Glossary
Every AI-DLC 2.0 term used in the lessons, in plain English.
- Abstraction boundary
- The orchestrator treats each stage as a black box and only checks its declared post-conditions. Lesson 4
- Additive extension
- Layering your own rules and post-conditions on top of the AWS baseline. Lesson 2
- AI-DLC
- The AI-Driven Development Life Cycle: a way of structuring how humans and AI agents share software delivery, from requirements through deployment. Lesson 1
- Audit trail
- The orchestrator's record of every decision: which stages ran, why and in what order. Lesson 4
- Bootstrap stage
- An initial setup stage. All non-bootstrap stages require explicit human approval (Guideline 4). Lesson 6
- Canonical stage definition
- The single source-of-truth file for a stage, referenced by filename from every tool-specific version. Lesson 5
- Compartment 1 (the What)
- Declares a stage's inputs, outputs and required intermediate artefacts. Declarative, not step-by-step instructions. Lesson 2
- Compartment 2 (How do we know it's right)
- The post-conditions: validation rules a correct output must satisfy. Also called the verification spec. Lesson 2
- Compartment 3 (What did we learn)
- Defines which runtime signals become candidate rules and how they are promoted. Also called the learning spec. Lesson 2
- Composability
- Adding entirely new stages, written with the same three-compartment construct, that AI stitches into the workflow. Lesson 2
- Compound engineering
- Skills improving from practice: human corrections become proposed rules that a human approves (Principle 9). Lesson 2
- Computational verification
- A check written as an executable (script, linter, test, static analysis) for rules with zero tolerance for ambiguity. Lesson 3
- Conductor
- The coordinator of agent work. Agents do not spawn other agents; the conductor does the coordinating. Lesson 6
- Cross-stage invariant
- A rule that spans the whole workflow, such as naming conventions or security policies, checked by the orchestrator. Lesson 4
- DRY
- Don't Repeat Yourself: keep one definition and reference it instead of copying it. Lesson 3
- Emitter
- The part of the package manager that generates one tool's artefacts and layers that tool's native features on top. Lesson 5
- Escalation
- Pausing and handing a problem to a human when AI is stuck or needs judgement, then resuming after it is resolved. Lesson 4
- Generate-Verify-Learn
- The spec's name for the three compartments as Definition, Verification and Learning specs. Lesson 3
- Guardrail / no-go zone
- A rule that must never be broken, such as never deleting a production CloudFormation stack. Lesson 2
- Halting condition
- The limit on a self-correcting loop: a maximum number of iterations or a token budget. When reached, AI escalates to a human. Lesson 2
- Human-in-the-loop
- A person reviews, clarifies or approves at a given point. AI-DLC 2.0 aims to need this only where it is genuinely required. Lesson 1
- Hydrate / hydration
- Gradually filling in Compartment 2 with more rules (guardrails, compliance, standards) so AI can verify more on its own. Lesson 2
- Inferential verification
- A check written as LLM instructions, for judgement calls that are not strictly pass/fail. Lesson 3
- Intent
- The business request that starts a workflow, such as “build a sales dashboard”. Always somewhat ambiguous at first (Principle 2). Lesson 2
- Knowledge library
- Reference material agents use to generate outputs and to self-correct. Lesson 3
- Lifecycle hooks
- Deterministic platform hooks used for halting and audit. Lesson 1
- Lowest common denominator
- Limiting a shared design to only what every tool supports. The spec explicitly avoids this. Lesson 5
- MCP
- Model Context Protocol: a portable way to deliver tools and context to agents. Lesson 1
- Multi-agent runtime
- A platform construct for modelling collaborating specialist agents. Lesson 1
- Orchestrator
- Composes stages into adaptive workflows, owns the goal, routes inputs, tracks state, enforces halting and keeps the audit trail. Lesson 4
- Orchestrator hints
- Metadata in a stage definition: sequencing constraints, conditional triggers, escalation policies. Lesson 4
- Package manager
- Packages tool-agnostic stage and orchestrator definitions into tool-specific artefacts for Kiro, Claude Code, Cursor or GitHub Copilot. Lesson 5
- Persona-Agent
- A specialist AI agent modelled on a team role, such as an architect or tester. Several collaborate in a multi-agent workflow. Lesson 3
- Post-condition
- A check that a stage's output must pass. Clear post-conditions are what make self-correction possible. Lesson 2
- Replacement extension
- Replacing specific baseline rules and taking ownership of that stage definition. Lesson 2
- SDLC
- Software Development Life Cycle: the familiar requirements, design, build, test and deploy stages. AI-DLC keeps them. Lesson 1
- Self-correcting loop
- AI plans, executes, verifies its output against post-conditions, and retries until it passes or hits a halting condition. Lesson 3
- Self-halt
- A stage finishing on its own once its checks pass. Only possible when it has computational checks, not LLM-only ones. Lesson 3
- Single-shot compression
- Trying to go straight from one prompt to production code. Rejected in favour of staged decomposition (Principle 6). Lesson 2
- Skill
- A packaged, discrete capability or expertise (database design, code review, security analysis). The unit of composition in AI-DLC 2.0. See agentskills.io/specification. Lesson 2
- Soundness and completeness
- Whether the post-conditions are correct (sound) and cover everything that matters (complete). Both grow over time. Lesson 2
- Stage
- A defined piece of work and the artefacts that must exist when it is complete. Stages and phases are flexible organising concepts above Skills. Lesson 3
No matching terms.