| name | lite-review |
| description | Lightweight code review of a GitHub PR or the local working diff. Runs a single sub-agent covering 4 fixed perspectives (CLAUDE.md compliance, bugs/correctness, code comment quality, security), scores findings 0-100, reports only score >= 50, and for the user's own PRs auto-fixes, commits, pushes, and updates the PR body. |
| argument-hint | [PR number or URL | omit to review the local working diff] |
| disable-model-invocation | false |
| effort | medium |
lite-review skill
Lightweight code review pipeline for changes too small to warrant the full
deep-review sweep. Reuses deep-review's scoring, autofix, and reporting
logic verbatim; only Step 4 (perspective reviews) differs.
Mode detection
Same as deep-review:
- PR mode (argument provided): extract PR number or URL, use
gh pr diff / view to get the diff. Eligible for autofix.
- Local diff mode (no argument): compute base with
git merge-base origin/<current-branch> HEAD and diff with git diff <base>..HEAD plus working-tree changes. Report only — no autofix or commit.
Steps
Execute the following steps strictly in order. lite-review's Steps 1-3
are identical to deep-review's Steps 1-3; lite-review's Steps 5-12
correspond to deep-review's Steps 6-13 (see
~/.claude/skills/deep-review/SKILL.md for their exact text) — only
lite-review's own Step 4 below replaces deep-review's Steps 4-5
(project-specific reviewer discovery and per-reviewer parallel review).
Step 1: Eligibility check (PR mode only)
Same as deep-review Step 1.
Step 2: Collect CLAUDE.md / rules content
Same as deep-review Step 2.
Step 3: Summarise changes
Same as deep-review Step 3.
Step 4: Single-agent perspective review (4 fixed perspectives)
Unlike deep-review's Step 5 (one sub-agent per reviewer file), launch a
single general-purpose sub-agent covering all 4 perspectives below in
one call:
- Read
~/.claude/skills/deep-review/reviewers/a-claude-md-compliance.md, b-bugs-correctness.md, e-code-comment-quality.md, and f-security.md directly (reuse their
## Scope bodies verbatim — do not duplicate the text into this file).
- Pass the sub-agent: the full diff, the change summary from Step 3, the
CLAUDE.md/rules content from Step 2 (with the "already read, don't
re-Read" instruction), the shared false-positive suppression
instructions (identical text to
deep-review SKILL.md's Step 5
suppression block), and all 4 reviewers' ## Scope bodies together,
instructing it to check the diff against all 4 perspectives in a
single pass and return findings tagged with which perspective(s) they
belong to.
Each agent returns findings as: problem summary + evidence + file:line reference.
Step 5: Confidence scoring (batched)
Same as deep-review Step 6: one Haiku sub-agent call for all findings,
same rubric, same Finding <N>: Score: <0-100> output format.
Step 6: Score filtering
Same as deep-review Step 7.
Step 7: Re-check eligibility (PR mode only)
Same as deep-review Step 8.
Step 8: PR author check (PR mode only)
Same as deep-review Step 9.
Step 9: Autofix (own PRs only)
Same as deep-review Step 10.
Step 10: Commit (own PRs only)
Same as deep-review Step 11.
Step 11: Update PR body (own PRs only)
Same as deep-review Step 12.
Step 12: Report results
Same as deep-review Step 13, with every occurrence of the report heading
changed from ### Deep Review to ### Lite Review (this applies across
all of Step 13's report variants — autofixed-own-PR, non-own-PR, and
no-issues-found — not just one) and the "no issues found" line changed to:
### Lite Review
No issues found. Checked for CLAUDE.md compliance, bugs/correctness, code comment quality, and security.
Formatting rules are identical to deep-review Step 13 (full SHA +
#L<line> links, no emoji, Score: <number> per finding, cite both code
and CLAUDE.md rule).