AI engineering delivery practice map

This guide focuses on how AI enters real software engineering delivery, rather than staying at the level of scattered prompt tricks. It is a delivery map: from requirement boundaries and Manifest contracts to implementation gates, audit evidence, and reusable collaboration data.

How to Use This Map

Start from the unstable part of your process:

• If requirements are vague and the agent keeps improvising, start with AI-TDD Manifest.
• If a public knowledge assistant is unstable or untraceable, start with the RAG system guide.
• If task slicing is unstable and implementation keeps drifting, use BMAD / Speckit workflow.
• If code appears complete but delivery evidence is weak, focus on delivery gates and delivery confirmation evidence.
• If the team wants to turn collaboration traces into reusable training or feedback assets, read the Coding Mentor data loop.

---

Core Module 1: RAG Public Knowledge Entry

RAG answers an engineering question: can a public assistant answer only from traceable, indexed, auditable site knowledge?

This guide covers three layers. First, it defines the edge architecture, three storage layers, and incremental indexing loop. Second, it goes into chunking, hybrid retrieval, intent routing, and current-page summaries. Third, it defines quality evaluation, safety gates, and release evidence for opening a public /chat endpoint.

Recommended reading:

• Building a RAG System: Architecture, Retrieval, Quality, and Release Gates

Use this path when you already have public content, documentation, or a knowledge hub and want the “Ask” entry point to be governed by release evidence instead of hopeful prompting.

---

Core Module 2: AI-TDD Manifest-Level Acceptance

AI-TDD answers a delivery question: when an AI agent can generate implementation quickly, how do we make sure the generated work stays aligned with the intended boundaries?

The method turns natural-language requirements into a machine-readable Manifest before implementation starts. The Manifest records core namespaces such as MUST, NEG, OUT, TRACE, EVD, ACC/E2E, FAIL/EDGE, CMD, ART, and TASK. The agent then implements inside this contract, while readiness and delivery gates verify the state transition.

Recommended reading:

• AI-TDD: Acceptance-Driven Development with a Manifest Contract

This is the deep-dive article for teams facing requirement drift, test coverage that does not equal requirement coverage, implementation scope creep, or weak delivery evidence.

---

Core Module 3: BMAD / Speckit Workflow

BMAD and Speckit focus on organizing delivery work: how to break requirements into executable task packets, how to confirm boundaries before implementation, and how to preserve audit evidence after completion.

AI-TDD can be treated as the acceptance-contract layer inside that workflow. BMAD / Speckit organize the work; AI-TDD turns the acceptance boundary into a checkable machine contract.

Use this layer when:

• A multi-step requirement needs executable task packets.
• The team needs explicit scope, constraints, and acceptance conditions before AI implementation.
• Planning, implementation, tests, review, and delivery evidence need to be connected.

---

Core Module 4: Delivery Gates and Evidence

AI engineering delivery cannot stop at “the code was generated” or “some tests passed.” A deliverable result needs to answer:

• Which requirements were verified?
• Which negative constraints were not violated?
• Which command, report, or artifact produced the evidence?
• Did the current result reuse stale evidence?
• Did a human reviewer make the final delivery decision?

AI-TDD Gate, TRACE rows, the EVD/ART/CMD evidence chain, and delivery confirmation gates form the delivery-gate layer. The point is not to add ceremony. The point is to prevent AI-generated work from becoming untraceable at the last mile.

---

Extended Module: AI Coding Mentor

AI-TDD focuses on preventing drift in a delivery run. AI Coding Mentor focuses on how teams continuously evaluate, guide, and improve AI coding systems.

Recommended follow-up reading:

• AI Coding Mentor Series
• From delivery to training: turning AI coding collaboration into a Coding Mentor data loop
• From engineering practice to training data: systematic SFT data generation for AI engineering

---

Recommended Adoption Path

1. Pick one small feature and write a minimal AI-TDD Manifest.
2. Add at least one acceptance test or verification command for each MUST.
3. Add NEG negative assertions and OUT scope boundaries for the main risk areas.
4. Run readiness checks before implementation and confirm the new acceptance tests are in expected RED state.
5. Let the agent implement inside the Manifest boundary, then run the delivery confirmation gate.
6. Preserve failures, human corrections, and delivery evidence as reusable team assets.

Related reading

The quantitative trading AI engineering chapter shows how specification-driven delivery is used in a real project. The AI Coding Mentor series explains how task design and feedback loops become evaluation assets. The platform-upgrade path series shows how these practices fit into a broader content and engineering system.