preview

name	preview
description	AI screening review for PRs - preliminary check before human review. Do NOT use for deep multi-reviewer code quality audits (use /audit instead).
when_to_use	スクリーニング, PRレビュー, プレビュー, preview PR, pre-review
allowed-tools	Bash(git:*) Bash(gh:*) Read AskUserQuestion Bash(ugrep:*) Bash(bfs:*)
model	opus
argument-hint	[PR URL or number]

/preview - PR Screening Review

Input

$ARGUMENTS is a URL, a number, or empty. If empty, detect from the current branch.

Execution

1. Identify PR: gh pr view $ARGUMENTS --json number,title,body,labels,files,url. On failure, retry without $ARGUMENTS
2. Abort if no PR found or working tree is dirty. Check via git status --porcelain
3. Checkout PR: gh pr checkout $PR
4. Gather PR context in parallel (§ PR context gathering)
5. Read each changed file in full, including code outside the diff hunks
6. Review per process: overview → per-file → dependency impact → findings
7. Output structured screening report

PR Context Gathering

Never include author in gh output fields.

# Diff
gh pr diff $PR

# Existing comments
gh pr view --comments $PR

# Inline comments
gh api repos/{owner}/{repo}/pulls/{number}/comments \
  --jq '.[] | {file: .path, user: .user.login, comment: .body}'

Comment Labels

Label	Meaning	Severity
[must]	Requires fix before merge	High
[want]	Should fix, not blocking	Medium
[imo]	Personal opinion, take or leave	Low
[ask]	Question needing clarification	-
[nits]	Minor style/formatting issue	Low
[info]	Context sharing, no action	-

Comment Tone

Rule	Detail
Format	[label] <observed behavior or risk>. <suggestion>. (file:line)
Concise	3 lines for [imo]/[nits]/[info]; up to 5 for [must]/[want] with evidence
Respectful	Acknowledge effort, avoid commands
Suggestive	"Consider..." not "This is wrong"
Author-targeted	Comments may be posted verbatim - calibrate detail for the PR author

Output

Generate the report from the ${CLAUDE_SKILL_DIR}/templates/screening-report.md skeleton and emit it to the conversation.

Rules

Rule	Detail
No auto-post	Never post comments to PR automatically
Abort on dirty	If uncommitted changes exist, warn and abort
Speed over depth	This is screening, not full audit
Verify before flag	Before [ask]/[want]+, trace the issue to a reachable runtime call site

References

Topic	File
Review Checklist	${CLAUDE_SKILL_DIR}/references/review-checklist.md