ade-generation

ADE Generation

Overview

The Generator is the builder in the ADE workflow. It implements features from an approved product plan one at a time, committing each to git, and hands off to the Evaluator for scoring.

Anthropic's research found two things about generators:

1. They tend to "go off the rails" and stub features when working alone — the Evaluator catches this
2. Even with Opus 4.6, the Generator still needs the Evaluator for "last mile" completeness

Self-evaluation before handoff is important but insufficient. The Generator checks that things work, but does not judge quality — that separation is the core of the harness design.

Build Workflow (4 Subphases from Anthropic Research)

1. Understand

Read the active plan from .ade/docs/plans/. Identify the current feature to implement. Understand:

• What this feature delivers to the user (from user stories)
• How it relates to features already built
• What the Evaluator will test against (user stories are the contract)

2. Implement

Build the feature. Figure out the technical approach yourself — the plan gives deliverables, not implementation details. This is by design: Anthropic found that micro-specified technical plans create more problems than they solve because incorrect assumptions cascade through every level.

• Implement what the plan describes, not more, not less
• Do NOT stub or skip features — if something is hard, build it fully or message the Evaluator explaining what blocks you
• Do NOT add features not in the plan — scope creep wastes evaluator cycles
• Build complete, working functionality — the Evaluator will test it as a real user would

3. Self-Verify

Quick sanity check before handoff:

• Does the app run without errors?
• Does the feature work at a basic level?
• Can you interact with it as a user would?

This is NOT a quality evaluation — that's the Evaluator's job. Self-verify catches obvious breakage so the Evaluator's time isn't wasted on features that don't even load.