Case Study, Part 2 of 5: Ideation Phase

Where we are: Building synth-data, a synthetic-data CLI. Setup is done, we kicked off with feature scope, and Initialization just finished. Now the workflow enters Ideation — the phase that turns a brief into a confirmed, scoped intent before any design or code happens.

In plain English: Before any code gets written, the AI first makes sure it actually understands the idea — who it's for, what "done" looks like, whether it's even worth building, and what's deliberately left out. Think of this phase as the conversation you'd have with a teammate before opening your code editor, just written down and agreed on instead of assumed.

Every prompt we actually type is highlighted like this.

Stage 1.1: Intent Capture (product agent)

The status line at the bottom of the terminal updates:

[AIDLC] IDEATION > Intent Capture [▓░░░░░░░░░] 1/7 -- product

The product agent asks how we want to work:

▸ Choose interaction mode:
  (1) Guide Me — agent asks structured questions
  (2) Edit File — write directly to the artifact
  (3) Chat — freeform discussion

You: 1

It asks its questions one at a time; we answer inline:

Claude Code: Who is this for — you personally, or a wider audience? You: Just me and my team at first, but I'd like to publish it publicly if it turns out useful.

Claude Code: What counts as "done" for the MVP? You: All four output formats work, and a schema with 5-10 columns generates 10,000 rows in under a few seconds.

Claude Code writes intent-statement.md and intent-capture-questions.md itself, then opens the gate:

# Intent Capture & Framing Complete

| Artifact | Contents |
|---|---|
| intent-statement.md | Problem statement, target users, success criteria |
| intent-capture-questions.md | 2 questions, all answered |

▸ How would you like to proceed?
  (1) Approve — Continue to Market Research
  (2) Request Changes

You: Approve

Behind the scenes: Each answer fires QUESTION_ANSWERED; writing the two files fires ARTIFACT_CREATED via aidlc-write-audit-log. The same write triggers aidlc-run-sensors, which checks the new files against whatever sensors this stage declares — a SENSOR_PASSED (or SENSOR_FAILED, which would block the gate) follows. Approving fires GATE_APPROVED, and aidlc-state.md's checkbox for this stage flips [?] → [x].

Stage 1.2: Market Research (product agent)

Claude Code looks at what already exists in this space — Faker as a library, Mockaroo as a hosted tool, SDV for statistically-realistic synthetic data — and writes market-research.md:

Claude Code: The gap: Faker generates realistic values but has no schema/constraint layer or multi-format export; Mockaroo does both but is a hosted web tool, not a CLI you can script into a test pipeline. synth-data sits in that gap — a scriptable CLI with a constraint-aware schema and native multi-format export. Approve

Stage 1.3: Feasibility & Constraints (product agent)

Claude Code: Feasible — every dependency (Faker, pandas, pyarrow, click) is mature and MIT/BSD licensed. One real constraint: Parquet output requires pyarrow, a heavier dependency than the others — worth making it an optional extra later, not blocking the MVP now. Written to constraints.md. Approve

Stage 1.4: Scope Definition (product agent)

Claude Code: scope-document.md confirms the MVP boundary from vision.md: schema definition, generation, four export formats. Explicitly out: a web UI, cross-table referential integrity, streaming generation. Approve

Stage 1.5: Team Formation (product agent)

For a solo project this stage still runs, just briefly:

Claude Code: Single owner, no team to form — recorded in team-formation.md as "solo, no delegation needed." If contributors join later, this file is where their areas of ownership would get defined. Approve

Stage 1.6: Rough Mockups (design agent)

Claude Code: A mockup of the CLI's actual usage, since there's no visual UI to sketch: text $ synth-data generate --schema schema.yaml --rows 10000 --format csv --out data.csv Generating 10,000 rows... done (0.8s) Wrote data.csv Written to rough-mockups.md. Approve

Stage 1.7: Approval & Handoff (product agent)

The last Ideation stage summarizes everything above into one handoff artifact:

# Ideation Complete — Handoff to Inception

| Artifact | Decision |
|---|---|
| intent-statement.md | Personal + eventual public tool |
| market-research.md | Fills a real gap vs. Faker/Mockaroo |
| constraints.md | Parquet needs pyarrow — flag as optional extra |
| scope-document.md | Schema, generation, 4 formats — no UI, no streaming |
| rough-mockups.md | CLI usage sketch |

▸ [Approve — Continue to Inception / Request Changes]

You: Approve

Behind the scenes: aidlc-state.md now shows Progress: 7/33 overall | 7/7 IDEATION stages complete. Every artifact this phase produced is committed under aidlc/spaces/default/intents/260909-synthetic-data-generator/ideation/ — nothing here is thrown away once we move on; Inception's Requirements Analysis reads intent-statement.md and scope-document.md directly rather than re-asking us the same questions.

‹ Back to: Part 1 — Setup & Kickoff Continue to: Part 3 — Inception Phase