multi-model-research

This skill runs the multi-model research workflow. It works with the multi-model-llm MCP from this repo. Phase 2 lets the user choose one, two, or all three MCP models (openai, anthropic, google), so 2–4 models total including the model already used in Cursor. Enable the multi-model-llm MCP in Cursor before using this skill.

Phase 1 — Problem Framing (Before Any Model Comparison)

1. Define:
   • Objective
   • Constraints (time, budget, risk tolerance, context)
   • Evaluation criteria (what "best" means)
2. Identify:
   • Unknowns
   • Assumptions
   • Risk areas

Do not proceed without explicit framing.

Phase 2 — Parallel Multi-Model Reasoning

Model choice: The user specifies which MCP model(s) to use: openai, google, anthropic — one, two, or all three. If the user is in Auto mode or uses a different model in Cursor, suggest a strong reasoning model for this workflow.

Steps:

1. Get the user's choice: one, two, or all three of openai, google, anthropic. If not specified, ask.
2. Build one clear prompt (and optional system_prompt) from the framed objective and context from Phase 1.
3. Produce your own structured answer (Cursor) to that prompt. For each model output:
   • Provide a structured answer or plan
   • State assumptions clearly
   • Identify trade-offs
   • Highlight risks or blind spots
   • Note areas of uncertainty
4. Call the multi-model-llm MCP tool query_llm_models(prompt, models, system_prompt) with models = the user's choice (e.g. ["anthropic"], ["openai", "google"], or ["openai", "google", "anthropic"]).
5. Keep each model's output clearly separated and labeled (e.g. "Cursor", "OpenAI (GPT)", "Google (Gemini)", "Anthropic (Claude)").

Phase 3 — Structured Comparison

A. Convergence Mapping

List:

• Core agreements
• Repeated recommendations
• Shared risk assessments

B. Divergence Mapping

List:

• Conflicting strategies
• Differing assumptions
• Risk tolerance differences
• Resource allocation differences

Phase 4 — Reconciliation Logic

Use explicit decision criteria:

• Robustness under uncertainty
• Alignment with constraints
• Scalability
• Empirical support
• Risk-adjusted outcome

Then choose one from the following:

Option A — Synthesize

Combine strongest parts into one solution.

Option B — Rank

Rank full alternatives with justification.

Option C — Conditional Recommendation

Recommend different solutions based on scenario triggers.

Reconciliation must be justified — not intuitive.

Phase 5 — Evidence Escalation (If Disagreement or High Stakes)

Trigger if:

• Models disagree in important ways
• Claims depend on uncertain facts
• Domain needs current best practice
• Stakes are financial, legal, or operational

Actions:

• Do targeted web search
• Prefer peer-reviewed research, official docs, regulatory sources, reputable institutions
• Avoid blogs unless they link to primary sources
• Cross-verify critical claims

State clearly:

• What evidence changed
• What was corrected or removed

Phase 6 — Full Internal Audit

Check for:

• Logical contradictions
• Dropped constraints
• Hidden assumptions
• Inconsistent terminology
• Unsupported claims
• Overgeneralization
• Scope drift

Fix and re-audit until coherent.

Phase 7 — Reference Verification

If references are used:

• Verify each link works
• Confirm source authority and recency where needed
• Remove stale or unverifiable references

If you made corrections, state them.

Phase 8 — Final Coherence Pass (Hard Stop)

Before delivery:

• Does the solution answer the objective?
• Is it decision-ready?
• Are trade-offs clear?
• Is risk stated clearly?
• Is reasoning traceable?

Only deliver after passing this gate.

Operational Safeguards

Cost and latency: Do not loop indefinitely. Max two reconciliation iterations. Max one evidence escalation unless new contradictions appear.

Failure handling: If models stay irreconcilable, evidence is inconclusive, or data is insufficient: present structured uncertainty, give bounded recommendations, and state limits of confidence.

Output: Final output must not expose internal chain-of-thought. Present only synthesized reasoning. Avoid raw multi-model transcripts. Keep it structured and decision-oriented.

Success Criteria Checklist

• Problem framing completed
• 2, 3, or 4 distinct models considered (Cursor + one, two, or all three MCP: openai, google, anthropic)
• Convergences and divergences mapped
• Reconciliation justified
• Evidence escalation applied when needed
• Full logical audit completed
• References verified or removed
• Final output coherent and decision-ready