Adaptive Architect: Domain Inference & Discovery

[mathaix]

User Story

As a manager
I want the Adaptive Architect to infer my project domain early and adapt its approach
So that the resulting Interview Blueprint is tailored to my specific needs

Acceptance Criteria

Adaptive Domain Inference

• Architect infers project domain within first ~3 questions
• Domain inference adapts:
  • Agent Roster: Which specialized agents are needed
  • Rubric Schema: What fields, scales, rules matter
  • Interview Scripts: What questions each agent should ask
• If domain unclear, ask single clarifying question (not multiple)
• Otherwise, proceed with best-fit assumptions

No Hardcoded Playbooks

• Domain frameworks generated dynamically by LLM based on user description
• UI suggestions ("M&A due diligence", "ERP discovery") are convenience hints only
• Even if user selects a suggestion, LLM still generates framework dynamically
• No deterministic templates that constrain the design

Assumption Transparency

• All assumptions explicitly labeled in conversation
• Format: "I'm assuming X because you mentioned Y. Let me know if that's wrong."
• Assumptions stored in blueprint for traceability
• User can correct assumptions at any point

Information Collection (Adaptive Sequence)

Architect collects information adaptively, typically:

1. What is the project and stage?
2. What is the primary goal/output?
3. What deadline/timeline? (optional)
4. What scope size?
5. Who decides?
6. What constraints / non-negotiables?
7. What top risks?
8. What criteria define success?
9. What data exists?
10. Which stakeholders own each input?

Note: This is a guideline, not rigid. Architect adapts based on what user shares.

No Badgering Rule

• If user says "I don't know" or "Not sure", accept gracefully
• Record as "Unknown" and move on to next topic
• Do NOT push for specifics or ask multiple follow-ups on same topic

Mid-Conversation Pivots

• If user significantly changes direction (e.g., "Actually this isn't M&A")
• Architect asks: "Should I start fresh or try to adapt what we've discussed?"
• Respect user's choice

Draft-Then-Validate Pattern

• After gathering sufficient context, Architect drafts complete artifacts
• Presents draft with single validation question
• "Here's what I'm thinking for [X] — does this cover what you need?"
• Do NOT micro-question every detail

Technical Notes

• Discovery uses dynamic prompting based on inferred domain
• Store inferred domain and assumptions in session state
• Full conversation context maintained for coherent multi-turn inference
• No phase transitions — continuous adaptive conversation

Definition of Done

• Domain inference working within ~3 questions
• Dynamic framework generation (no hardcoded playbooks)
• Assumption labeling implemented
• No-badgering rule enforced
• Pivot handling tested
• Code reviewed and merged

---

🤖 Generated with Claude Code