Skip to content

Knowledge Base

Agent Skills

cl24s/knowledge-base

Agent Skills¶

Agent Skills are reusable capability packages that help an AI system perform a defined class of tasks consistently. A skill may contain instructions, domain knowledge, examples, tools, typed contracts, constraints, and evaluation criteria.

A useful mental model is:

Skill
├── instructions
├── knowledge / context rules
├── tools
├── input contract
├── output contract
├── examples
├── constraints / permissions
└── evaluation

A skill is therefore broader than a prompt, but smaller and more focused than a complete agent.

The exact term skill is not universal across AI platforms. In this knowledge base it is used as an engineering abstraction for a reusable task-level capability and its contract.

Roadmap¶

1. Skill Mental Model ¶

Understand what a skill is and separate prompt vs skill vs tool vs workflow vs agent.

Status: first detailed draft complete.

2. Skill Anatomy and Contracts ¶

Break a skill into its reusable parts: purpose, instructions, knowledge, tools, inputs, outputs, constraints, and success criteria.

Status: first detailed draft complete.

3. Instructions, Knowledge and Examples ¶

Decide what should be packaged with a skill and what should arrive dynamically at runtime through context, retrieval, or tool results.

Status: first detailed draft complete.

4. Tool-Backed Skills ¶

Design skills that can request external actions while keeping tool execution, permissions, validation, and side effects under application control.

Status: first detailed draft complete.

5. Typed Inputs and Structured Outputs ¶

Define machine-readable skill contracts and map LLM-facing DTOs into application/domain objects without coupling the domain model to the model provider.

Status: first detailed draft complete.

6. Skill Discovery, Selection and Routing ¶

Make skills discoverable and help a runtime choose the right capability using descriptions, metadata, routing logic, confidence, and escalation.

Status: first detailed draft complete.

7. Skill Composition and Reuse ¶

Compose focused capabilities without creating giant skills. Cover chaining, shared primitives, dependencies, orchestration, and reuse boundaries.

Status: first detailed draft complete.

8. Runtime State, Context and Memory ¶

Understand what state a skill needs during execution, what belongs in transient context, and when persistent memory should remain outside the skill itself.

Status: first detailed draft complete.

9. Reliability, Failure Handling and Deterministic Boundaries ¶

Add retries, timeouts, idempotency, deterministic validation, stop conditions, fallback behavior, and explicit failure semantics around probabilistic execution.

Status: first detailed draft complete.

10. Security, Permissions and Trust Boundaries ¶

Apply least privilege, authorization, secret isolation, input trust classification, human approval, and safe tool access at the skill boundary.

Status: first detailed draft complete.

11. Skill Evaluation and Observability ¶

Evaluate skills with representative datasets and measure task success, tool use, failures, latency, cost, traces, and regressions.

Status: first detailed draft complete.

12. Versioning, Lifecycle and Design Patterns ¶

Treat skills as maintained software artifacts: version them, roll changes out safely, deprecate them, and recognize useful patterns and anti-patterns.

Status: first detailed draft complete.

Learning order¶

Completion state¶

The Agent Skills workstream is complete at first-draft depth. Future changes should primarily refine these chapters through questions, implementation experience, and evaluation rather than extend the roadmap linearly.

Next step¶

Build the detailed Agentic Loops roadmap, then expand it in learning-order blocks.