self-review

Self-Review

Review the current branch's diff against its base branch using a model with zero context of the current conversation. Simulates a senior engineer MR review before you submit.

Input

Optional focus area in quotes.

Examples: /self-review, /self-review "focus on error handling"

Phase 1: Validate Context

Verify we're inside a git repo:

git rev-parse --is-inside-work-tree 2>/dev/null && echo "git repo" || echo "not a git repo"

If not a git repo, STOP: "This skill requires a git repository. Run it from inside a worktree."

Get the current branch:

git branch --show-current

If on main or master, STOP: "You're on the main branch. Switch to a feature branch first."

Auto-detect the base branch:

git rev-parse --abbrev-ref @{upstream} 2>/dev/null

If upstream is set, extract the remote branch name (e.g., origin/main → use origin/main). If no upstream, fall back to origin/main.

Store as [BASE].

Verify there are commits to review:

git log [BASE]..HEAD --oneline

If empty, STOP: "No commits found between [BASE] and HEAD. Nothing to review."

Tell the user what you're reviewing:

Reviewing branch [BRANCH] against [BASE]

Phase 2: Generate Diff & Stats

Capture the diff stats:

git diff [BASE]...HEAD --stat

Capture commit messages:

git log [BASE]..HEAD --oneline

Capture the full diff:

git diff [BASE]...HEAD

Count diff lines. If over 3000 lines, warn:

This diff is large (~[N] lines). The review may miss details — consider splitting your MR.

Proceed regardless.

Store the diff as [DIFF], stats as [STATS], and commit log as [COMMITS].

Phase 3: Build Review Prompt

Start with the base framing:

You are a senior engineer reviewing a merge request. You have no prior context about this project beyond the diff and commit messages below. Your job is to find what the author missed — not to praise what they got right. Be specific: quote the code, state the issue, explain why it matters.

Append the review instructions:

Review this merge request for:

Correctness & Logic

• Bugs, logic errors, off-by-one mistakes
• Unhandled error paths and failure modes
• Race conditions or concurrency issues
• Security vulnerabilities (injection, auth bypass, data leaks)

Code Quality

• Unclear intent — code a new reader would misunderstand
• Missing validation at system boundaries
• DRY violations — duplicated logic that should be extracted
• YAGNI violations — over-engineering, unnecessary abstractions, speculative features
• Unnecessary comments that restate the code

Change Hygiene

• Commit messages: are they clear and well-scoped?
• Is the change appropriately scoped, or does it mix unrelated concerns?
• Are there test coverage gaps for new/changed behavior?
• Are there leftover debug statements, TODOs, or commented-out code?

Structure your review as:

1. Summary — one paragraph on what this MR does and your overall assessment
2. Critical issues — must fix before merge
3. Suggestions — would improve but not blocking
4. Nits — minor style/preference items

If the user provided a focus area, append:

Additionally, pay special attention to: [FOCUS_AREA]

Append the commit log and diff:

Commits

[COMMITS]

Diff

[DIFF]

Store the full prompt as [PROMPT].

Phase 4: Run Review

cursor agent --model gpt-5.3-codex -p "[PROMPT]"

Capture the full output.

If the command fails or times out, report the error and suggest the user check that cursor CLI is installed and accessible.

Phase 5: Present Results

Present the output under a clear header:

## Self-Review: [BRANCH] → [BASE]
**Files changed**: [FILE_COUNT] | **Lines**: [INSERTIONS]+/[DELETIONS]- | **Model**: gpt-5.3-codex

---

[cursor agent output]

---

*Review by cursor agent (gpt-5.3-codex) — independent perspective, no shared context.*

Error Handling

• Not a git repo: Stop with clear message suggesting to run from a worktree
• On main branch: Stop, explain to switch to a feature branch
• No commits: Stop, explain nothing to review
• cursor command not found: Tell the user to install Cursor CLI or check PATH
• Timeout: Suggest a smaller diff or splitting the MR
• Empty output: Retry once, then report the issue