View on GitHub
On this page
Related Technologies Related Topics
AI MCP Dev Tools Prompt Engineering

Artificial Intelligence

GenAI the model; core capability
Context window a hard constraint on how much the GenAI can consider at once
Tooling (tool use) / Tool Calling a pattern to bypass or mitigate context limits by letting the model query external systems instead of encoding everything in the prompt
AI agents orchestrate GenAI + tooling + memory + planner to accomplish complex, multi-step tasks
MCP: Model Control Plane governs, routes, secures, and observes GenAI + tools + agent orchestration at org scale
MCP: Model Context Protocol a structured envelope for what you send to / receive from a model -- it's about input/output data shape, provenance and runtime constraints

GenAI: Generative AI

  • large models that generate content (text, code, images, audio)
  • Role: the core “reasoning + generation” capability
  • Limits:
    • hallucination/incorrect facts
    • sensitivity to prompt phrasing
    • bounded by training data,
    • constrained by runtime context (context window) Context Window Problem
  • the limitation on how many tokens the model can attend to in a single prompt/response
  • Symptoms:
    • truncated history
    • degraded reasoning on long documents
    • loss of earlier context Tool Calling
  • the pattern of letting the model call external tools (search, DBs, code runners, calculators, APIs) rather than putting everything into the context window
  • effectively “extends” the model’s capability beyond the token limits by offloading data/state/compute or access to specialized systems AI agents
  • orchestrations that combine a GenAI model with planning, tool selection, memory, and action execution
    • e.g., planner -> choose tool -> execute -> observe -> update plan
  • Use cases: multi-step tasls (research, booking, complex automation) MCP: Model Control Plane
  • Control-plane infrastructure that manages models, routing, policy, observability, audit logs, access control, cost/budgeting, versioning, and possibly deployment of tool adapters
  • analogue: like the operations center that keeps your AI models running in production (deployment, scaling, monitoring) MCP: Model Context Protocol
  • it’s about input/output data shape, provenance and runtime constraints
  • analogue: like a universal adapter that lets AI models plug into any data source or tool they need to answer questions better
  • companies who want AI agents too work with their products use this method

AI Models & Tools

Claude

  • Claude Code
    • a terminal/IDE assistant that executes agentic actions (editing files, running tests, making commits)
      • while the Claude model (Opus/Sonnet/Haiku - depending on configuration) supplies reasoning and generation
  • Claude Opus
    • flagship
    • highest reasoning and code capability
    • best for deep/complex tasks
    • slowest and most expensive
  • Claude Sonnet
    • balanced “workhorse”: good accuracy, speed, and cost for general tasks including coding and analysis
  • Claude Haiku
    • fastest and cheapest
    • for high-volume low-complexity tasks and real-time use
    • lower depth

Claude Code Magic Tricks

for better development of software solutions upon user requests

compacts

  • when it runs out of context window, it “compacts” prior conversation history and takes notes about exactly where it was when it stopped
  • then it clears its context window and fresh version of the conversation starts with reviewing the notes taken skills
  • picks up skills as needed from a prior configured curated list of skills in natural language
    • skills -> instructions -- prompts, sets of tools needed for specific tasks
  • free list of skills
  • Claude Code slash command /skills lets you create or download skills subagents
  • spawn specialized subagents for specific tasks (e.g., code analysis, testing, documentation)
  • allows to run many different processes at once in parallel
  • Claude Code slash command /agents lets you set up subagents