-Agent & Orchestration·INTERMEDIATE·7 min·Practice availableLoop versus graphUse an agent loop when one controller can own iteration and state; introduce a graph only when explicit branches, joins, specialized responsibilities or independent recovery boundaries create measurable value.Used in 1 courses · Multi-Agent & OrchestrationRead Guide Security & Human Control·FOUNDATION·7 min·Practice availableRuntime enforcementRuntime enforcement turns policy from advice into executable constraints by controlling which actions, data and side effects can actually pass through the system, even when the model proposes something unsafe.Used in 1 courses · Production AI ReliabilityRead Guide Production Architecture·ADVANCED·7 min·Practice availableSHIP / BLOCK / INCONCLUSIVERelease evaluation should end in an explicit decision state: SHIP when evidence clears required gates, BLOCK when a veto or unacceptable regression is present, and INCONCLUSIVE when evidence is too weak to decide safely.Used in 1 courses · Production AI ReliabilityRead Guide Model Engineering·ADVANCED·7 min·Practice availableAdapter fine-tuningAdapter fine-tuning changes a small trainable component while keeping most base-model weights fixed, making targeted behavior adaptation cheaper and easier to version than full fine-tuning when a measured task gap justifies training.Used in 1 courses · Model EngineeringRead Guide Vibe Coding & Agentic Coding·INTERMEDIATE·7 min·Practice availableRepository contextRepository context is the code, conventions, tests, dependency boundaries and local history an AI coding agent needs to make a change that fits the actual system rather than producing plausible isolated code.Used in 1 courses · Vibe Coding & Agentic Software EngineeringRead Guide Vibe Coding & Agentic Coding·INTERMEDIATE·7 min·Practice availableDependency and migration verificationDependency upgrades and data or schema migrations require verification across compatibility, rollout and rollback boundaries because a locally correct code change can fail when versions or persisted state differ in production.Used in 1 courses · Vibe Coding & Agentic Software EngineeringRead Guide LLM Application Engineering·INTERMEDIATE·7 min·Practice availableAPI rate-limit resilienceRate-limit resilience keeps an LLM application useful when providers or internal services enforce quotas by coordinating concurrency, backoff, queues, budgets and fallbacks instead of turning every 429 into an uncontrolled retry storm.Used in 1 courses · LLM Application EngineeringRead Guide RAG & Retrieval·INTERMEDIATE·7 min·Practice availableMetadata retrievalMetadata retrieval uses structured fields such as tenant, document type, date, authority or product to constrain and rank evidence before semantic similarity is allowed to dominate the result set.Used in 1 courses · RAG & Knowledge SystemsRead Guide Search & Knowledge Graphs·INTERMEDIATE·7 min·Practice availableKnowledge graph structureA knowledge graph represents entities and typed relationships explicitly so retrieval and reasoning can follow meaningful connections that are difficult to recover from text similarity alone.Used in 1 courses · RAG & Knowledge SystemsRead Guide Computer & Browser Use·INTERMEDIATE·7 min·Practice availableAction groundingAction grounding requires an agent to tie proposed actions to current observable environment state, valid targets and verified preconditions instead of acting from stale assumptions or model-generated descriptions.Used in 1 courses · Agent EngineeringRead Guide Computer & Browser Use·INTERMEDIATE·7 min·Practice availableSandbox permissionsSandbox permissions bound what code or an agent can read, write, execute and reach on the network so experiments and generated actions cannot silently inherit the full authority of the host environment.Used in 1 courses · Agent EngineeringRead Guide Multi-Agent & Orchestration·ADVANCED·7 min·Practice availableDelegation stateDelegation state defines what information, ownership and progress move from a parent agent or workflow to delegated workers and what evidence must return before the parent can safely continue.Used in 1 courses · Multi-Agent & OrchestrationRead Guide Evaluation & Reliability·INTERMEDIATE·7 min·Practice availableCompensation and recoveryCompensation and recovery handle workflows where an earlier side effect succeeded but a later step failed, using reconciliation, compensating actions and durable state instead of pretending the whole transaction can simply be retried.Used in 1 courses · Production AI ReliabilityRead Guide Evaluation & Reliability·ADVANCED·7 min·Practice availableRelease economicsRelease economics treats latency, inference cost, operational burden and expected failure loss as part of the release decision alongside quality, because a model improvement that cannot be served sustainably is not a production win.Used in 1 courses · Production AI ReliabilityRead Guide Observability·INTERMEDIATE·7 min·Practice availableFailure attributionFailure attribution identifies which layer—model, prompt, context, retrieval, tool, state, orchestration or runtime—most likely caused an observed bad outcome so the fix targets the mechanism instead of the symptom.Used in 1 courses · Production AI ReliabilityRead Guide Model Engineering·ADVANCED·7 min·Practice availableFine-tuning dataset designFine-tuning dataset design turns a measured behavior gap into representative training examples with clear labels, coverage and provenance while keeping evaluation cases separate enough to tell whether training actually generalized.Used in 1 courses · Model EngineeringRead Guide
Use AI
19 Guides
