Refactor Assistant
Open the app: Plan a refactoring → (products.documentize.ai)
In one sentence
Refactor Assistant scans a C# repository on GitHub for oversized classes that hide separate responsibilities, ranks them, and, when you choose, extracts the chosen parts into new classes and opens a pull request.
What makes it unique
It works in two separate steps with a human decision in between. Scan (called "suggest") is read-only and changes nothing. Apply then extracts only the "seams" you queued. Each class is shown with its seams (a seam is one hidden responsibility that could move out), a score, and a verdict, so you decide with evidence rather than letting an AI rewrite whatever it likes.
When to use it / when not to
- Use it on C# classes that have grown into "does everything" classes and you want a ranked shortlist of what to split.
- Not for documentation → Repo Docs. Not for a full structural map → Code Graph.
- Not for languages other than C#.
- Not as a build check. The page states plainly: Documentize never builds or runs your code, so your CI on the pull request is the build check.
Before you start
- A GitHub repository (
owner/nameor URL) with C# code. Private repository: givearise-projectaccess first, as the page hint says. - For Apply: you must be signed in, and the page says Apply needs a paid plan and takes 10 to 30 minutes per seam.
- Privacy: the code is read to analyse it; Apply writes a branch and pull request as
arise-project. Treat the repository as shared with that account and with the language-model service used for the analysis (verify the exact data flow on the live app).
Step by step
- Enter the GitHub repository and optionally a Branch.
- Choose how many Candidates to show (10, 25 or 50), or type one class name to inspect (optional).
- Run the scan. Read the tiles (classes scanned, candidates, seams found, already refactored, rejected, too large) and the list.
- Filter by project or verdict, open a class to see its outline, and open each seam: score, cohesion, call sites to delegate, what moves with it, what stays behind.
- Preview the plan: the contract the new class is written against and what moves. Nothing is generated yet.
- Add to apply queue the seams you want (at most 5 per run), pick One pull request, one commit per seam or One pull request per seam, and press Apply.
- Follow the steps (Plan, Contract, Generate, Fix loop, Build, Review, Commit) and open the pull request. Use your CI to build it.
Options explained
| Option | What it does | Default | Recommended |
|---|---|---|---|
| Branch | Branch to analyse and open the pull request against | default branch | Use a stable branch |
| Candidates to show | Size of the ranked list: 10, 25 or 50 | 10 | 10 first |
| Inspect one class | Scan one class only, showing every seam, settled ones included | empty | When you already suspect a class |
| Verdict (filter) | Worth splitting, Infrastructure, Too coupled, Settled, Not judged | all | Start with "Worth splitting" |
| Project (filter) | Restrict the list to one project | all | Use in large solutions |
| Mode (Apply) | One pull request with one commit per seam, or one pull request per seam on its own branch | one PR | One PR per seam if your team reviews small |
What you get
- Scan: an on-page report, saved as a run you can reopen. There is also a History of everything decided about the repository (rejected, proposed, extracted), visible when signed in.
- Apply: a pull request with the extracted classes, with a summary of files and checks. Outcomes per seam: In the pull request, Not worth a review, Failed, Skipped, Not started, Timed out. If nothing was kept, no pull request is opened.
Quality & limits
- Scan is analysis only; Apply is AI-generated code, not compiled by Documentize. Review every pull request and let CI run before merging.
- Scan has a time limit; a repository too large to scan in time is refused with a message. Apply takes several minutes per seam.
- At most 5 seams per Apply run.
Works well with
- Before: Code Graph to see the shape of the solution.
- After: Repo Docs to document the new structure.
- Related: Trust, verification and limits of AI output.
Common mistakes & fixes
- "Choose at least one seam" → scan first, queue a seam, then Apply.
- "At most 5 seams" → queue the rest for a second run.
- "A seam reference is not valid" → the result is stale; scan again.
- Expecting a compiled result → the pull request is proposed code; your CI is the build.
- "This result is no longer available" → scan again.
Same thing elsewhere
Not verified for this app. See channels.
Try it
Scan a public C# repository with Candidates to show = 10. Open the top class marked "Worth splitting", read its first seam, and press Preview plan. Do not apply; judge whether the proposed split matches how you would have cut the class.