Discovery Tool for AI-DLC: One Core, Every Platform

Discovery Tool for AI-DLC ships as a platform-agnostic skill system — run the same guided interview in Kiro, Claude Code, or Amazon Quick with zero code changes.

Francisco Fagas

Senior Specialist SA, GenAI/ML

Published Jun 3, 2026

Last modified Jun 9, 2026

AI-DLC Discovery replaces the blank page with a guided, role-based interview that produces the structured inputs AI-DLC needs before Inception. In 10–35 minutes, a PM and a Tech Lead walk through focused questions and emerge with a Product-Definition/ folder — vision, constraints, open questions — ready to hand off.

The v1 blog post covered the why and the 7-step flow. That core method hasn't changed. What changes in the last release is how it's packaged and where it runs.

"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 new release

Three forces drove the rewrite:

AI-DLC v2 (Under development move to skills, check official updates from AI-DLC Github repo) — the monolith of rules became composable skills following the agentskills.io spec. Discovery needed to speak the same language.
Customers asked "does it work in my IDE?" — AI-DLC v1 shipped as a generic rules bundle. It worked, but adapting it to Kiro vs. Claude Code vs. Amazon Quick meant manual surgery every time.
Roles should run in parallel — in AI-DLC v1, Business and Technical run sequentially. Real teams have the PM and Tech Lead working concurrently on different machines.

What Changed (and What Didn't) Discovery Tool for AI-DLC

Unchanged — the method

Preserved from v1	Still works exactly the same
7-step interview flow	Role → Type → Depth → Interview → Gate → Visual → Handoff
Question → Doc → Approval pattern	`[Answer]:` tags, `ready` to validate
Quick vs. Full depth	~10 min vs. ~25–35 min
Resumable state	Stop after any step, continue later
Output: `Product-Definition/` folder	vision + constraints + open-questions
Visual Sketch (optional)	Mermaid user journeys + HTML mockups

Changed — the architecture

Discovery Tool for AI-DLC v1	Discovery Tool for AI-DLC v2
Monolith rules file	Modular skills (`agentskills.io` spec)
Generic bundle, adapt manually	Core + Adapters + `build.js` → `dist/<target>/`
Roles run sequentially (Both-Sequential)	Roles run in parallel (Both-Parallel)
Language detection module	Eliminated — agent follows user's language naturally
Best-effort validation	Optional deterministic verification (`process_checker`)
1 platform (generic)	4 Tier-1 platforms: Kiro, Claude Code/Cowork, Amazon Quick

Architecture: Core + Adapters + Build

The key insight: the interview is 100% portable because it's text instructions. What changes per platform is how the agent registers, installs hooks, and transports questions.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
aidlc-discovery/
├── src/
│   ├── skills/                        # ── CORE (agentskills.io, 100% portable) ──
│   │   ├── discovery-orchestrator/    #    mini-orchestrator
│   │   ├── product-discovery/         #    PM interview → vision
│   │   ├── tech-discovery/            #    Tech Lead interview → constraints
│   │   ├── open-questions/            #    Cross-role reconciliation
│   │   └── visual-sketch/            #    Optional visual artifacts
│   ├── common/                        # ── Shared protocols ──
│   │   ├── protocols/                 #    builder / validator / orchestrator
│   │   ├── conventions/               #    question-format, state-schema
│   │   └── scripts/process-checker.js #    Deterministic gate (optional)
│   └── targets/                       # ── ADAPTERS (only part that changes) ──
│       ├── kiro/                      #    .kiro/skills + agents + hooks
│       ├── claude/                    #    Personal Skills + MCP + hooks
│       └── amazon-quick/              #    Quick skill (multi-step + tools)
└── build.js                           # → dist/<target>/ ready to deploy

Write once, build per target. The build.js script assembles src/skills + src/common + the target-specific adapter into a deployable package. No copy-paste, no drift.

