Agent Skills: Independent Review

Run N identical parallel Subagents on the same Task for independent Verification. Use when the Architect wants consensus through redundancy — multiple Subagents independently analyze, review, or test the same thing, then results are compared for agreement, disagreements, and unique finds. Invocation /independent-review [count] "task"

UncategorizedID: heyJordanParker/dotfiles/independent-review

Install this agent skill to your local

pnpm dlx add-skill https://github.com/heyJordanParker/dotfiles/tree/HEAD/packages/agents/skills/independent-review

Skill Files

Browse the full folder contents for independent-review.

Download Skill

Loading file tree…

packages/agents/skills/independent-review/SKILL.md

Skill Metadata

Name
independent-review
Description
Run N identical parallel Subagents on the same Task for independent Verification. Use when the Architect wants consensus through redundancy — multiple Subagents independently analyze, review, or test the same thing, then results are compared for agreement, disagreements, and unique finds. Invocation /independent-review [count] "task"

Independent Review

  • N identical Subagents do the same Task independently.
  • The value is consensus through redundancy: agreement, disagreements, and unique finds.

1. Capture current changes for Review Tasks

Current Changes:

!git changes

Full Diff:

!git diff HEAD

2. Parse the Architect's input

The first numeric argument is the Subagent count. Default to three Subagents when no count is provided. Everything else is the Task.

Example: /independent-review "is this migration safe?" runs three Subagents.

Example: /independent-review 5 "review for security vulnerabilities" runs five Subagents.

3. Build one Prompt

Every Subagent receives the exact same Prompt with no variation.

Template: Story: {Task — what the Architect wants analyzed, reviewed, or tested, and why}

Business: {Rules — codebase Context, stack, and what matters}

Goal: Perform this analysis independently. Be thorough. Document every finding with Evidence: file paths, line numbers, and concrete examples. Do not hedge; state conclusions directly.

Verification:
- Every finding includes Evidence, not just assertions.
- Conclusions are stated directly, not hedged.
- Output is structured with clear sections.

Architecture:
{Task scope. For a Review of code changes, use Current Changes and Full Diff above; mark files to inspect with *.}

Process:
1. Read the Task scope and every file marked * in the Architecture block.
2. Complete the analysis, Review, or test independently.
3. For each Verification item, run Verification and paste the observed output.
4. If a Verification item fails, fix the work and re-verify.
5. Post a completion summary: what was checked, what was verified, and what was tricky.

Keep Subagent inputs identical

Use the same subagent_type, Prompt, and tools for every Subagent. Reasoning variation provides the natural diversity.

Never: assigning different lenses, files, or specialties to each Subagent.

4. Dispatch N Subagents in parallel

Use run_in_background: false so all results are collected. Name Subagents reviewer-1, reviewer-2, and onward.

Keep the Skill Task neutral

This Skill wraps any Task the Architect provides: code Review, bug analysis, Architecture assessment, test adequacy, or migration risk.

5. Synthesize the results

Compare all Subagent returns before reporting.

  • Consensus means findings that two or more Subagents independently identified. List each finding and which Subagents found it.
  • Unique finds means findings only one Subagent caught. These need Architect judgment because each could be an insight or a false positive.
  • Disagreements means Subagents contradict each other. Present both sides with Evidence.
  • Confidence means the N-of-N agreement ratio, such as "3/3 Subagents agree" or "2/5 Subagents found this".

Require Evidence for every finding

Every finding includes concrete Evidence. "Might be a problem" is not a finding.