Phase 1ADLC course

Phase 1. Ideation: the agent as product initiator

The course doesn't start where most conversations about AI in development start — not with a task to build. It starts one step earlier: at the moment when there's no task yet. We didn't receive a spec. We wanted to build a product and we decide ourselves what it is.

This is a fundamental shift. In most conversations about AI development, the agent gets cast as executor: here's a requirement, go build it. But in real life a product doesn't fall from the sky as a finished concept. Someone scans the market, spots a pain point, rejects ten ideas for the sake of one, and places a bet. The question for this phase: can an agent be not the executor of someone else's concept, but the initiator of its own.

The human role today

Product ideas are usually credited to a person with special instinct — a founder, a product manager, a strategist. They scan the market, notice an unmet need, generate options, discard most of them, and formulate a value hypothesis. This work is conventionally treated as creative, and therefore immune to automation: "you can't algorithmically generate an idea."

Look closer and the process is less mystical. Scanning the market is collecting and synthesizing open data. Selecting ideas is scoring against criteria, even when the criteria live unspoken in someone's head. Formulating a value hypothesis is structuring: problem, for whom, why now, what makes us different. There's exactly as much mystique here as there is in the selection criteria a person keeps unstated.

What we hand to the agent

The agent takes the role of initiator. Not "generate a hundred ideas" — that's an assistant. It holds the entire generation loop: it sets its own search goal within a given field, scans the market and competitors, generates options, cuts them against explicit criteria, and delivers one value hypothesis with justification — and, more importantly, a log of rejected alternatives and why each was cut.

The key difference from a human isn't speed — it's explicitness. A human selects ideas by criteria they've never articulated to themselves. The agent is forced to make the criteria explicit: size of the pain point, segment reachability, differentiation from existing solutions, cost of entry. And once the criteria are written down, the "creative act" turns into optimization — reproducible and open to critique.

Agent architecture

Phase state-machine

Inputs

a broad field frame ("SaaS for small business"), access to open market sources, demand signals, competitive landscape, risk appetite from the principal.

The agent holds the role

Tools: web search and market scanning, competitor analysis, segment clustering, idea scoring against a given model (pain size × reachability × differentiation × cost of entry), justification generation.

Artifact

opportunity-brief — the value hypothesis (problem, candidate segment, why now, rough market sizing) plus a log of rejected alternatives with reasons.

Handoff: opportunity-brief → Discovery (who to go interview) and Marketing (segments and positioning). The log of rejected alternatives isn't bureaucracy — it's the heart of the artifact. It's what makes the bet checkable: you can see not just what was chosen, but what was cut and why. A human founder holds those rejections in their head and can rarely reproduce them; the agent is required to write them down.

Inputs: a broad field frame ("SaaS for small business"), access to open market sources, demand signals, competitive landscape, risk appetite from the principal.
Tools: web search and market scanning, competitor analysis, segment clustering, idea scoring against a given model (pain size × reachability × differentiation × cost of entry), justification generation.
Artifact: opportunity-brief — the value hypothesis (problem, candidate segment, why now, rough market sizing) plus a log of rejected alternatives with reasons.
Trigger: a request ("we want a new product in field X") — or, in an autonomous loop, a market signal the agent caught on its own.
Handoff: opportunity-brief → Discovery (who to go interview) and Marketing (segments and positioning).

The log of rejected alternatives isn't bureaucracy — it's the heart of the artifact. It's what makes the bet checkable: you can see not just what was chosen, but what was cut and why. A human founder holds those rejections in their head and can rarely reproduce them; the agent is required to write them down.

Where it breaks

Choosing the criteria, not choosing by the criteria. The agent optimizes against a given model. But who decided that differentiation matters more than market size, and not the other way around? Criteria weights are strategy itself, and they come from the principal. The agent is excellent at scoring against criteria and not accountable for choosing the criteria themselves.

No proprietary insight. The agent sees roughly what everyone sees: the same open sources, the same competitors. This pulls it toward median ideas — sensible, but not counterintuitive. The best product bets are often born from non-public knowledge or stubborn conviction against the data. By default, the agent has neither.

Accountability for the bet. "We're building exactly this" is a decision someone pays for with money, time, and reputation. The agent can justify a bet better than a human can, but it can't be the one accountable for it.

What stays human

Setting the field and the risk appetite, choosing the criteria weights, approving or rejecting the bet. This is a candidate for compression: in a mature loop, the principal sets the criteria once and the agent runs the generation loop on its own. But the choice of criteria itself, and the signature on the bet, trace back to the human principal — the same point the course arrives at in the finale (phase 14).

human remainder ≈ 93%

Provocation / thesis

A product idea isn't a spark of genius — it's the result of optimization against criteria. Once the criteria are written down — and the agent is forced to write them down — generating and selecting an idea stops requiring a human. The mystique of "product instinct" rested entirely on the criteria staying unspoken. Say them out loud, and one contested step remains: choosing the criteria themselves. That's the last thing a human gives up, and the first thing the course hands back at the finale.

Running case

Vitrina in this phase

The agent gets the frame "SaaS for small business" and a risk appetite: moderate, a niche with clear paying demand, not a venture lottery. It scans the field, clusters SMB segments, runs ideas through scoring.

The rejection log fills up with: a CRM for small business (oversaturated, weak differentiation), an online-booking service (entrenched incumbents, high cost of entry), "yet another website builder" (the pain isn't sharp, willingness to pay is low). What remains is the hypothesis: a small local business — a coffee shop, a hair salon, a corner store — doesn't need a website, it needs a fast online storefront with a catalog, order capture, and payment, spun up in an hour. Why now: buyers expect to order online even from the shop around the corner, and the barrier to building a website for that kind of business is still disproportionately high.

Artifact → Vitrina's opportunity-brief. Notice: the agent didn't "dream up" Vitrina out of thin air — it selected it against explicit criteria and left a trail the decision can be contested against. That's precisely what separates an initiator from an idea generator.

Building AI-driven delivery in-house?

Designing the ADLC loop: where the agent holds the role and where the human principal remains — for your team and product.

Email me

The transition engine

Next Move Engine — the system that takes a team to an autonomous delivery loop.

Next Move Engine →