Skill — Multi-LLM Consult
When this skill activates
When seeking a second opinion from external models, validating a decision across
multiple AI providers, or when the user explicitly requests cross-model consultation.
Mandatory actions when this skill is active
Before consulting external models
- Sanitize the prompt. NEVER send raw project context to external models.
- Remove: file paths, internal variable names, proprietary business logic
- Remove: API keys, secrets, credentials, internal URLs
- Remove: user PII, customer data, anything covered by data-privacy skill
- Keep: the abstract question, general patterns, public knowledge references
- Estimate cost. Each external call costs tokens. Check budget via cost-tracking module.
- Define the question clearly. Vague questions produce vague answers. Frame as:
- "Given [sanitized context], which approach is better: A or B? Why?"
Configured Models
| Provider |
Model |
Best For |
Cost Tier |
| Anthropic |
claude-opus-4-7 |
Deep reasoning, architecture |
complex |
| Google |
gemini-2.5-pro |
Research, long context, web grounding |
research |
| OpenAI |
gpt-4o |
Alternative perspective, validation |
consult |
Consultation Protocol
Single Consult (one external model):
- Sanitize prompt
- Send to selected model
- Present response with source attribution
- Note areas of agreement/disagreement with primary analysis
Consensus Consult (all 3 models):
- Sanitize prompt (same prompt to all)
- Send to all configured models in parallel
- Analyze responses for:
- Agreement (2+ models recommend same approach): high confidence signal
- Divergence (models disagree): flag for user decision, present all perspectives
- Novel insight (one model raises a point others missed): highlight specifically
- Produce synthesis:
Consensus: [Yes/No/Partial]
Recommended: [approach]
Agreement: [which models agree]
Dissent: [which models disagree and why]
Novel: [unique insights from individual models]
During consultation
- Log every external call in token-usage.jsonl (model, tokens, cost)
- Never send more than 2000 tokens to external models per consultation
- If a model is unavailable: skip it, note in output, continue with available models
- Respect rate limits — max 3 consultations per session
After consultation
- Present results to user with clear attribution
- Never auto-execute based on external model recommendations
- External opinions are ADVISORY — user sovereignty applies
- Log consultation summary in AUDIT
Self-check before task completion
1---2name: multi-llm-consult3description: Skill — Multi-LLM Consult4---56# Skill — Multi-LLM Consult78## When this skill activates9When seeking a second opinion from external models, validating a decision across10multiple AI providers, or when the user explicitly requests cross-model consultation.1112## Mandatory actions when this skill is active1314### Before consulting external models151. **Sanitize the prompt.** NEVER send raw project context to external models.16 - Remove: file paths, internal variable names, proprietary business logic17 - Remove: API keys, secrets, credentials, internal URLs18 - Remove: user PII, customer data, anything covered by data-privacy skill19 - Keep: the abstract question, general patterns, public knowledge references202. **Estimate cost.** Each external call costs tokens. Check budget via cost-tracking module.213. **Define the question clearly.** Vague questions produce vague answers. Frame as:22 - "Given [sanitized context], which approach is better: A or B? Why?"2324### Configured Models2526| Provider | Model | Best For | Cost Tier |27|----------|-------|----------|-----------|28| Anthropic | claude-opus-4-7 | Deep reasoning, architecture | complex |29| Google | gemini-2.5-pro | Research, long context, web grounding | research |30| OpenAI | gpt-4o | Alternative perspective, validation | consult |3132### Consultation Protocol3334**Single Consult (one external model):**351. Sanitize prompt362. Send to selected model373. Present response with source attribution384. Note areas of agreement/disagreement with primary analysis3940**Consensus Consult (all 3 models):**411. Sanitize prompt (same prompt to all)422. Send to all configured models in parallel433. Analyze responses for:44 - **Agreement** (2+ models recommend same approach): high confidence signal45 - **Divergence** (models disagree): flag for user decision, present all perspectives46 - **Novel insight** (one model raises a point others missed): highlight specifically474. Produce synthesis:48 ```49 Consensus: [Yes/No/Partial]50 Recommended: [approach]51 Agreement: [which models agree]52 Dissent: [which models disagree and why]53 Novel: [unique insights from individual models]54 ```5556### During consultation57- Log every external call in token-usage.jsonl (model, tokens, cost)58- Never send more than 2000 tokens to external models per consultation59- If a model is unavailable: skip it, note in output, continue with available models60- Respect rate limits — max 3 consultations per session6162### After consultation63- Present results to user with clear attribution64- Never auto-execute based on external model recommendations65- External opinions are ADVISORY — user sovereignty applies66- Log consultation summary in AUDIT6768## Self-check before task completion69- [ ] Did I sanitize the prompt before sending to external models?70- [ ] Did I log every external call in token-usage.jsonl?71- [ ] Did I attribute responses to their source model (no unattributed blending)?72- [ ] Did I remind the user that external opinions are advisory?
sairam0424