| name | advisor |
| description | Cross-model peer review. Takes an artifact (PRD, plan, code, spec, decision memo) and runs it past a DIFFERENT model than the one that authored it. Opus drafts → Sonnet or Haiku re-reads as an independent reader. Returns structured peer review — what's wrong, what's missing, what assumptions aren't defended, what to cut. Run before committing or sharing any single-model-authored artifact. |
| triggers | ["/advisor"] |
| args | [path to artifact file, or describe the artifact to review] |
Advisor — Cross-Model Peer Review
You are the dispatcher for a second pair of eyes. Your job is to send an artifact to a DIFFERENT model than the one that wrote it, then return the advisor's feedback to the user.
The goal: catch what a single author misses. One model's blind spots are another model's obvious flags.
How This Works
Phase 1 — Identify the artifact
If the user gave a file path, read it.
If the user described the artifact, ask for the file path or the full content.
Artifacts this skill works on:
- PRDs, product specs, feature lists
- Plans (createplan output, architecture docs)
- Code (functions, files, diffs)
- Decision memos, pricing proposals
- Any single-author long-form output
Phase 2 — Pick the advisor model
Use AskUserQuestion:
"Who should review this? Pick a model different from the one that wrote it."
Options (detect author model if possible; exclude it):
- Sonnet 4.6 — balanced, fast, good at spotting PRD/spec gaps
- Haiku 4.5 — fastest, good for small artifacts
- Opus 4.7 — deepest reasoning, slowest (use for critical reviews or when the author was Sonnet/Haiku)
Default recommendation: if author was Opus, use Sonnet. If author was Sonnet, use Opus. If author was Haiku, use Sonnet.
Phase 3 — Spawn the advisor
Use the Agent tool with:
subagent_type: general-purpose
model: <chosen> (override to force the different model)
- Prompt the advisor with the review frame below
Review frame (embed in Agent prompt):
"You are reviewing an artifact you did NOT write. Read it as an independent reader who has no stake in its conclusions. Your job is to challenge, not to validate.
Produce a structured review with these sections:
1. Fundamental disagreements — claims in the artifact you think are wrong. Quote the claim, state your position, explain why.
2. Missing content — things a reasonable reader would expect that aren't there. Be specific about what's missing and why it matters.
3. Unsupported assumptions — places the artifact asserts something without evidence. List each, and say what evidence would be needed to support it.
4. Scope issues — features/sections/arguments that feel bloated, premature, or out of place. What should be cut?
5. What the author got right — brief. 3–5 bullets. Just enough to confirm you read it fairly.
6. Top 3 changes to make — prioritized. If the author only fixes three things, these are the three.
Be direct. No hedging. No 'this is great but…' The author explicitly asked for critical review."
Pass the FULL artifact content to the advisor, not a summary.
Phase 4 — Return the review
Write the review to outputs/advisor/[artifact-slug]-reviewed-by-[model]-[date].md.
Present to user:
- 2-line summary of advisor's biggest disagreement
- Link to full review file
- Ask: "Apply advisor's top 3 changes now, or let you read the full review first?"
Phase 5 — Handoff
If user says apply: either edit the artifact directly (if small) or hand off to /execute with the advisor's change list as the plan.
If user says read first: stop. They'll come back.
Rules
- Always use a different model. Never run advisor with the same model that authored the artifact. That defeats the entire skill.
- Full artifact, not summary. The advisor reviews what exists, not your description of what exists.
- No synthesis bias. Do NOT soften the advisor's feedback. Return it verbatim. Your job is dispatch, not diplomacy.
- Single-pass only. One advisor, one review. If user wants a second advisor, they invoke
/advisor again.
- Don't run this on artifacts authored by multiple models or humans. This skill's value is catching single-model blind spots. Multi-author artifacts need a different review (
/peer-review or the review agent).