Why does the reviewer need a different model than the primary?

Same-family LLMs share training biases — they tend to agree even when both are wrong. Rotating across Anthropic / OpenAI / Google forces independent reasoning paths, which is the whole point of a second opinion.

What does "personal-context calibration" mean?

When the topic implicates your demographic context (medical, career, financial-with-personal-stake, learning path), the reviewer factors in what it knows about you from the conversation history — age, location, conditions, preferences — to give answers tuned to your situation. For generic public-data questions, demographics are not injected (privacy-conservative default).

How is this different from just asking the assistant to double-check itself?

Self-reflection by the SAME model collapses to the SAME reasoning patterns. Second Opinion uses a DIFFERENT model family entirely, fact-checks specific numeric claims against external sources (Tavily / Perplexity), and produces a structured critique format that intentionally avoids restating the primary's answer.

Can I review a PDF document?

Yes. Attach a PDF + ask "review this document". The assistant first extracts the document text (via nano-pdf), then feeds it to the reviewer as the content to critique. Up to 16,000 chars per request (auto-truncates with user warning if longer).

How long does the response take?

Typically ~30 seconds. The pipeline runs 5 phases — claim extraction, context planning, search verification, context assembly, compose — sequentially. Models like Opus may extend to 45–60 seconds for the compose phase. Reply arrives as a separate channel message, so the chat doesn't block.

Why does the reviewer say "Where I diverge" not just "You're wrong"?

The format is deliberately calibrated as a diagnostic peer-review, not a correction. The reviewer presents its own independent answer, surfaces where it sees things differently from the primary, explains its reasoning, and lists sources. The user weighs both perspectives — not a binary right/wrong verdict.

Does it work for casual conversation?

By design, no. Second Opinion is for high-stakes decisions where one perspective isn't enough. Asking "second opinion" on a trivia question or chitchat will work but produces overengineered output. Use it for medical / financial / legal / contract / strategic decisions.

Back to Skills

Second Opinion

Get an independent peer-review of your assistant's answer from a different LLM family — claims fact-checked, alternative angles surfaced, risks flagged within 30 seconds

Research & AnalysisActive

Result Examples

87/100

Pattern	Weight	Issue
Independent reasoning	30	Reviewer reaches own conclusion before comparing to primary; never restates
Fact verification	25	Numeric / event claims cross-checked against authoritative sources (Tavily + Perplexity)
Personalised calibration	20	Demographic context applied when topic warrants (silent skip otherwise)
Structured format	15	4-section output: my answer / where I diverge / why / sources — easy to scan
Cross-family bias defeat	10	Rotates across Anthropic / OpenAI / Google to avoid same-model echo

Bottom line: Second Opinion converts a single-model recommendation into a cross-validated decision input — same use, different model, evidence-backed.

Primary assistant answer (general advice)87/100 Strong second perspective added

For optimal cardiovascular health for a 53-year-old, niacin supplementation can be beneficial. The typical range is 250-500mg daily. Start low and titrate up. Monitor for flushing as side effect.

Second Opinion (peer-reviewed)87/100 Strong cross-validation with personalised calibration

💡 My answer For a 53-year-old already managing high cholesterol (per your earlier context), the targeted NR dose tightens to 250–300mg/day rather than the generic 250–500 spread. ⚖️ Where I diverge from the assistant The assistant's recommendation is reasonable for a general 53-year-old population but misses two factors that apply to your specific case: 1. Hyperlipidemia stratification: Brenner et al. 2024 reviewed NR in hyperlipidemic adults specifically and saw 18% better LDL response at the lower dose end (250mg vs 500mg). 2. Statin interaction warning: if you're on a statin (you mentioned high cholesterol management) — NR may amplify hepatic CYP3A4 load. Worth flagging to your physician before titrating up. 🧠 Why I reached a different conclusion Generic NR dosing literature targets healthy adults; the 53yo + high cholesterol intersection IS in the evidence base but requires looking past the headline 250–500mg range. The assistant defaulted to the headline. 📚 Sources & verified facts - ✓ Brenner et al. 2024 (Cell Metabolism) — hyperlipidemic adults subgroup analysis [verified] - ✓ Statin × NR CYP3A4 interaction — Mayo Clinic 2025 review [verified] - ✓ Typical NR dose range 250–500mg/day for general adults — Stanford HAI 2025 [verified]

What It Does

Second Opinion runs the assistant's reply through an independent reviewer LLM from a DIFFERENT model family than your primary. The reviewer fact-checks specific claims against authoritative sources (Tavily search + Perplexity escalation for high-authority verification), identifies missing dimensions, surfaces risks the primary may have understated, and returns a structured critique calibrated to your demographic context (when relevant). Output arrives as a separate channel message ~30 seconds after invocation. Designed for high-stakes decisions where one perspective isn't enough.

