Skip to content
NTPU AI Learn

AI 103 · Module 5

Knowledge Architecture & Agent Skills

From "having data" to "the agent knows how to use it"

Start Section 1
  • Duration55 min
  • LevelCore
  • PrerequisiteAI 102 recommended, plus Modules 1 to 4 of this course
0 / 6 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor

Info

After you join, your completed lesson numbers are sent to your instructor together with your learner code, so they can see your progress. Use the codes your instructor gave you, and do not enter your name, company secrets or other personal data. Progress is kept with the rest of that session's data.

Learning objectives

  • Distinguish Knowledge (what to know) from Skills (how to perform) in one sentence
  • Describe sources, layers, and structured and unstructured knowledge, and judge trust by source authority and freshness
  • Explain an Agent Skill's purpose, boundaries, reuse and composition
  • Complete an Agent Skill Card using the seven parts from Purpose to Verification

The problem and the core distinction

10 min

Does having a knowledge base mean the agent can do the work?

The earlier modules let an agent find the right small piece of knowledge. But a new hire who receives the full company handbook is not yet able to handle complaints. Knowing "what the rules are" and knowing "how to do the job" are different things.

Core question

Does having a Knowledge Base mean the agent can do the work? The answer is no.

Knowledge: what to know

  • Policies, SOPs, domain knowledge
  • Examples and historical cases
  • Live data
  • Answers "What to know"

Skills: how to perform

  • The method and steps for a class of task
  • What information is needed and which tools to use
  • What format to output and how to self-check
  • Answers "How to perform a task"

Info

An agent needs both: Knowledge provides the basis, Skills provide the method. Without Skills, an agent often has the data but does not know how to use it.

Key takeaway: Knowledge answers "what to know" and Skills answer "how to do it"; neither is enough alone.

Knowledge architecture

10 min

Knowledge architecture: sources, layers and types

Knowledge Architecture is the overall plan for the knowledge an agent needs: which sources it comes from, which layers it has, what form it is stored in, and who maintains it.

Common knowledge types (simulated examples)
Knowledge typeExampleRate of changeCommon form
PoliciesLeave policy, data-use rulesLow to mediumDocuments (unstructured)
SOPsComplaint handling procedureMediumDocuments (unstructured)
Domain KnowledgeIndustry terms, regulatory backgroundLowDocuments, glossaries
ExamplesA model report, a good replyLowDocuments, templates
Historical CasesPast complaint cases and outcomesKeeps growingTables, records (structured)
Live DataStock, today's prices, order statusVery highDatabase or system data (structured)

Structured Knowledge has fixed fields and suits precise queries and filters; Unstructured Knowledge such as documents and articles relies on chunking and retrieval. In practice knowledge is often layered: stable principles on top, and fast-changing data below, fetched live.

Info

Different types of knowledge suit different ways of storing and obtaining them. Fast-changing live data should not be copied into a document that is rarely updated.
Check yourself

Which of these three looks most like a Skill rather than Knowledge?

Which of these three looks most like a Skill rather than Knowledge?

Key takeaway: Knowledge comes in types with different rates of change; choose how to store and obtain each accordingly.

10 min

Trusted Sources, Source Authority and Knowledge Freshness

Not everything in a knowledge base is equally trustworthy. An agent needs to judge which one to believe, and a person must design that rule in advance.

Three questions for judging a knowledge source
ConceptQuestionExample
Trusted SourceIs this source authorized and reviewed?An official policy from HR outweighs a forwarded screenshot in a chat group
Source AuthorityWhen two sources conflict, which prevails?An official notice outranks meeting minutes, which outrank personal notes
Knowledge FreshnessHow current is it and how often must it update?A price list updates daily; an organization charter may not change for years

For each kind of knowledge, name an owner, an update frequency and a priority order for conflicts. Written down, these rules give the agent a basis and tell people whom to ask when something goes wrong.

Warning

Knowledge that is of unknown origin, out of date or contradictory is more dangerous than none, because an agent will use it in a confident tone. A person must verify the source of any high-impact conclusion.
Check yourself

In the knowledge base, an official notice and an old meeting slide say different things about the same rule. What is the best approach?

