Foundational Disciplines¶
The four practitioner disciplines that determine agent output quality — independent of which tool you use.
These modules teach the conceptual frameworks behind effective AI-assisted development. Each discipline addresses a different layer of the system: what you say (prompts), what the agent sees (context), what it can do (tools), and what catches mistakes (harness). The capstone shows how the four compound.
Core Modules¶
| Module | Topic | Duration |
|---|---|---|
| Prompt Engineering | System prompt altitude, polarity, rule- vs example-driven instructions, compliance ceiling, domain-specific prompts, negative-space instructions | 30–45 min |
| Context Engineering | Context window mechanics, attention sinks, lost-in-the-middle, compression strategies, dynamic context assembly, JIT loading, prompt caching | 30–45 min |
| Harness Engineering | Repo legibility, mechanical enforcement, constrained solution spaces, backpressure, feedback loop quality, pre-completion checklists, convergence detection | 30–45 min |
| Tool Engineering | Tool description quality, token-efficient design, schema design, MCP architecture, skill authoring, tool minimalism, poka-yoke tools | 30–45 min |
| How the Four Disciplines Compound | The multiplication model, diagnosing failures by discipline, investment progression, decision frameworks | 30–45 min |
Complementary Modules¶
| Module | Topic | Duration |
|---|---|---|
| Eval Engineering | Pass@k metrics, LLM-as-judge, golden query pairs, incident-to-eval synthesis, behavioral testing, anti-reward hacking | 30–45 min |
| Autonomous Research Loops | Experimentation vs information research loops, termination design, doom loop prevention, context management, grounding strategies, control surfaces | 30–45 min |
| Earned-Complexity Agent Maturity Ladder | Nine diagnostic layers from single-shot tool calling to multi-agent delegation — each rung exposes the failure modes the next pretends to solve, with anchor pages for every layer | 30–45 min |