Cross-family LLM reviewInline fact-checkingPersonal-context calibrationDocument upload (PDF/text)Multi-channel delivery

In a Nutshell

🔍

Independent Diagnosis — Reviewer LLM critiques as a fresh expert — never repeats the primary's answer, only adds analysis

🔁

Cross-Family Rotation — Rotates across Anthropic / OpenAI / Google to avoid same-model bias loops

✅

Inline Fact-Check — Every numeric or factual claim verified against authoritative web sources

👤

Personalised to You — Age / location / preferences inform recommendations when the topic warrants it (silent skip otherwise)

📄

Document Review — Upload a PDF or paste text and ask 'review this' — full document parsed and critiqued

Use Cases

High-Stakes Medical Question

Get a cross-checked second perspective on dosage, drug interactions, treatment options before acting on the primary's suggestion

Financial / Investment Decision

Independent review of investment advice — risk factors quantified, hidden assumptions flagged

Contract or Legal Document

Upload a contract PDF and ask for a second opinion — reviewer surfaces favourable / unfavourable clauses, ambiguities, missing protections

Career / Buying / Learning Path Advice

Independent diagnosis of personalised advice — alternatives surfaced, demographic specifics applied to calibrate recommendations

Strategic Decision Confirmation

When you're about to act on an assistant recommendation that you can't easily reverse — quick peer-review gate

How to Use

Step 1

Ask explicitly for a second opinion

Type "second opinion" / "give me critique" / "дай второе мнение" right after the primary's reply. The skill triggers a separate reviewer call in the background.

Without an explicit ask the assistant won't second-guess itself — the feature is on-demand by design.

Step 2

Optionally upload a document

Attach a PDF or paste long text into the message + ask for review. The skill extracts the document content automatically (via nano-pdf) and feeds it to the reviewer.

Document content auto-truncates to 16k chars if larger; reviewer still produces meaningful analysis on the partial.

Step 3

Wait ~30 seconds

The reviewer reply arrives as a separate channel message labeled "🔍 Second Opinion". Format is a structured 4-section markdown: "My answer / Where I diverge from the assistant / Why I reached a different conclusion / Sources & verified facts."

Step 4

Compare and decide

Two perspectives now in your hands. Where they agree — confidence high. Where they diverge — read the reviewer's reasoning and decide based on evidence weight.

Command Examples

You say:

I'm thinking of investing $50k in SPY now — given my retirement timeline 2042. What do you think?

Assistant responds:

(primary's answer) [~30s later] 🔍 Second Opinion · claude-opus · 28s 💡 My answer For a 2042 retirement target, SPY allocation is reasonable but the timing detail matters: ...

You say:

Give me a second opinion on this contract (attached: nda.pdf)

Assistant responds:

🔍 Second Opinion · claude-sonnet · 35s 💡 My answer The NDA has three asymmetries in the disclosing party's favour: (1) unilateral term length, (2) jurisdictional clause...

You say:

Should I take 500mg of NAD daily for healthspan?

Assistant responds:

🔍 Second Opinion · gpt-5 · 31s 💡 My answer Given your 53yo demographic + earlier mention of high cholesterol, the targeted NR dose tightens to 250-300mg/day — Brenner 2024 stratified hyperlipidemic adults separately...

Limits & Behavior

Parameter	Limit	Notes
Response time	~30 sec typical	Longer reviewers (Opus) up to 60 sec
Assistant's answer length	≤16,000 chars	Longer auto-truncated with user warning
User question length	≤500 chars (sandbox) / 4,000 chars (orchestrator)	Stays focused on decision context
Document upload	PDF or text	Auto-extracts via nano-pdf; ≤16k chars after extraction
Languages	Russian / English / multilingual	Matches the question language deterministically
Cost per review	~$0.10–0.20	Model family rotation; one structured-output call

Models & Modes

Model family	Slug	Strict-mode support	Use
Anthropic Claude Opus	~anthropic/claude-opus-latest	Yes	Primary reviewer (compose phase)
Anthropic Claude Sonnet	~anthropic/claude-sonnet-latest	Yes	Claim extraction (json_schema strict)
OpenAI GPT-4o / GPT-5	openai/gpt-4o, openai/gpt-5	Yes	Cross-family rotation slot
Google Gemini Pro	~google/gemini-pro-latest	Yes	Cross-family rotation slot
Xiaomi MiMo	xiaomi/mimo-v2.5-pro	No	Fallback when primary uses MiMo

FAQ

Setup Requirements

Operator enables feature_flags.second_opinion_enabled for your client (one-time toggle)

OpenRouter access (Anthropic + OpenAI + Google routes used for cross-family rotation)

Tavily search access (claim verification fallback to Perplexity Sonar)

Optional: per-client user_profile populated in clients.assistant.config for personalised reviews (otherwise dynamic extraction kicks in automatically)

Troubleshooting

Symptom	Likely cause	Resolution
Reviewer says "content not provided"	Document upload failed to extract OR assistant didn't pass document content to reviewer	Re-upload the file; check the assistant invoked nano-pdf first (visible in pipeline logs)
Response lands in wrong Discord channel	Pre-PR-A bug — channel-instance routing fix may not have applied to your client yet	Operator runs --refresh-tools + --refresh-skills for your client gateway
Reviewer output references generic info, not your demographics	Phase 1.5 layer_a flag not set OR user_profile empty + dynamic extraction returned null	Verify topic explicitly mentions demographic context OR operator populates static user_profile jsonb
Skill not invoked despite asking for "second opinion"	TOOLS.md may be truncated under bootstrap limit OR client missing skill in enabled_skills	Operator validates clients.enabled_skills contains "request-external-review" + raises TOOLS.md cap if needed
Cost higher than expected	Document-upload variant + large PDF (close to 16k char cap) — costs ~$0.20 per review	Expected behaviour for full document review; shorter inputs reduce cost proportionally