In the knowledge base, an official notice and an old meeting slide say different things about the same rule. What is the best approach?

Key takeaway: Assign trusted sources, an authority order and update rules to knowledge; a person settles conflicts.

Agent Skills

10 min

Agent Skills: reusable ways of doing work

An Agent Skill organizes "how to complete a class of task" into a reusable design. A good Skill gives different people, on different occasions, a steady and consistent method instead of rethinking a prompt every time.

  • Reusable Skills: usable again and again for one class of task, for example "complaint analysis"
  • Skill Instructions: explicit steps, decision criteria and output requirements
  • Skill Resources: the knowledge, templates and tools the Skill needs
  • Skill Boundaries: what it can do, what it cannot, and when to hand over to a person
  • Skill Composition: small Skills combine into larger flows, for example "gather data" plus "analyze" plus "write a summary"
A boundary example (simulated)

A "complaint analysis Skill" can organize cases, summarize causes and draft a reply; but for a decision involving compensation, the boundary is "hand over to a person for approval", and the Skill may not conclude on its own.

Warning

A Skill without boundaries tends to overstep. Writing down "what it cannot do" and "when to hand over to a person" makes Human–AI–Human concrete: a person defines the task, AI assists, and a person reviews and stays responsible.
Check yourself

A "complaint analysis Skill" is designed to approve refund amounts automatically. What is the biggest problem?

A "complaint analysis Skill" is designed to approve refund amounts automatically. What is the biggest problem?

Key takeaway: A Skill is a reusable method with clear boundaries and points for human review.

10 min

Skill structure and the Agent Skill Card

A complete Skill can be described in seven parts. An Agent Skill Card arranges those seven parts on a card a person can review.

Agent Skill structure
  1. Purpose: what the Skill is meant to accomplish
  2. Instructions: steps and decision criteria
  3. Required Context: what information is needed to begin
  4. Knowledge: which knowledge and sources it relies on
  5. Tools: which tools it may use (if any)
  6. Output Format: the format and content of the output
  7. Verification: how the result is checked, and who reviews it
Agent Skill Card example (simulated: complaint analysis)
PartContent
PurposeOrganize one complaint, summarize causes and draft a suggested reply
InstructionsConfirm the facts, check the policy, then classify severity and draft a reply
Required ContextThe complaint text (de-identified), order status, service tone guidelines
KnowledgeThe current refund policy, the complaint SOP, similar past cases
ToolsOrder lookup (read-only)
Output FormatFact summary, causes, suggested reply, items needing human review
VerificationCheck the policy version and cited sources; a person approves any compensation
  1. Pick a Skill
    Choose from complaint analysis, ESG report review, investment research, healthcare document summarization or course material preparation, or use your own simulated scenario.
  2. Fill in the seven parts
    Complete each from Purpose to Verification and mark the boundaries.
  3. Mark human review points
    State which outputs a person must confirm before they are used.
  4. Check for sensitive information
    In practice, use simulated data only. Do not enter company secrets or client personal data.
Check yourself

A Skill Card has Purpose, Instructions and Output Format but no Verification. What is the biggest gap?

A Skill Card has Purpose, Instructions and Output Format but no Verification. What is the biggest gap?

Key takeaway: A Skill Card states method, basis and review points in seven parts, Purpose to Verification.

Recap

5 min

Recap

Done

Knowledge tells an agent what to know; Skills tell it how to perform. Both need trusted sources, clear boundaries and human review.
Knowledge and Skills at a glance
KnowledgeSkills
AnswersWhat to knowHow to perform a task
ExamplesPolicies, SOPs, cases, live dataComplaint analysis, report review, research flow
Watch forSource authority and freshnessBoundaries, output format and verification
OwnerThe knowledge owner and maintainerThe Skill designer and reviewer

An agent now has Memory, Knowledge and Skills. But more information does not mean better performance. The next module asks: with so much information, which should be handed to the agent, and when?

Key takeaway: Knowledge and Skills complement each other, and both need trusted sources, boundaries and human review.

Save as PDF

What next?