Research with AI·INTERMEDIATE·6 min·Practice availableTriangulate independent sourcesConfidence should increase when independent evidence paths converge, not when many links repeat the same upstream claim.Used in 2 courses · Research with AI · Write a Book with AIRead Guide Long-Form Creation·INTERMEDIATE·7 min·Practice availableLong-form consistencyCoherent long documents externalize claims, terminology, voice and decisions instead of relying on one giant prompt to remember everything.Used in 1 courses · Write a Book with AIRead Guide Work Automation·FOUNDATION·6 min·Practice availableChoose the automation boundaryAutomate stable, observable work and keep humans where ambiguity, accountability or exception cost remains high.Used in 1 courses · Run a Solo Business with AIRead Guide Documents & Data Analysis·FOUNDATION·6 min·Practice availableDocument extraction as evidence-preserving transformationDocument extraction should convert messy source material into structured evidence without erasing location, provenance or uncertainty.Used in 2 courses · Build an AI Knowledge Base · Data Analysis with AIRead Guide Long-Form Creation·INTERMEDIATE·6 min·Practice availableEditorial voice controlConsistent voice comes from explicit editorial constraints and review artifacts, not from repeatedly asking a model to 'sound like us'.Used in 2 courses · Create a Course or Knowledge Product with AI · Write a Book with AIRead Guide AI Knowledge Base·INTERMEDIATE·7 min·Practice availableKnowledge base lifecycleA useful AI knowledge base is a maintained evidence system with ingestion, authority, update, retirement and evaluation policies.Used in 3 courses · AI Customer Support · Build an AI Knowledge BaseRead Guide Media Creation·FOUNDATION·6 min·Practice availableVisual briefing for generated mediaA strong visual brief turns an aesthetic intention into constraints that can be generated, reviewed and revised consistently.Used in 1 courses · Create a Course or Knowledge Product with AIRead Guide Research with AI·FOUNDATION·6 min·Practice availableResearch question decompositionA strong research workflow converts a broad question into smaller claims and evidence needs before searching, so synthesis is driven by what must be established rather than by whichever sources appear first.Used in 2 courses · Research with AI · Run a Solo Business with AIRead Guide Media Creation·INTERMEDIATE·6 min·Practice availableMultimodal iterationText, image, audio and other modalities need iterative review against cross-modal meaning and production constraints, not independent generation followed by assembly.Used in 1 courses · Create a Course or Knowledge Product with AIRead Guide Support & Operations·INTERMEDIATE·7 min·Practice availableCustomer support copilotA support copilot should increase operator leverage by retrieving evidence, structuring cases and drafting actions while keeping source authority and consequential decisions explicit.Used in 2 courses · AI Customer Support · Run a Solo Business with AIRead Guide Courses & Knowledge Products·FOUNDATION·7 min·Practice availableLearning objective designA useful learning objective states what the learner should be able to decide, explain or perform with observable evidence, rather than naming a topic the course intends to cover.Used in 1 courses · Create a Course or Knowledge Product with AIRead Guide Long-Form Creation·INTERMEDIATE·7 min·Practice availableFact-check and revisionAI-assisted long-form work becomes trustworthy when claims are extracted, checked against authoritative sources and revised with provenance instead of asking the same model to 'fact-check itself'.Used in 1 courses · Write a Book with AIRead Guide Documents & Data Analysis·INTERMEDIATE·7 min·Practice availableVerification for AI-assisted data analysisAI can accelerate extraction, transformation and interpretation, but every important numerical conclusion should be reproducible from explicit data and calculations outside the model's narrative.Used in 1 courses · Data Analysis with AIRead Guide Courses & Knowledge Products·INTERMEDIATE·7 min·Practice availableCurriculum decompositionCurriculum decomposition turns learning objectives into an ordered set of teachable decisions, concepts and practices, preserving prerequisites and transfer instead of generating a pile of loosely related lessons.Used in 1 courses · Create a Course or Knowledge Product with AIRead Guide Work Automation·INTERMEDIATE·7 min·Practice availableAutomation maintenanceAutomation creates leverage only when its monitoring, exceptions, dependency drift and repair cost stay lower than the work it removes; maintenance is part of automation ROI, not an afterthought.Used in 1 courses · Run a Solo Business with AIRead Guide Long-Form Creation·FOUNDATION·7 min·Practice availableResearch to outlineResearch-to-outline turns collected evidence into a structure of claims, questions and sections so long-form generation follows what the evidence can support instead of inventing a narrative first and searching for citations afterward.Used in 1 courses · Write a Book with AIRead Guide Research with AI·INTERMEDIATE·7 min·Practice availableClaim–evidence matrixA claim–evidence matrix maps important conclusions to supporting and contradicting sources, provenance, freshness and confidence so research can be audited before it is compressed into a narrative.Used in 1 courses · Research with AIRead Guide Media Creation·INTERMEDIATE·7 min·Practice availableMedia quality reviewMedia quality review evaluates generated images, audio, video and presentation assets against instructional purpose, factual accuracy, accessibility, brand constraints and technical delivery rather than accepting output because it looks polished.Used in 1 courses · Create a Course or Knowledge Product with AIRead Guide Product, Marketing & Growth·FOUNDATION·7 min·Practice availableCustomer problem researchCustomer problem research gathers evidence about recurring user pain, current workarounds, consequences and willingness to change before a solo builder automates or markets a solution.Used in 1 courses · Run a Solo Business with AIRead Guide