AI-DLC 2.0Workflows, explained

Home / Lesson 2 of 6

The Nine Principles

The big ideaNine beliefs explain every design choice in AI-DLC 2.0. The core one: give AI clear inputs, outputs and checks, and it can correct its own work, with a human safety net.

Video, 5:11. On-screen captions are built in.

The principles are the “why” behind the whole framework. You do not need to memorise them. Read the one-line version of each, then open any card for an example. Three of them carry most of the weight: 3 (AI corrects itself against checks), 4 (every task has three compartments) and 7 (autonomy arrives in safe increments).

Explore the nine principles

Open any principle for an example.

1Human judgement can be distilled into machinesMuch of the effort humans spend judging, validating and steering AI can be captured as rules a machine runs.

Every time a reviewer leaves the same comment (“add input validation here”), that is a judgement pattern waiting to become an automated check. The spec treats today’s heavy human involvement as a starting point to optimise, not a permanent state.

Read Principle 1 in the spec
2Every intent starts ambiguous, and that's OKRequests are always fuzzy at first, so AI must ask clarifying questions instead of assuming.

“Build me a dashboard for sales performance” hides many questions: which metrics, what time range, for whom, why? AI-DLC uses the human-in-the-loop to clarify. Curating a clear intent statement up front makes this step much faster.

Read Principle 2 in the spec
3AI as a self-correcting solver, with a safety netGiven clear inputs, outputs and checks, AI can try, check and fix its own work until it passes, within a budget.

It works when the checks run as a program the AI cannot modify, success is tractable, and retries are cheap, for example Build & Test. Every loop has a halting condition (max iterations or token budget). If AI cannot converge, it escalates to a human instead of looping forever.

Read Principle 3 in the spec
4The three-compartment modelEvery task is described by What (inputs/outputs), How do we know it's right (post-conditions), and What did we learn.

Without clear post-conditions AI cannot self-correct; without clear inputs and outputs it does not know the problem. At first the checks are incomplete, so AI asks humans more often (for example about dashboard colours). Add your UI standards as a check, and it stops needing to ask.

Read Principle 4 in the spec
5Every SDLC stage fits the constructThe same compartment model works for any stage, not just coding and testing.

Requirements clarification, architecture design, deployment and monitoring can all be framed as: defined inputs, expected outputs, explicit post-conditions. Creative, analytical or mechanical, it is the same shape.

Read Principle 5 in the spec
6Staged decomposition over single-shot compressionBreak delivery into stages; do not try to go from one prompt to production code.

One giant step causes post-condition explosion (every rule in one place) and combinatorial search explosion (endless loops). Version 1’s fixed stages were too opinionated, so version 2 introduces Skills: small, composable units of expertise such as database design, code review or security analysis.

Read Principle 6 in the spec
7Autonomy in safe increments, not a big bangOrganisations turn their rules into machine checks gradually, and autonomy grows with them.

Start with what you have (coding standards, naming conventions, basic security rules). Hydrate incrementally by adding more guardrails and compliance rules as checks. Expand the safe increment: as the rule library matures, AI can run through more stages on its own.

Read Principle 7 in the spec
8ExtensibilityYou can extend, override or add to the AWS baseline with the same tools and guarantees.

Additive: layer your own rules on top of the baseline. Replacement: swap a baseline rule where your needs differ, and own that stage from then on. Composability: author entirely new stages with the same three-compartment construct and let AI stitch them in.

Read Principle 8 in the spec
9Learning from practice (compound engineering)Skills improve from real use: human corrections become proposed new rules, approved by a human.

Today a correction is applied once and forgotten. Under this principle, if a user keeps rewriting the same section, the stage proposes a new post-condition; if they keep overriding a design choice, it proposes a guardrail. A human approves before anything becomes active.

Read Principle 9 in the spec

The Three-Compartment Model, at a glance

Principle 4 is the shape every Skill and stage takes. If you remember one diagram, make it this one.

Compartment 1

The “What”

Inputs and outputs, plus any intermediate artefacts that must be produced. Example: user stories in, domain model out.

Declarative: it says what must exist, not how to do it.

Compartment 2

“How do we know it’s right?”

Post-conditions: the rules and quality checks a correct output must pass. Example: every entity in the model traces back to a user story.

Without these, AI has no way to self-correct.

Compartment 3

“What did we learn?”

Which signals from real use (human corrections, re-runs) become candidate new rules, and how they get promoted.

This is how the checks get better over time.

The spec names these the Definition Spec, Verification Spec and Learning Spec, together called Generate-Verify-Learn. Principle 5’s heading says “Two-Compartment”, but its text and the rest of the spec describe the three-compartment construct.

Try this at work

Pick one recurring review comment your team makes (for example, "every API needs auth"). Writing it down as a check is exactly what the spec calls hydrating Compartment 2.

Want the original wording? Read Section 2 of the specification.

Quick check

Three questions. Get them all right to mark this lesson done.

Q1. What must every self-correcting loop include?

Q2. Why does the spec reject going straight from “build me this app” to deployed code in one shot?

Q3. A user keeps rewriting the same section of AI-generated requirements. What does Principle 9 say should happen?