- Home
- Skills
- Code Quality & Review
- Quill Review: Verified Code Review on Your Local Model
More screenshots
Works with the AI Agents you already use
Quill Review: Verified Code Review on Your Local Model
Your own local model reviews the changes first, then your agent checks every finding against the code, so you only see real problems, ranked by severity.
$5
Quill Review: Verified Code Review on Your Local Model
Example session with this skill installed
Review my uncommitted changes.
- Read your context and instructions
- Compiled the quill review
Code review: uncommitted changes (4 files, +23 -7)
High
src/auth.js:6: isAdmin assigns instead of comparing. return user.role = 'admin'; overwrites the user's role and always returns a truthy value, so every caller passes the admin check and the stored role is corrupted. Fix: return user.role === 'admin';
src/auth.js:10: requireAdmin does not await getUser. getUser is async, so user is a Promise. The !user guard never fires and isAdmin receives the Promise, not the user. Fix: const user = await getUser(db, id);
src/files.js:7: the path traversal check was removed fromreadUpload.path.joindoes not stop../segments, so a name like../../etc/passwdreads files outsideuploads. Fix: restore the check that resolves the path and rejects it unless it starts withuploads + path.sep.src/list.js:3:paginatedrops the last item of every page.end = start + size - 1makesslicereturnsize - 1items, sopaginate([1, 2, 3, 4, 5], 0, 2)returns[1]. The existing test attest/list.test.js:6fails on this. Fix:const end = start + size;
Low
New functions have no tests: addDays (src/dates.js), pageCount (src/list.js) and requireAdmin (src/auth.js). I found this during verification; the local model did not report it. Fix: add tests for normal, boundary and error cases, starting with the admin check.
How this was checked: first pass by qwen3.5-35b-a3b-vl through Quill Code (full focus), the selected model, run without --model. I verified every finding against the code: 4 of 4 reported were confirmed, 0 rejected. I corrected the line numbers (for example, the paginate bug is on line 3, not 1 to 4) and raised the paginate finding from medium to high because it breaks normal pagination. I read the full diff and the existing test, and ran node --test test/list.test.js, which fails with the [1] result above. The working tree status was the same before and after the review, and no code was changed.
I can fix the four confirmed findings if you want. The files.js fix is the one that most needs your review, since it changes how names are resolved.
Connects securely to your tools. The creator never sees your data.
About this skill
Ask your agent for a code review as usual. Instead of spending cloud tokens reading every line, this skill has the model running on your own machine do the first pass through Quill Code (free, MIT, https://github.com/Smoffyy/quill-code). Your agent then checks every finding against the actual code before you see it: real problems are kept and ranked, false positives are rejected and counted, wrong line numbers and severities are corrected.
What you get
- Any scope: uncommitted changes, staged changes, the last commit, a commit range, a branch or pull request against its base, or chosen files and folders.
- Focus modes: full review, security, performance, tests, or a focus you name.
- Verified findings: nothing reaches you that your agent has not checked in the code. Problems the local model missed but your agent spots while checking are added and marked.
- Severity you can act on: High, Medium and Low by a fixed rubric, with corrected locations and a suggested fix for each.
- Thorough mode: a second pass with another local model; findings both models report get high confidence.
- Large changes handled: big diffs are split into groups automatically so small models stay accurate.
- Three output formats: a review report, a pull request comment ready to paste, or JSON for CI.
- Fix mode: on request, fixes only the confirmed findings, runs your tests, reviews the fix once more and never commits.
- Read-only by design: the review runs with Quill's ask-first access, which declines every edit and command except read-only git, and your agent checks that nothing changed.
- Follows your project's rules: AGENTS.md, CONTRIBUTING and linter settings are passed to the review.
How it is tested
Every release passes 8 automated scenarios run end to end with Claude Code and Qwen 3.5 35B-A3B on llama.cpp: finding four planted bugs (an off-by-one, an assignment in a condition, a missing await and a path traversal) without accusing correct code next to them; a clean change with no invented problems; a branch review with a security focus; review and fix with the tests passing; valid JSON output; a 20-file change split into groups; thorough mode with two models; and a machine without Quill, where it explains how to install it instead of faking a review. All 8 also pass on Qwen 3.5 9B. On Qwen 3.5 4B all 8 passed too, and verification caught bugs the 4B missed and rejected its false alarms, but on the 20-file change its first pass missed the planted bug in one of two runs. For large changes, use a 9B model or larger, or thorough mode.
Requirements
- Quill Code 1.0.1 or newer
- git
- A local model server (llama.cpp, LM Studio, Ollama, vLLM, Jan, GPT4All or any OpenAI-compatible server) running a model with tool calling. 9B or larger is recommended; 4B works, with more for your agent to verify.
Install Quill Code
macOS and Linux:
curl -fsSL https://github.com/Smoffyy/quill-code/releases/latest/download/install.sh | sh
Windows (PowerShell):
irm https://github.com/Smoffyy/quill-code/releases/latest/download/install.ps1 | iex
Then open a new terminal in a project and run quill once to pick your model.
What's inside
SKILL.md: the review workflowreferences/brief.md: the brief the local model gets, with the checks for each focusreferences/report.md: the severity rubric and the report, pull request comment and JSON formatsreferences/large-changes.md: how big changes are splitreferences/fixing.md: fix mode
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Code Quality & Review.
Needs Quill Code, git and a running local model server. Review quality depends on your local model: your agent's verification removes false positives and often catches what the model missed, but not always, especially on large changes with small models (under 9B). A review takes about one to four minutes on a local model, longer in thorough mode.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 27 days ago
- Software
- Quill Code (quill CLI)
Git
llama.cpp, LM Studio, OllamaorvLLM
Permissions
The agent runs quill and read-only git commands; the review itself never edits files, and in fix mode only the confirmed findings are changed. Quill talks only to the local model server you configured, usually on localhost. The skill makes no network calls of its own and reads no credentials.