Parallel Roles — No Races

Previous release "Both" mode forced PM → then → Tech Lead in sequence. The new release runs them concurrently on a shared Product-Definition/ folder with strict partitioning:

1
2
3
4
5
6
7
Product-Definition/
├── state/
│   ├── session-index.md        # shared: project-type, depth, mode, role status
│   ├── business-state.md       # written ONLY by PM session
│   └── technical-state.md      # written ONLY by Tech Lead session
├── interview/{business,technical}/
└── audit/{business,technical}-audit.md

Single-writer per role eliminates race conditions. The last session to complete triggers a JOIN phase that reconciles contradictions between vision and constraints, surfacing them as open questions — something v1 never did.

Four Platforms, One Tier

Every target is Tier 1: full skill + agent + automation + deterministic verification. The only variance is question transport:

Platform	Adapter	Question Transport
Kiro IDE/CLI	`kiro`	File with `[Answer]:` tags
Claude Code	`claude`	File with `[Answer]:` tags
Claude Cowork	`claude` (same)	File with `[Answer]:` tags
Amazon Quick	`amazon-quick`	Conversational in batches of 3–5

It Works Today

Here's Discovery Tool for AI-DLC v2 running the same "Ideation Portal" project across all three platform families (see Getting Started below to try it yourself):

Claude Code (Bedrock)

Claude Code loads the aidlc-discovery skill, reads the protocols, and begins the orchestrated interview — all from start aidlc-discovery.

Amazon Quick

Amazon Quick activates the discovery-orchestrator skill, pre-fills shared selections based on user preferences, and produces the Vision Document in the Session Tab — conversational transport, same output.

Kiro

Kiro activates the discovery-orchestrator skill, searches the workspace for protocols, reads the state schema and question format, then drives the interview through its spec-session workflow.

Same project. Same questions. Same Product-Definition/ output. Three different AI platforms.

Discovery Is Independent — AI-DLC Is a Consumer

An important positioning clarification: AI-DLC Discovery is not part of AI-DLC. It's a standalone product-ideation tool. AI-DLC is one of its consumers, not its container.

This means:

Discovery has its own orchestrator, its own front door
The output (Product-Definition/) can feed AI-DLC, or a team that just wants a clean product definition

Getting Started with this release

Install

Paste this automated install prompt into your AI agent. It detects your platform, downloads the correct pre-built package, and installs it — no manual steps. Full reference: install.md

1
2
3
4
5
6
7
8
9
10
11
Set up aidlc-discovery v2 in this project. Do every step yourself and report back when done:

1. Detect which platform you are running in: Kiro, Claude Code, or Amazon Quick desktop.
2. Download the matching asset from the v2.0.0 release (replace <platform> with kiro | claude | quick):
   https://github.com/aws-samples/sample-aidlc-discovery/releases/download/v2.0.0/aidlc-discovery-<platform>-v2.0.0.zip
3. Unzip and install to the platform's location:
   - kiro   -> merge the bundle into ./.kiro/
   - claude -> merge the bundle into ./.claude/
   - quick  -> copy the aidlc-discovery/ skill folder into ~/.quickwork/profiles/federate-prod/skills/
4. Confirm the install path and the invocation phrase
   (Kiro / Quick: "start aidlc-discovery"; Claude Code: "/aidlc-discovery").

Then invoke:

1
start aidlc-discovery

For a full end-to-end walkthrough with examples, see how-to-use.md .

What's New in the Invocation

Same trigger. The orchestrator now asks one additional shared question: mode (single / sequential / parallel). After that, each role's interview proceeds independently or in sequence based on your choice.

Migration from Discovery v1

If you're already running Discovery v1 (the rules-based install), here's what changes:

What you keep

Any existing Product-Definition/ folder with completed interviews — v2 reads the same output format
The start aidlc-discovery invocation trigger — same phrase, same behavior
All answers you've already captured — state files are forward-compatible

What you replace

v1 artifact	v2 replacement	Action
`.aidlc/aidlc-discovery/` (rules bundle)	`dist/<target>/` (skills package)	Delete `.aidlc/`, install v2
Rules/steering file (CLAUDE.md, .kiro/steering/, etc.)	Platform-native skill registration	Rebuild with `node build.js --target <your-platform>`
`aidlc-discovery-state.md` (single state file)	`state/session-index.md` + per-role state files	Auto-migrated on first v2 run
Sequential "Both" mode	Parallel mode (default)	No action — new sessions default to parallel

