Core Concepts
Before diving deeper, let’s establish shared vocabulary.
Key Terms
Section titled “Key Terms”| Term | What It Means |
|---|---|
| Model | The AI itself (e.g., GPT-5.2, Claude Sonnet 4.5). Determines capability and quality |
| Provider | Company hosting the model’s API (e.g., OpenAI, Anthropic). Determines pricing and terms |
| Agentic | AI that acts autonomously: reads files, runs commands, iterates on errors |
| Context Window | How much text the AI can “see” at once (measured in tokens). More ≠ better |
| BYOK | Bring Your Own Key. Use your own API keys instead of a subscription |
| MCP | Model Context Protocol. A standard for connecting AI tools to external services |
| Skill | A reusable instruction set, playbook, or convention that helps an agent perform a class of tasks more reliably |
Context Concepts
Section titled “Context Concepts”| Term | What It Means |
|---|---|
| Context Engineering | Building systems to provide the right information and tools to the model |
| Context Rot | When too much irrelevant context makes the AI “dumber” |
| Context Budget | A practical limit for how much context to load before quality starts to degrade |
| Subagent | A separate AI instance launched for investigation, keeping main context clean |
Workflow Concepts
Section titled “Workflow Concepts”| Term | What It Means |
|---|---|
| Composer/Agent Mode | Multi-file editing mode (vs. single-file autocomplete) |
| Prompt | Your instruction to the AI. Quality of prompt = quality of output |
| Verification | Having the AI check its own work (tests, linter, type checker) |
| Close the Loop | Design workflow so the agent can verify its own output |
Context Budget
Section titled “Context Budget”There is no strong primary-source basis for a universal 40% threshold. Treat hard percentages as heuristics, not laws.
The safer idea is simpler: quality drops before the window is full, especially when context is noisy. Use selective retrieval, compaction, and project context files instead of stuffing everything into one prompt.
If you want a working rule, think in ranges rather than one magic number. The practical target depends on the model, the task, and how clean the context is.
Models vs Providers
Section titled “Models vs Providers”Models determine capability — how smart the AI is, how fast it responds.
Providers determine terms — pricing, privacy policy, legal jurisdiction.
The same model can be available through multiple providers:
- Claude Sonnet 4.5 is available via Anthropic, AWS Bedrock, and Google Vertex
- GPT-5.2 is available via OpenAI and Azure OpenAI
This matters for enterprise (data residency, compliance) and cost optimization.
Next Steps
Section titled “Next Steps”- Understand the Five Levels — where you are and where to aim
- Learn Context Engineering — the key skill
- MCP Deep Dive — the connectivity layer
- Skills Deep Dive — reusable instructions and playbooks