SuperAPI AI Directory

code-review-multi-angle

Coding & development

View repository

Updated 11-11 22:14

About this skill

code review multi angle

Review the specified diff, PR, selected range, file, or current Git changes in read-only mode.

When to Use

  • Use: when you want to review code changes before merge, before commit, or before a PR
  • Use: /review -style defect review and /security-review -style security review at the same time
  • Use: when asked to "review from multiple angles," "review strictly," or "check for any major oversights"
  • Use: when fixes after review are explicitly requested, such as "review and fix," "fix," or "correct it"
  • Don't use: for structural review of writing style, documentation, articles, or workflow customization. In those cases, use the dedicated prompt / agent / skill
  • Don't use: for full-scale pre-release QA, iterative improvement of the entire product, release / publish / version sync. For those, use refine-product-100 Use

Mode

  • Default is review-only: do not edit code, only output findings and fix policy
  • fix / 修正して / 直して only when review-and-fixis explicitly specified: perform review, critic review, fixes, and focused validation
  • For destructive operations, external publishing, history rewriting, push, and broad refactors, confirm before executing even in fix mode

Fix mode classification and scope

In fix mode, classify the problems found in a single review as follows.

  • Fix now: only those within the requested scope, non-destructive, locally verifiable, and whose correctness can be explained. Fix with the minimal diff
  • Guard now: those where the main fix is difficult but the risk can be reduced through at least one of tests, static checks, docs, or verification procedures. Produce that artifact
  • Block: those involving destructive operations, external exposure, specification decisions, secrets, production data, or insufficient permissions. Report without executing
  • Do not leave Fix now that cannot be safely edited as review-only; make themGuard now or Block Set to
  • Sweep only horizontally expands to problems caused by the same root as the fixed perspective (search by string, function name, type, error message, config key). Improvements that are merely conceptually similar or quality improvements on a different axis go to Next Steps
  • If validation fails, read the cause and perform fixes and re-validation for up to 2 cycles per same cause. Anything beyond that is left as Block Keep as
  • Only present learnings as candidates for Learnings Candidate . Reflect them into instructions or review-learnings files only when the user explicitly says so

Scope Detection

Determine the review target in order of priority.

  1. PR URL, comparison branch, file, selected range, or pasted diff specified by the user
  2. staged changes: git --no-pager diff --staged
  3. unstaged changes: git --no-pager diff
  4. If the working tree is clean: git --no-pager diff main...HEAD

main In repositories where a base branch other than the default is natural, check the branch name and upstream before using it.

Review Lanes

Examine the same target from at least 3 independent perspectives.

1. Correctness and Regression

  • Whether specifications, existing behavior, or boundary conditions have been broken
  • Whether it breaks with null / empty / missing / duplicate / out-of-order / concurrent input
  • Whether compatibility of public APIs, schemas, CLI contracts, or output formats has been broken
  • Whether insufficient error handling causes crashes, data loss, or partial updates

2. Security and Trust Boundary

  • Whether there is injection, path traversal, SSRF, unsafe deserialization, or prototype pollution
  • Whether authentication, authorization, permission boundaries, or tenant / repo / workspace boundaries are crossed
  • Whether secrets, tokens, PII, internal URLs, or stack traces are leaked into logs or output
  • Whether untrusted input is mixed into LLM prompts, tool routing, shell commands, URLs, or file paths

3. Reliability, Operations, and Maintainability

  • Whether retry, timeout, idempotency, cleanup, rollback, or resume is missing where needed
  • Whether performance impact has measurably worsened
  • Whether false positives from build artifacts, generated files, or large diffs are avoided
  • Whether it deviates from existing patterns, nearby implementations, or testing policies

Noise Filter

Only report findings that the reader actually has value in fixing. Classify them as Critical / High / Medium / Low, and report Low as well if there is rationale and fixing value.

What not to report.

  • Style, format, naming, preferred design
  • Minor improvements to grammar, spelling, or comment text
  • "If possible" or "generally speaking" level best practices
  • Refactors that do not affect the requested outcome even if fixed

Do not report findings you are unsure about. Instead, briefly note them in Assumptions / Not Reviewed .

Critic Review

For non-trivial changes, security findings, multi-file changes, or fix mode, double-check the review results through an available rubber duck / critic path.

  • If available, use native Rubber Duck,duck-critic skill, a read-only subagent, or a different-model critic
  • Keep the critic read-only and do not ask it to make fixes or state changes
  • Deduplicate the critic's points and reflect only those actually adopted in findings or fixes
  • If the critic path is unavailable, perform a second-pass review yourself and write the reason it was unavailable in Verification write in

Verification

To the extent possible, perform read-only or low-risk verification before reporting a finding.

  • Check surrounding code, existing tests, call sites, schemas, README, and configuration
  • Run compile / typecheck / lint / focused test as needed
  • If unable to run, clearly state the reason and alternative verification
  • In review-only, do not edit code. Do not create, modify, or delete files
  • In fix mode, perform minimal fixes after critic review, then run focused validation

Output Format

Present findings first, in order of severity.

## Review Findings

### 1. {短いタイトル}
- Severity: Critical | High | Medium | Low
- Lane: Correctness | Security | Reliability
- File: {path}:{line}
- Problem: {実際に起きる問題}
- Evidence: {差分、周辺コード、テスト、仕様から確認した根拠}
- Suggested Fix: {実装方針。コードは変更しない}

## No Significant Issues

{finding がない場合だけ、重大な問題が見つからなかった理由を 1 行で書く}

## Assumptions / Not Reviewed

- {未確認範囲、実行できなかった検証、前提。なければ None}

## Fixes Applied

- {fix mode の場合だけ、変更したファイルと理由。分類(Fix now / Guard now / Block)と Sweep の結果を含める。review-only では None}

## Learnings Candidate

- {永続化したい学び候補。なければ None}

## Verification

- {実行した確認、または未実行理由}

Do not include praise, long work logs, or exhaustive lists of files viewed.

Completion Criteria

  • Identified the target diff or specified scope
  • Checked from 3 or more perspectives
  • Limited findings to actionable ones only and classified them as Critical / High / Medium / Low
  • Performed critic review for non-trivial reviews or fix mode, or explicitly stated an alternative second pass
  • Verified evidence or performed focused validation to the extent possible
  • Did not change code in review-only
  • In fix mode, performed focused validation after minimal fixes