human-ai-system-builder（人类 AI 系统构建者）

Human/AI system builder

The agent MUST build a project-contract before implementation. Start from references/project-contract.md and assets/project-contract.template.yaml.

The agent MUST make every decision below explicit:

• Actors.
• Canonical store.
• Human surface.
• AI surface.
• Sync direction and trigger.
• Conflict detection and resolution.
• Validation timing and location.
• Generated artifacts.
• Human startup flow.

The agent MUST use references/gui-selection.md to choose one of these patterns:

• references/pattern-native-gui-reuse.md
• references/pattern-custom-gui.md
• references/pattern-hybrid-gui.md

For core architecture choices, the agent MUST also use:

• references/canonical-store-selection.md
• references/sync-and-conflict.md
• references/validation-and-launch.md
• references/minimal-questions.md
• references/deliverables.md

For document, form, spreadsheet, and layout-sensitive systems, the agent MUST also read references/domain-structured-docs.md.

Workflow

1. Classify the domain and the primary constraints.
2. Draft the project-contract.
3. Choose the human GUI pattern.
4. Decide the canonical store and AI editing surface.
5. Specify sync, conflict, validation, artifacts, and launch flow.
6. Ask only the irreducible questions that materially change architecture.
7. Choose or adapt the nearest reusable template under assets/templates.
8. Produce the architecture summary, chosen pattern, contract, scaffold plan, and implementation sequence.

Design rules

• The agent MUST keep one canonical source of truth.
• The agent MUST prefer machine-readable, diff-friendly canonical data for AI.
• The agent MUST prefer the most natural human GUI for the job, even when it is an existing application.
• The agent MUST reuse an existing GUI when that GUI is the truth for layout, wrapping, print, or visual constraints.
• The agent MUST show validation in the human primary work surface whenever practical.
• The agent MUST prefer diagnostics in the same active human GUI via inline markers, anchored panels, or in-surface sheets before introducing a separate diagnostics window.
• The agent MUST separate generated artifacts from canonical data.
• The agent MUST NOT rely on silent overwrites for human/AI conflicts.
• The agent MUST make the human launch path low-friction. The agent MUST prefer launcher or double-click flows over CLI for routine human work.

Outputs

The agent MUST produce these outputs unless the user explicitly narrows scope:

• Architecture summary.
• project-contract.
• Chosen GUI pattern and rationale.
• Scaffold layout.
• Implementation plan.
• Acceptance criteria.

The agent MUST use assets/project-contract.schema.json when a machine-readable schema is needed. The agent SHOULD use the example contracts under assets/templates when the request is close to an existing pattern.