second-opinion

second-opinion is a small Codex skill for running a critical review loop on any task.

It tells the agent to:

1. capture the goal, current work, and proof so far
2. ask a fresh reviewer to independently inspect the work
3. have that reviewer research comparable practices when useful
4. triage the feedback by severity
5. fix the accepted findings
6. ask the reviewer to re-review the updated result, or start a fresh reviewer with the prior findings when the surface cannot resume
7. close with evidence, remaining risks, and any deferred work

Use it when you want a second opinion, adversarial review, independent review, or an improve-and-re-review loop before calling work done.

Install With Codex

Ask Codex to install the skill from this repo:

Install the second-opinion skill from GitHub repo kaskilling/second-opinion at path second-opinion.

Or run the Codex skill installer helper directly:

python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py \
  --repo kaskilling/second-opinion \
  --path second-opinion

Restart Codex after installing so the new skill is loaded.

Install With skills.sh

If you use the cross-runtime skills CLI:

npx skills@latest add kaskilling/second-opinion --skill second-opinion

Run this from the repo or workspace where you want the skill installed.

If your setup installs skills by copying folders, copy the second-opinion/ folder into your skills directory.

Use

Ask naturally:

Use second-opinion on this change.

Spawn a reviewer to be critical of this work, then fix the findings and ask for
a re-review.

For review-only work:

Use second-opinion to review these docs. Do not edit files; just give findings
and recommended fixes.

The skill is generic. It can review code, docs, setup instructions, plans, repositories, workflows, and generated artifacts.

Expected Loop

A typical run looks like this:

User: Use second-opinion on this setup guide. Fix the accepted findings.

Agent:
1. captures the goal, changed files, and current proof
2. asks a fresh reviewer to critique the guide
3. checks setup commands while the reviewer works
4. classifies findings as Blocker, High, Medium, Low, or Reject
5. fixes the accepted Blocker/High findings
6. runs the narrowest proof command
7. asks the reviewer to re-review the updated state
8. reports what changed, what passed, and what risk remains

Layout

second-opinion/
├── README.md
├── LICENSE
└── second-opinion/
    ├── SKILL.md
    └── agents/openai.yaml