cursor
/cursor
Have Cursor do a sandboxed, read-only review of the current git repository's changes (for example, what Codex just did). Runs only when the user types /cursor.
- Author
- hdeng26 ↗
- Repository
- claude-cursor-review-skill ↗
- Licence
- MIT
- Last repository change
- 11.11.2026
Assessment and security
- Files in this package
- 4
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 11.11.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
Have Cursor do a sandboxed, read-only review of the current git repository's changes (for example, what Codex just did). Runs only when the user types /cursor.
Arguments
/cursor [--dir PATH] [--full] [--repo] [--task ID | --base REF] [review focus...]
Original author description
Have Cursor do a sandboxed, read-only review of the current git repository's changes (for example, what Codex just did). Runs only when the user types /cursor.
How it works
- 1. Parse the arguments
- 2. Decide what to review
- 3. Write the context brief
- 4. (--full only) Create the conversation excerpt
- 5. Run the review
- 6. Report
SKILL.md preview
---
name: cursor
description: Have Cursor do a sandboxed, read-only review of the current git repository's changes (for example, what Codex just did). Runs only when the user types /cursor.
disable-model-invocation: true
argument-hint: "[--dir PATH] [--full] [--repo] [--task ID | --base REF] [review focus...]"
allowed-tools: Bash(python3 ${CLAUDE_SKILL_DIR}/scripts/cursor_review.py *) Bash(python3 ${CLAUDE_SKILL_DIR}/scripts/transcript_excerpt.py *)
---
# /cursor: one external read-only review
User arguments: `$ARGUMENTS`
This command does one thing: run `cursor_review.py` once and report the result to the user faithfully.
Do not fix problems, retry, merge, call Codex, or do anything else along the way.
Talk to the user in the language they use with you.
The scripts are in `${CLAUDE_SKILL_DIR}/scripts/`.
## 1. Parse the arguments
- `--dir PATH`: directory of the repository or git worktree to review, for example another session's worktree. Written `<DIR>` below; defaults to the current directory
- `--full`: attach a condensed conversation transcript (only the text written by you and the user)
- `--repo`: attach a read-only snapshot of the whole repository (`--scope repo`), for changes that need context to understand
- `--task ID`: review the delegated task recorded in `runs/ID/` (needs a `runs/ID/meta.json` produced by a delegation script; see the README)
- `--base REF`: the starting point to compare against
- Any remaining text is the review focus; pass it verbatim to `--focus`
## 2. Decide what to review
Write every git command below as `git -C <DIR> ...`. First run `git -C <DIR> rev-parse --show-toplevel`.
If it fails, tell the user that `<DIR>` is not a git repository (`/cursor` must be used inside a git repository, or pointed at one with `--dir`), then stop.
Pick the first that applies:
1. `--task ID` given → `--task ID`
2. Run `git -C <DIR> status --porcelain -- . ':(exclude)runs'`. If it prints anything (there are uncommitted changes, for example from the Codex plugin) → `--base <REF or HEAD> --worktree` (here "worktree" means the working tree, not a git worktree)
3. Clean working tree → `--base <REF or HEAD~1>` (review the latest commit)
Say in your report which one you picked and why.
## 3. Write the context brief
Run `git -C <DIR> rev-parse --absolute-git-dir` to get `<GITDIR>`, and write the brief to `<GITDIR>/cursor-reviews/context-${CLAUDE_SESSION_ID}.md` (inside `.git`, so it never enters version control or the diff). Contents:
- What the user wants: quote the user's own words from the conversation where possible
- The concrete requirements given to the implementer (Codex or anyone else), and constraints agreed in the conversation: which files may change, what must not be touched, style requirements
- Risks you know of, or things the user is worried about
Only include what is relevant to this change; usually no more than a page. Never include keys, passwords or tokens.
Do not conclude here that the change is correct: this is background for the reviewer, not your review.
## 4. (`--full` only) Create the conversation excerpt
```bash
python3 ${CLAUDE_SKILL_DIR}/scripts/transcript_excerpt.py --session ${CLAUDE_SESSION_ID} -o <GITDIR>/cursor-reviews/transcript-${CLAUDE_SESSION_ID}.md
```
## 5. Run the review
```bash
python3 ${CLAUDE_SKILL_DIR}/scripts/cursor_review.py --dir <DIR> <arguments from step 2> [--scope repo] \
--context <GITDIR>/cursor-reviews/context-${CLAUDE_SESSION_ID}.md \
[--context <GITDIR>/cursor-reviews/transcript-${CLAUDE_SESSION_ID}.md] \
[--focus "<review focus>"] [--timeout 600]
```
- With `--repo` or `--full`, add `--timeout 600` and set the Bash timeout to 660000 ms
- Exit codes: 0 = PASS, 1 = FAIL or UNCLEAR, 3 = the review itself failed. Never rerun in any of these cases
- Never add `CURSOR_ISOLATION=none`. If the sandbox fails, report the failure and let the user decide
## 6. Report
Read the `review_cursor.md` named in the output, and `review_cursor.json` in the same directory (on failure, look at its `error` and `stderr`), then report in this format:
```
cursor review: PASS / FAIL / UNCLEAR / failed (<error>)
repository: <repository root for DIR> (for a git worktree, also give the branch)
range: <base>..<HEAD or working tree>, scope=<diff|repo>, context: <context files>
definite problems: each with file:line
suspicions: ...
record: <path to review_cursor.md>
```
You may finish with one or two sentences on whether you agree with Cursor's conclusion, marked as your own view.
Do not modify the review record.
Files in this package (4)
- .gitignore 1 KB
- LICENSE 1 KB
- README.md 7 KB
- SKILL.md 4 KB
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · hdeng26/claude-cursor-review-skill | 95 | MIT |
| G1Joshi/Agent-Skills | 100 | MIT |
| Dicklesworthstone/agent_flywheel_clawdbot_skills_and_integrations | 95 | MIT |
Database snapshot: 13.11.2026 · package version 9ab55921df83 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100cursor-git-integration
Manage integrate Git workflows with Cursor IDE. Triggers on "cursor git", "git in cursor", "cursor version control", "cursor commit", "cursor branch". Use when working with cursor git integration functionality. Trigger with phrases like "cursor git integration", "cursor integration", "cursor". - 100cursor-git-integration
Integrate Git workflows with Cursor IDE: AI commit messages, @Git context, diff review, and conflict resolution. Triggers on "cursor git", "git in cursor", "cursor version control", "cursor commit", "cursor branch", "@Git". - 100cursor-model-selection
Configure and select AI models in Cursor. Triggers on "cursor model", "cursor gpt", "cursor claude", "change cursor model", "cursor ai model". Use when working with cursor model selection functionality. Trigger with phrases like "cursor model selection", "cursor selection", "cursor". - 100cursor-local-dev-loop
Optimize daily development workflow with Cursor IDE using Chat, Composer, Tab, and Git integration. Triggers on "cursor workflow", "cursor development loop", "cursor productivity", "cursor daily workflow", "cursor dev flow". - 100cursor-cli
Run Cursor CLI for one-shot cross-AI code review when a Cursor perspective is explicitly wanted. Invoke ONLY when the user asks ('cursor review', 'ask cursor', 'run cursor'), or another skill explicitly delegates to Cursor. Do not invoke it implicitly for a generic review. Not… - 100cursor-cli
Run Cursor CLI for one-shot cross-AI code review when a Cursor perspective is explicitly wanted. Invoke ONLY when the user asks ('cursor review', 'ask cursor', 'run cursor'), or another skill explicitly delegates to Cursor. Do not invoke it implicitly for a generic review. Not…