From skill-set
Execute peer reviews from other LLM CLI tools (Gemini, Codex, Claude) in parallel and synthesize actionable insights. Use when user requests feedback from other LLMs, peer review, or external validation — e.g., "get feedback from gemini", "ask codex to review", "ask claude to review", "peer review this", "what do other LLMs think", "get a second opinion", "validate with codex".
How this skill is triggered — by the user, by Claude, or both
Slash command
/skill-set:consulting-peer-llmsThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Get feedback from other LLM CLI tools (Gemini, Codex, Claude) on your current work. This skill executes multiple LLM reviews in parallel and synthesizes their responses into one actionable report.
Get feedback from other LLM CLI tools (Gemini, Codex, Claude) on your current work. This skill executes multiple LLM reviews in parallel and synthesizes their responses into one actionable report.
Core principle: Use peer LLMs for external validation and diverse perspectives on implementation quality.
Use this skill when the user requests external LLM review:
This skill runs CLI tools in parallel, which takes 5-30 minutes. Only trigger on explicit user request, not as a routine step.
Supported CLI tools:
gemini - Google Gemini CLIcodex - OpenAI Codex CLIclaude - Anthropic Claude Code CLIDetection:
command -vPrompt minimalism principle: CLIs run in the same repository. They can git diff, git log, and read any file. Never duplicate what they can discover themselves.
Bare prompt (no context available — e.g., slash command without arguments):
Review all changes on the current branch vs origin/main.
Use git diff origin/main...HEAD and read files directly.
With conversation context (agent knows what was implemented):
Review all changes on the current branch vs origin/main.
Use git diff origin/main...HEAD and read files directly.
{1-2 sentence summary of what was implemented and why}
With explicit review focus (user specifies files or areas):
Review all changes on the current branch vs origin/main.
Use git diff origin/main...HEAD and read files directly.
{1-2 sentence summary, if available}
Focus on: {user's specific requirements — paths or areas only if user explicitly asked}
What goes in the prompt:
Keep out of the prompt (CLIs run in the same repo and discover all of this themselves):
Avoid running git commands to gather context for the prompt. If there is no conversation context, use Tier 1 (bare prompt) as-is rather than fabricating context. Shorter prompts produce more focused reviews.
Full template: See reference/prompt-template.md
Always use the bundled script rather than calling gemini, codex, or claude directly:
bash "$SKILL_DIR/scripts/peer-review.sh" execute "$PROMPT"
Run in background. The script handles CLI detection, correct flags per CLI, parallel execution, and timeout.
CLI flag semantics are unintuitive and differ between tools — for example, codex -p means --profile (not prompt), and codex without exec enters interactive mode. These have caused repeated failures when invoked directly. The script encapsulates the correct invocations. Direct CLI tool permissions are intentionally excluded from allowed-tools to prevent bypassing it.
Output contract: The script writes each CLI's full response to $PEER_REVIEW_DIR/<cli>.txt (default /tmp/peer-review-$$/) and prints only a bounded status block to stdout — never response bodies. Do not pipe the script through head/tail/grep; truncating stdout cannot save tokens here, it only loses status lines. Stdout looks like:
PEER_REVIEW_DIR=/tmp/peer-review-12345
Responses (read each file individually — do NOT pipe through head/tail/grep):
gemini OK /tmp/peer-review-12345/gemini.txt (5421 lines)
codex OK /tmp/peer-review-12345/codex.txt (3892 lines)
claude EMPTY /tmp/peer-review-12345/claude.txt (0 lines, stderr: ...)
Details: See reference/cli-commands.md
After the script finishes, read each per-CLI file directly with the Read tool — one Read call per file. Then present them to the user in sequence for transparency:
# Gemini Review
{contents of $PEER_REVIEW_DIR/gemini.txt}
---
# Codex Review
{contents of $PEER_REVIEW_DIR/codex.txt}
---
# Claude Review
{contents of $PEER_REVIEW_DIR/claude.txt}
---
If a CLI's status was EMPTY or FAILED, also surface its stderr file (<cli>.txt.err) so the user can diagnose auth or network issues.
Always synthesize - even for single CLI responses.
Synthesis principles:
Report example: See reference/report-format.md
Apply the autofixing-and-escalating skill to the synthesized report items. The synthesized items are the external source input — classify each as OBVIOUS or AMBIGUOUS, auto-fix obvious ones, and escalate ambiguous ones for user decision.
Integration point: The synthesized report from Step 4 replaces the raw peer responses as the authoritative item list. Do not re-classify raw CLI output — only the deduplicated, validated synthesis.
After synthesis, rm -rf "$PEER_REVIEW_DIR". The script does not auto-delete; explicit overrides (e.g. .context/peer-review-...) accumulate until cleaned.
Commands:
/skill-set:consulting:review <requirements> - Auto-detect all installed CLIs and review with the given requirementsBundled script: scripts/peer-review.sh — Handles CLI detection, parallel execution with timeout, and result collection. Bash 3.2+ compatible (macOS/Linux).
scripts/peer-review.sh check — Show installed CLIs and timeout availabilityscripts/peer-review.sh execute "prompt" [cli1 cli2] — Run review with specified or all available CLIsTypical execution time: 5-30 minutes (parallel)
Output files: $PEER_REVIEW_DIR/<cli>.txt (response) and <cli>.txt.err (stderr). Default $PEER_REVIEW_DIR is /tmp/peer-review-$$/. Files persist after the script exits — Read them with the Read tool, then optionally clean up.
gemini, codex, or claude directly instead of using the bundled scripthead, tail, grep, awk, or any line-cap — response bodies are in files, not stdout. Truncation here only loses status lines and signals you're treating the output incorrectly.--full-auto, -q, --model, etc.)Some CLIs fail: Continue with successful ones, note failures in report
Timeout (exit 124): Reduce prompt size, check CLI responsiveness
No retries: Keep execution fast and simple
"codex failed", "unexpected argument", or "profile not found"
codex directly instead of using the script. Use bash "$SKILL_DIR/scripts/peer-review.sh" execute "$PROMPT"codex, gemini, and claude calls are intentionally excluded from allowed-tools — if the user is prompted for Bash approval, you are calling the CLI directly instead of using the scriptcodex -q, codex -a full-auto, codex -p — these are not valid one-shot invocationscodex exec, codex review — these work, but calling them directly bypasses timeout and parallel execution. Use the script."Empty response from CLI"
gemini -p "test", codex exec -o /tmp/test.txt "test", or claude -p "test""All CLIs failed"
gemini --version && codex --version && claude --version"Response is truncated"
$PEER_REVIEW_DIR/<cli>.txt directly (not bash stdout, which is the status block); if the file itself is short, reduce prompt length and retry.npx claudepluginhub ether-moon/skill-set --plugin skill-setCreates bite-sized, testable implementation plans from specs or requirements, with file structure and task decomposition. Activates before coding multi-step tasks.