Full catalog
Search 中文
More
AI VISTA / SITE NAVIGATION

Courses and topics open directly from the main navigation.

Open the full content catalog

Courses

4
  1. 01Build your first testable AI workflow5 lessons
  2. 02Make AI answers reliable enough to use14 lessons
  3. 03Design answers that can show their evidence11 lessons
  4. 04Automate with permission, recovery, and cost controls37 lessons
View all courses →

Topics

9
  1. 01Task framing4 guides
  2. 02Context and prompts6 guides
  3. 03Evaluation7 guides
  4. 04Grounded answers11 guides
  5. 05Agent architecture12 guides
  6. 06Agent safety5 guides
  7. 07Local model deployment7 guides
  8. 08Model operations7 guides
  9. 09Field cases4 guides
View all topics →

DECISION MAP / NINE CRITICAL CHOICES

Name the symptom. Find the decision underneath.

Do not start with a tool name or trend. Locate the layer where the job is stuck, then leave with a decision tool you can use.

See all 9 topics ↓Browse the full catalog →
AI VISTA / ROUTING SHEET01—03
SYMPTOMLAYERARTIFACT
  1. 01Should this task use AI?→Task framing→Fit matrix
  2. 02Why is the answer unstable?→Context / evaluation→Failure tree
  3. 03How far may automation go?→Agent safety→Permission ladder
?ENTER BY QUESTION
INPUTDECISIONARTIFACTNO GLOSSARY ROUTES

ALL TOPICS / NINE DECISIONS

Start with a topic or the problem you see.

Each topic collects every published guide for one kind of decision. Choose the question that matters now, or browse every article in the catalog.

See every article →
01Task framingWhat exactly should AI do—and should it do it at all?You have not decided whether AI belongs4 guides02Context and promptsWhat must the model see and return for the next step to work?Output quality changes across inputs6 guides03EvaluationWhat evidence would make this workflow safe to expand?You cannot locate the failing layer7 guides04Grounded answersWhen does external evidence improve the answer enough to justify a retrieval system?Answers must show evidence11 guides05Agent architectureWho owns each agent step, and how can a reviewer verify what actually happened?Work now needs delegation, tool calls, or durable memory12 guides06Agent safetyHow much authority can this agent safely hold?An agent will change external state5 guides07Local model deploymentCan this model serve our workload on local hardware, and when is that deployment worth it?You must size GPU memory and choose local or cloud serving7 guides08Model operationsWhich model makes this workflow dependable and affordable in operation?You need a model and an operating cost boundary7 guides09Field casesWhat actually changes when an AI workflow meets a real review, a source update, or a payment?The case reads well, but the operating boundary is unclear4 guides

LEARNING ROUTE / STAGE 02 · DECISIONS

This stage turns a vague symptom into a question you can learn and test.

You do not need a model name, technical term, or course title. Describe where the work gets stuck and the map routes you to the most relevant decision.

BRING
The clearest symptom in one real job
DO HERE
Locate the earliest decision layer
LEAVE WITH
A topic entry and a course direction
WHEN THIS STAGE IS DONE Continue to courses Choose the syllabus closest to the problem.

HOW ONE JOB MOVES THROUGH THE MAP

One support-routing workflow creates four different decisions in sequence.

The map does not assign a permanent category. It finds the earliest decision worth resolving now. If that decision stays vague, evaluation and automation inherit the wrong assumptions.

AI VISTA / WORKED ROUTEEXAMPLE: ROUTE SUPPORT EMAILS
  1. 01
    Task framing

    Set the task boundary first

    Classify support emails, but never reply. Low-confidence cases must go to the support lead.

    ↗
  2. 02
    Context and prompts

    List what the model must see

    Provide only the email, approved queue definitions, and urgency rules—not unrelated customer data.

    ↗
  3. 03
    Evaluation

    Test whether it actually works

    Use ordinary, ambiguous, cross-queue, and high-risk emails; record the label, reason, confidence, and human override.

    ↗
  4. 04
    Agent safety

    Decide how far automation may act

    Begin with suggestions only. Consider automatic routing only after acceptance passes, with recall and human takeover.

    ↗
RULE OF USE

Stop at the earliest unresolved step. Produce that topic’s artifact before deciding whether to move deeper.

NINE PROBLEMS / FIELD REFERENCE

Start from a signal, prepare enough evidence, then enter the matching course.

This is not a maturity score. It identifies the earliest unresolved decision and makes clear what to bring and what to produce before moving on.

SIGNALEVIDENCE TO BRINGARTIFACT TO LEAVENEXT COURSE
  1. 01
    Task framingThe goal is still ‘improve efficiency’ and no exact AI action or next-step owner is named ↗

    Current human steps, real inputs, and failure consequences

    Fit matrix, task brief, and control-loop decisionOpen course →
  2. 02
    Context and promptsThe same prompt omits fields or resolves conflicts differently across inputs ↗

    Complete input, source precedence, and required output fields

    Context-flow map, acceptance contract, and injection test setOpen course →
  3. 03
    EvaluationThe team can only say output is ‘mostly fine’ and cannot isolate failing cases ↗

    Real task slices, failure labels, and severity

    Evaluation-set blueprint and diagnostic treeOpen course →
  4. 04
    Grounded answersConsequential answers need evidence, but citations may not support the claim or may be stale ↗

    Authoritative sources, version, effective date, and access scope

    RAG decision gate, retrieval trial, and citation-freshness auditOpen course →
  5. 05
    Agent architectureWork is delegated but tool and shared-state ownership is unclear ↗

    Task dependencies, tool effects, run budgets, and memory sources

    Delegation map, tool contract, loop guard, and memory lifecycleOpen course →
  6. 06
    Agent safetyThe system is about to write, notify, or change external state ↗

    Actions, consequences, approval scope, and failure states

    Permission ladder, human gates, and recovery runbookOpen course →
  7. 07
    Local model deploymentThe model file appears to fit, but peak load and quality are unknown ↗

    Model, precision, context, concurrency, real cases, and cloud baseline

    VRAM budget, quantization trial, load record, and deployment matrixOpen course →
  8. 08
    Model operationsModel choice relies on a leaderboard and unit price while operating cost is unknown ↗

    Task cases, quality floor, latency, retries, and human cost

    Model scorecard, cost map, cache trial, and drift watchboardOpen course →
  9. 09
    Field casesThe cases look plausible, but the team cannot name which step to change ↗

    Your own actions, consequences, evidence, and approvers

    Case comparison sheet, permission ladder, and recovery drill recordOpen course →

A USEFUL BOUNDARY

First check whether the problem needs a system at all.

If one deterministic fact is missing, look it up first. Move to a more complex path only when the job needs judgment, context, or a sequence of actions.

THREE QUICK CHECKS01—03
  1. 01
    Look it up firstIs the answer already in a reliable source?
    →
  2. 02
    Check the variationDoes changing input require ongoing judgment?
    →
  3. 03
    Set the boundaryCan the work stop and hand off when it fails?
    ✓
Search published guides↗