Second Opinion

Second Opinion

Purpose

Ask two strong models for independent advice on one concrete matter:

• GPT-5.6 Sol with xhigh reasoning
• Claude Fable 5 with max effort

This is the dial-a-friend workflow. It answers “Given these facts, what would you do and why?” It does not create a durable internet-research artifact; use create-guide for that.

kokoro-coreml runs Kokoro-82M TTS on Apple Silicon via Core ML / ANE. Second opinions usually concern stage placement (ANE vs GPU vs CPU), Core ML conversion traps, latency/quality tradeoffs, or Swift pipeline design — not inventing trainer WER claims for other repos.

Invocation and authority

Run only when the user explicitly asks for a second opinion, dial-a-friend, outside advice, both named models, or equivalent. Explicit natural-language requests count; the literal skill name is not required.

Invocation authorizes two read-only external model calls and local artifacts under tmp/second-opinions/. It does not authorize code edits, commits, pushes, or other mutations.

Procedure

1. Frame a rich investigation brief

The brief is an entry point, not a substitute for repo research. Sparse context produces generic advice and anchors both reviewers to the parent agent's possibly stale framing.

Write a detailed markdown brief containing:

1. North-star outcome: the higher-level system or product result we are actually trying to achieve, why it matters, and how success is measured. Include non-negotiable quality constraints. Example: “minimize end-to-end TTS latency on M-series without degrading audio quality by choosing which Kokoro stages run on ANE, GPU, or CPU.”
2. Tactical question: the exact decision, diagnosis, or tradeoff currently under consideration.
3. Why this tactic might serve the north star: the hypothesized causal link and why this is believed to be the current bottleneck.
4. Current understanding: measured evidence, assumptions, unresolved contradictions, and why the question is difficult.
5. Constraints: fixed policy, budget, compatibility, safety, and non-goals. Never include live credentials, .env secrets, or customer scrape payloads in the brief.
6. Options considered: including rejected paths and why they were rejected.
7. Repo starting points: relevant code, tests, plans, notes, guides, notes, plans, receipts, and recent diffs or commits.
8. Requested output: multiple potential directions, each direction's steelman and strongest counter-argument, a final recommendation, risks, confidence, and the evidence that would reverse the recommendation.

Do not hide inconvenient facts. Distinguish measurements from assumptions. Name starting paths, but do not curate the evidence boundary so narrowly that reviewers cannot discover newer or contradictory repo state.

Explicitly invite both reviewers to:

• restate the north-star outcome before evaluating the tactical question
• test whether the tactical question attacks the real bottleneck or merely optimizes a local component
• propose a different lever, decomposition, or architecture when it would reach the north star more directly
• orient themselves in the repository before answering
• inspect all code, tests, plans, notes, guides, notes, plans, and receipts relevant to the question—not only files linked in the brief
• search for newer evidence that supersedes the brief
• reconcile documentation against implementation and firsthand measurements
• use git status, diffs, and recent history when they affect the answer
• report which repo evidence materially drove the recommendation
• respect secrets: do not quote credentials or raw customer payloads

The agent preparing the brief should inspect the repo first and provide strong starting points. The delegated reviewers must still perform their own repo investigation; do not ask them to reason only from the summary.

Save the brief under tmp/second-opinions/. Do not place it in README/Guides/ or README/Notes/ unless the user separately asks for a durable document.

2. Assign an expert persona

Give both reviewers the same domain-appropriate persona so the comparison tests reasoning, not role ambiguity.

For ML, Core ML, ANE, TTS, evaluation, or Apple Silicon questions, default to the repo's Ilya Sutskever persona:

• persona skill
• persona source

Use the persona's actual text—not merely “act like Ilya.” The useful stance is learning-first and empirical: respect the Bitter Lesson, prefer simple objectives and decisive experiments, treat data and scale as first-class, and avoid false certainty.

For other non-ML questions, choose another existing repo persona only when it genuinely fits. Otherwise use a neutral skeptical principal-engineer persona. Pass a non-default persona with:

SECOND_OPINION_PERSONA_FILE=<persona-path> \
  scripts/run-second-opinion.sh <brief-path>

3. Launch fresh independent reviewers

Run:

scripts/run-second-opinion.sh <brief-path>

The helper launches separate fresh CLI threads in parallel:

• codex exec --model gpt-5.6-sol with model_reasoning_effort="xhigh" and a read-only sandbox
• claude -p --model claude-fable-5 --effort max in plan mode with --tools "Read,Glob,Grep,Bash" — Bash for read-only investigation only

Both receive the same prompt. Neither sees the other opinion. Never seed one reviewer with the first reviewer’s answer.

The helper defaults to the full Ilya persona and tells each reviewer to perform a broad, read-only repo investigation before answering. The brief's linked files are starting points, not a whitelist.

The generated prompt puts the full persona before task instructions. It requires reviewers to reason at two levels:

1. the north-star outcome the project is trying to achieve
2. the tactical question currently proposed as a way to get there

Reviewers must challenge whether the tactic attacks the true bottleneck and surface broader alternatives when another lever reaches the north star more directly. The prompt also states the review's purpose: improve the owner's decision, provide a fresh perspective, expand the option space, and identify decisive evidence. Each reviewer must propose multiple plausible directions when warranted and, for each, provide:

• the strongest steelman
• the strongest counter-argument and failure conditions
• the smallest observation or experiment that discriminates it from alternatives

Reviewers must not invent weak alternatives to meet a quota; they should say when evidence leaves only one viable path.

The model and effort pins are part of the contract. If either exact model or effort is unavailable, report that reviewer as unavailable. Do not silently substitute another model or lower effort.

4. Read and adjudicate

Read:

• gpt-5.6-sol-opinion.md
• claude-fable-5-opinion.md
• status.txt

Then produce your own synthesis. Do not concatenate the answers and do not use majority vote. Check each argument against local evidence and canonical repo docs (README, notes, guides, Core ML receipts).

Use this compact structure:

## North-star outcome

## Shared conclusion

## Where they disagree

## My adjudication

## Unknowns that could change the answer

State plainly when:

• both models agree but rely on the same unsupported assumption
• one opinion is stronger because it uses better evidence
• local firsthand evidence overrides either model
• the question remains genuinely unresolved
• one reviewer failed or was unavailable

5. Preserve the boundary

The output is advice for the current matter, not source-of-truth documentation. If the run reveals durable local learning, use write-notes separately. If it reveals a missing external-knowledge domain, use create-guide separately with an internet-synthesis brief.

Do Not Use When

• The user wants an “afternoon of Googling” synthesis → create-guide
• A raw research export needs integration → guide-ingest
• The user wants implementation → use the appropriate implementation workflow
• The question is routine and a second model would add cost without meaningful independent judgment
• The user asks for only one named model → invoke that model directly instead of pretending this two-opinion contract ran

Failure handling

• Missing CLI, auth failure, empty output, parse failure, or timeout: read status.txt and the per-reviewer opinion stubs the helper wrote; continue with the available reviewer and disclose that the result is not a two-model opinion.
• Do not silently substitute another model or lower effort.
• Both fail: report the blocker and stop; answer locally only if the user asks for a non-delegated opinion.

References

• Helper: ../../../scripts/run-second-opinion.sh
• Cross-agent precedent: ../execute-plan/SKILL.md
• Internet-synthesis boundary: ../create-guide/SKILL.md
• Default persona: ../ilya-sutskever/ilya-sutskever.md