Migration steps

1
2
3
4
5
6
7
8
9
10
11
12
13
# 1. Remove v1
rm -rf .aidlc/aidlc-discovery/

# 2. Clone and build v2
git clone -b v2 https://github.com/aws-samples/sample-aidlc-discovery.git /tmp/discovery-v2
cd /tmp/discovery-v2
node build.js --target kiro   # or: claude, quick

# 3. Copy to your project
cp -r dist/kiro/ /path/to/your-project/

# 4. Invoke — v2 detects existing Product-Definition/ and offers to continue or start fresh
start aidlc-discovery

If v2 detects a v1-era Product-Definition/, it offers three choices: continue the existing session (migrating state), start a fresh parallel session alongside it, or archive and begin new.

Sample Output: What Discovery Produces

Here's a trimmed example of the actual vision-document.md generated by v2 for the "Ideation Portal" project:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
# Vision Document — Ideation Portal

## Problem Statement
Teams using AI-DLC Discovery need a centralized workspace to manage
multiple discovery sessions, track progress across roles, and review
generated artifacts before handoff.

## Target Users
- Product Managers running discovery for 2–5 concurrent projects
- Tech Leads reviewing technical constraints across team initiatives
- Engineering Managers tracking ideation-to-inception pipeline

## Success Criteria
1. PM completes a full discovery session in under 30 minutes
2. Tech Lead can work in parallel without blocking PM progress
3. Handoff prompt generates in <5 seconds after both roles approve

## MVP Scope (In)
- Multi-project dashboard with session status
- Parallel role execution with live progress indicators
- One-click AI-DLC handoff with pre-validated prompt

## MVP Scope (Out)
- Team permissions / RBAC (post-MVP)
- Integration with external project management tools
- Custom question banks (uses default bank only)

And the corresponding open-questions.md — ambiguities surfaced during the interview:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# Open Questions — Ideation Portal

## Cross-Role Contradictions
1. PM envisions a web-based dashboard; Tech Lead constraint says
   "desktop-only, no server component." → Resolve: clarify if Electron/Tauri
   qualifies or if "web" means browser-accessible.

## Unresolved During Interview
2. Authentication model: SSO via corporate IdP or standalone credentials?
   (PM answered "TBD", Tech Lead deferred to security team)
3. Data persistence: local filesystem vs. cloud-synced workspace?
   (Both roles answered differently — PM: cloud; Tech: local-first)

## Recommended Resolution
- Items 1–3 should be resolved in AI-DLC Requirements Analysis
  with the full cross-functional team present.

These files drop directly into AI-DLC's Inception phase. The handoff prompt references them by path so AI starts reading with full context from token zero.

Three Takeaways

1. One core, every platform

Write the interview once as portable skills. Adapters handle the platform-specific wiring. No more "fork the rules for each IDE."

2. Parallel roles = faster, richer discovery

PM and Tech Lead work concurrently. The JOIN phase catches contradictions between vision and constraints that sequential mode missed.

3. Discovery is a product, not a phase

It's independent. AI-DLC consumes its output. So can any team that needs structured product ideation — regardless of what methodology they follow after.

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.

Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.

# agentic-ai # agents # ai # ai-assisted # ai-dlc

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

Builder Center

Discovery Tool for AI-DLC: One Core, Every Platform

Francisco Fagas

Why new release

What Changed (and What Didn't) Discovery Tool for AI-DLC

Unchanged — the method

Changed — the architecture

Architecture: Core + Adapters + Build

Parallel Roles — No Races

Four Platforms, One Tier

It Works Today

Claude Code (Bedrock)

Amazon Quick

Kiro

Discovery Is Independent — AI-DLC Is a Consumer

Getting Started with this release

Install

What's New in the Invocation

Migration from Discovery v1

What you keep

What you replace

Migration steps

Sample Output: What Discovery Produces

Three Takeaways

1. One core, every platform

2. Parallel roles = faster, richer discovery

3. Discovery is a product, not a phase