AI-DLC Discovery: Define What to Build Before AI Builds It
AI-DLC Discovery generates the Vision and Technical Environment documents AI-DLC needs before Inception — replacing the blank page with a guided interview.
Teams adopting AI-DLC consistently ask the same question: "How do I write inputs that produce production-grade output?"
AI-DLC Discovery answers that question. It's a guided, role-based interview that generates the Vision Document and Technical Environment Document AI-DLC expects — in 10–35 minutes, with a full audit trail.
"AI-DLC Discovery is a community contribution, not an official part of the AI-DLC project. Any opinions in this article are those of the individual author and may not reflect the opinions of AWS."
Why This Matters
The AI-Driven Development Lifecycle (AI-DLC) is a reimagined, AI-native methodology that positions AI as a central collaborator in software development. Its white paper defines three phases — Inception, Construction, and Operations — where AI initiates and directs workflows while humans provide oversight at critical decision junctures.
After accompanying multiple customers through their first AI-DLC engagements across LATAM, I noticed the same pattern in roughly 8 out of 10 teams:
The quality of AI-DLC's output is directly proportional to the quality of its input context.
This isn't a bug — its context engineering applied to software development at scale. Teams struggle not because they lack domain knowledge, but because writing a good input document from scratch requires you to simultaneously think about business intent, technical constraints, user personas, success criteria, and architectural trade-offs. That's a lot to hold in your head while staring at a blank page.
The Problem, Precisely
In an AI-driven methodology, prompt and context engineering isn't a developer convenience — it's a project management discipline.
| Input Quality | What AI-DLC Does | Result |
|---|---|---|
| No inputs | Generates generic questions from scratch | First hour of Inception wasted on basics |
| Vague inputs | Produces broad decomposition, asks many clarifying questions | Teams spend Mob Elaboration (The collaborative refinement session in AI-DLC Inception) narrowing scope |
| Rich, structured inputs | Immediately proposes contextually relevant User Stories, NFRs, Units | Mob Elaboration focuses on refinement and edge cases |
AI-DLC Discovery: The Front Door
AI-DLC Discovery is a role-based interview workflow that replaces the blank page with a guided conversation. It produces the exact artifacts AI-DLC expects — formatted, validated, and with known ambiguities pre-declared as open questions for AI-DLC to resolve during Inception.
- Business context — who's the user, what problem are we solving, what does success look like
- Technical context — what's the stack, what exists already, what are the integration points
How It Works
The workflow operates through a simple loop that mirrors AI-DLC's own mental model: AI proposes, human validates.
AI-DLC Discovery - 7 step workflow
The seven steps are:
- Role Selection — pick Business (Vision Document), Technical (Technical Environment Document), or both sequential.
- Project Type — brand-new project, new feature on an existing product, or migration/modernization. Shapes which follow-up questions appear.
- Depth — Quick pass (~10 min, core questions) for POCs, or Full interview (~25–35 min) for production-grade work.
- Interview — the tool writes questions into a markdown file with
[Answer]:tags. You fill them in, replyready, and the tool validates. - Gate — review consolidated answers, then approve or request changes before final document generation.
- Visual Sketch (optional) — five short questions, then the agent generates Mermaid user-journey flowcharts and HTML mockups. Adds ~10–15 min.
- Generate + Handoff — produces the final documents and a ready-to-paste prompt that gives AI-DLC full context from the first interaction.
Steps 1–5 are the core interview (~10–25 min). Step 6 adds visual artifacts if needed. The entire flow is resumable — stop after any step and pick up later.
What It Produces
After a full session, you have a
Product-Definition/ folder:1
2
3
4
5
6
7
8
Product-Definition/
├─ vision-document.md ← AI-DLC input #1: what, for whom, success metrics, MVP scope
├─ technical-environment.md ← AI-DLC input #2: stack, constraints, security, testing
├─ open-questions.md ← Ambiguities for AI-DLC to resolve in Inception
├─ aidlc-discovery-state.md ← Session state (enables resume)
├─ audit.md ← Append-only interaction log
├─ interview/ ← Questions + answers per role
└─ visual/ (optional) ← Mermaid flowcharts + HTML mockups
The three output files at the top are what you handoff to AI-DLC. Everything else is working state and audit trail.
The Handoff Prompt
When Discovery completes, it generates a ready-to-paste prompt for AI-DLC:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
I'm starting a new project. Please read the following inputs before
beginning the AI-DLC workflow:
REQUIRED inputs:
• Product-Definition/vision-document.md
The Vision Document — what we're building, who it's for, success metrics.
• Product-Definition/technical-environment.md
The Technical Environment — binding constraints for Construction.
• Product-Definition/open-questions.md
Pre-declared ambiguities to resolve in Requirements Analysis.
Now start the AI-DLC Inception phase.
This gives AI-DLC maximum context density from the first interaction.
