Independent Review
- N identical Subagents do the same Task independently.
- The value is coverage through redundancy: N independent passes surface findings one pass misses.
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.
Write the Prompt with /delegate. What this Skill adds to that Template:
Goal: perform this analysis independently. Document every finding with Evidence —
file paths, line numbers, concrete examples. State conclusions directly.
Verification:
- Every finding includes Evidence, not just assertions.
- Conclusions are stated directly, not hedged.
- Output is structured with clear sections.
Architecture: the Task scope. For a Review of code changes, use Current Changes and
Full Diff above; mark files to inspect with `*`.
Keep Subagent inputs identical
Use the same Agent, 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 /delegate to dispatch every reviewer in one message. Each returns its report on its own; collect all of them before synthesizing.
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. Triage the returns with /triage
Collect every return, then run /triage over the combined findings. Consensus proves nothing; the assessment against the code does. A finding only one Subagent caught gets the same assessment as one all of them caught, and a disagreement between Subagents is settled by reading the code, never by counting votes.