CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

second-opinion

Gets a blind review of a diff or a PR from a model of another provider (Claude, GPT, Gemini), then checks and compares its findings. Use when the user asks for a second opinion, an adversarial review, or a review by another model.

Quality

88%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

SKILL.md
Quality
Evals
Security

Second opinion

A second opinion is a blind review: a model from another provider, in a fresh session, judges the change without seeing your view. Its value is independence, and another provider has other blind spots than yours. So the brief carries the goal and the change, and keeps your conclusions back.

Steps

  1. Pin the change. The base ref and head (git diff <base>...HEAD), or the PR number. Done when you can name it by ref or URL.

  2. Pick the reviewer. A model of another provider than yours, at its default effort. When no other provider is available, use a stronger model of your provider (for example Fable instead of Opus), or your own model at a higher reasoning effort. The user's choice wins.

  3. Write your own view first. A few lines: the risks you see and your verdict. Keep it out of the brief. It is your baseline for step 6.

  4. Write the brief to a temporary file:

    • Context: what the change is for, in two or three sentences, and the spec (the issue, or the PR body).
    • Change: the refs and the diff command, or the PR URL.
    • Known and accepted: findings already judged and left on purpose, so the reviewer skips them.
    • Rule: the section "Edge cases and guards" from the code-review rules, pasted in full.
    • Scope, verbatim: "You may read files and run read-only commands. Change nothing, and post nothing to GitHub: this is a review."
    • Output, verbatim: "Reply with a one-line verdict, then numbered findings, most severe first. For each: severity (blocker, major, minor, nit), file:line, the problem, a concrete failure scenario (inputs, then the wrong result), the producer of its input, and your confidence. End with the files and commands you checked."
  5. Run the reviewer from the checkout, read-only, in the background. Use the first that your environment has:

    • A delegation tool of your agent environment that can start a session on another provider (for example T3 Code's delegate_task).
    • The other provider's CLI: codex exec -s read-only -m <model> -o <reply-file> - < <brief-file>, or claude -p --permission-mode plan --model <model> < <brief-file> > <reply-file>.
    • A subagent on the model from step 2, when no other provider is available. Give it the brief only, not this conversation.

    Tell the user which model reviews. A reviewer of your own provider shares more of your blind spots, so say so in the report.

  6. Check and compare. The reviewer can be wrong. Check each finding against the code yourself, and give it a verdict: confirmed, or disputed (with the reason). For a finding that names no producer of its input, judge it as "Edge cases and guards" says. Then report: both verdicts, new confirmed findings your baseline missed, shared findings, disagreements, and what neither of you checked.

Done when every finding has a verdict with a reason.

Repository

storybookjs/storybook

Path

.agents/skills/second-opinion/SKILL.md

Last updated
First committed

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.