Design Principles
- Natural language, no jargon. The user picks "brand-new project" not "greenfield."
- Question → Document → Approval. Fill
[Answer]:tags, replyready, tool validates. Same pattern AI-DLC uses internally. - Per-question state. Stop after any batch and resume next week. Business and Technical can work in separate sessions.
- Two depths per role. Quick (~10 min) for POCs. Full (~25–35 min) for production-grade work.
- Complete audit trail. Every interaction logged with ISO-8601 timestamps.
The Context Engineering Argument
Why does a 10–35 minutes interview make such a measurable difference in AI-DLC's output quality?
The difference comes down to context density. A well-structured Vision Document gives AI:
- Bounded problem space — no wasted tokens on excluded directions
- User personas with specificity — stories match actual users
- Success criteria — AI self-validates proposals
- Non-negotiable constraints — boundaries respected from the start
- Existing technical decisions — AI builds on your stack
AI-DLC's "Reverse the Conversation Direction" principle works best when AI's initial proposals are close to right. Close-to-right proposals come from rich context.
Pre-Loading Context: AI Helping You Help AI
The AI agent can read existing documentation before starting. The interview becomes a validation exercise instead of an authoring exercise:
1
2
3
4
5
6
7
Before starting, read the following as context:
- /workspace/docs/product-brief.docx
- /workspace/docs/architecture-overview.md
Then start aidlc-discovery. Propose answers from files; ask me only
for things the files don't cover.
Context engineering applied recursively: AI assembles the context that AI-DLC uses to drive the development process.
What Changes for Teams
Before Discovery
1
2
3
Business: "We need a recommendation engine."
→ AI-DLC Inception starts → First 45 min: clarifying questions
→ Construction starts with residual ambiguity
After Discovery
1
2
3
Business: "We need a recommendation engine."
→ Discovery interview (25 min) → Documents produced
→ AI-DLC starts with full context → Construction with clarity
Total time similar or shorter. Quality of starting context categorically different.
What AI-DLC Discovery Is Not
- Not a replacement for AI-DLC — it's its front door
- Not a validator of business logic — it asks the questions, you answer
- Not a code generator — no source code produced
- Not an automated extraction tool — your agent can read files, but you validate
Three Takeaways
1. Input quality determines output quality
The bottleneck is always "did we give AI enough context?" Discovery systematizes solving that bottleneck.
2. Prompt engineering is a team sport
A Product Owner answers, "who and why" while a Tech Lead answers, "what stack and what constraints." Both become the context that shapes every AI proposal.
3. The front door matters as much as the methodology
If the first experience is a blank page, teams stumble. If it's a guided conversation, teams start building.
Pro tip: If a question takes too long, skip it
If you're stuck on a question for more than 2–3 minutes, that's a signal — not a failure. Use
X) Other or write "TBD" in the [Answer]: field. Discovery captures these as pre-declared open questions in open-questions.md, which AI-DLC resolves during Requirements Analysis with the full team. A 70%-complete Discovery session with honest "I don't know" markers produces better AI-DLC output than a 100%-complete session with guesses.Getting Started
1. Install (one-time per project)
Paste the install prompt into your AI agent (Claude Code, Kiro, Cursor, Amazon Q Developer, or any agent with shell access):
1
2
3
4
5
Set up aidlc-discovery in this project:
1. Download latest release from GitHub API
2. Extract to .aidlc/aidlc-discovery/
3. Create rules/steering file for your IDE
4. Add .aidlc/ to .gitignore
Full prompt: install.md
2. Invoke
1
start aidlc-discovery
The agent reads the rules file, loads the core workflow, and begins the interview.
3. Complete the interview
Answer questions in the markdown files, reply
ready after each batch. Stop and resume anytime.4. Hand off to AI-DLC
Paste the generated handoff prompt into a fresh AI-DLC context. Inception starts with full context.
5. Learn more
- AI-DLC blog post — methodology overview
- AI-DLC white paper — full method definition
- AI-DLC workflows — IDE rules implementation
- AI-DLC Discovery — source, examples, changelog
Francisco Fagas is a Senior Specialist Solutions Architect for Generative AI at AWS, based in San Juan, Puerto Rico, covering LATAM NOLA enterprise accounts. He created AI-DLC Discovery as a personal contribution based on his experience accompanying multiple customers through AI-DLC engagements.
Reviewers
- César Bautista — Telco Solutions Architect at AWS, based in Mexico City. Passionate about generative AI and modern developer tooling, César leverages his deep technical expertise to drive cloud adoption and digital transformation across Latin America through AI-DLC engagements.
- Gonzalo Vasquez — Solutions Architect at AWS, based in Santiago, Chile, covering ISV accounts across SOLA. He focuses on developer experience, AI-assisted development workflows, and helping customers adopt modern software delivery practices.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article