Survey an existing codebase for the structural change worth making, then rank the candidates, crown one, and write it up as a reviewable proposal. Use when the user says "where should I refactor", "this code is hard to change", or "audit the module boundaries". It proposes and never edits code.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Survey an existing codebase for the structural change worth making, then rank the candidates, crown one, and write it up as a reviewable proposal. Use when the user says "where should I refactor", "this code is hard to change", or "audit the module boundaries". It proposes and never edits code.
license
MIT
allowed-tools
Bash, Read, Grep, Glob, Write, Task, Agent
metadata
{"internal":false}
refactorkit
The survey you run on a codebase you already have. refactorkit reads the repo, meaning its churn, its module shapes, and its tests, finds where the structure is costing you, ranks the candidates against two gates, crowns exactly one, and writes it up as a proposal somebody can argue with.
It is a single procedure, optionally scoped to a subtree (/refactorkit src/payments). There are no modes: everything it does, it does the same way on every run.
It proposes, it never edits
refactorkit writes exactly one Markdown file and touches no source code. No moves, no renames, no extractions, no "small tidy-up while I'm in here." The name says refactor; the skill only ever says where and why, and the doing belongs to whoever picks the proposal up. That boundary is the whole reason the report can be trusted, because a survey that also edits has already decided, and you're reading a rationalization rather than a recommendation.
Edit is deliberately absent from this skill's tools. The one file it creates is new, so Write covers it.
When this fires
"Where should I refactor this", "what's wrong with this codebase's structure", "find refactoring opportunities", "this code is hard to change", "improve our architecture", "audit the module boundaries", "the seams here are a mess", "/refactorkit".
Four boundaries, one line each:
Not a code review. A review reads a diff for defects in work someone just did. refactorkit reads the tree for structure that predates the diff. A "review my changes" request is not this skill.
Not documentation. Describing the architecture as it stands is a different job; refactorkit only ever describes an architecture that doesn't exist yet.
Not visual design. Nothing here concerns UI, styling, or a design system, despite what "design" suggests when it sits next to "refactor."
Not a decision record. refactorkit produces a proposal, not a ratified decision. When the proposal supersedes an existing architecture decision, it says so and routes to whatever records decisions in this repo.
The vocabulary
Use these terms precisely; they're what keeps the findings from collapsing into generic smell-listing.
Module. Anything with an inside and an outside: a file, a class, a package, a service. Not a language construct, a boundary.
Interface. Everything a caller must know to use the module correctly. Not just the signature: also the ordering constraints, the required config, the side effects, the error contract, and the invariants nobody wrote down. Interfaces are almost always bigger than they look.
Depth. How much behaviour a module hides, relative to how much interface it makes you learn. A deep module buys a lot of hidden work for a small interface. A shallow one charges nearly the same either way.
Seam. The place a system can be split without tearing: where one side can be replaced, faked, or tested without the other.
Adapter. A module whose entire job is reshaping data between two others. Adapters aren't wrong; a cluster of them is a symptom that two sides disagree about a shape.
Locality. How much of one conceptual change lands in one place. Good locality means a feature change touches few files; poor locality means it fans out.
Leverage. How much future change a fix makes cheaper. It's what breaks ties between candidates that are otherwise equally sound.
Use the project's own domain terms throughout. If the codebase says ledger, tenant, shipment, the proposal says those words. Falling back on generic "component," "service," "API," or "layer" is the single clearest sign a finding was pattern-matched rather than read, and it's what makes architecture advice sound plausible while saying nothing.
The four friction patterns
A closed set. These four are what refactorkit looks for, and it looks for nothing else. The closure is load-bearing: an open-ended hunt for "problems" produces the same generic list on every codebase, which is exactly the output this skill exists to replace. All four reduce to depth, since each one is a different way an interface costs more than the behaviour behind it.
Pattern
What it is
The shape to look for
Who finds it
Shallow interface
the interface costs about as much to learn as the implementation costs to read
a unit whose members each forward to exactly one member of exactly one other unit
subagent
Adapter prevalence
modules that exist only to reshape data between two others, where the finding is that two sides disagree on a shape, never "delete the adapters"
a cluster of translation-only units sitting on one boundary, translating in both directions
subagent
Untested coupling
behaviour you can only test by standing up its collaborators
tests that construct more collaborators than they make assertions, or an untested unit whose neighbours are all tested
subagent
Poor locality
one conceptual change fans out across many files
files that repeatedly change in the same commits while living in different parts of the tree
main session
Signals are shapes, never names. Do not hunt for *Mapper, *Adapter, *DTO, *Service, or any other suffix. Those conventions belong to one or two language communities and mean nothing in a Go, Rust, Elixir, or PHP repo, and a named example is precisely what an agent pattern-matches on instead of reading. Derive the repo's own conventions first (the guardrails read below usually hands them to you), then look for the structure described above under whatever names this codebase happens to use.
The two gates
Every candidate clears both before it is listed at all. A candidate that fails either one is discarded silently: it does not appear as a weaker finding, and it does not appear with a caveat.
The deletion test. If this module vanished and its callers absorbed what it did, would complexity concentrate where it belongs, or merely relocate? Concentration is the win. Relocation is churn wearing a refactor's clothes, and printing it is how a report gets long and stops being read. State the verdict with one sentence of evidence: which callers absorb what, and why that is or isn't a better home.
Interface as test surface. After the proposed change, can the behaviour be tested through the new interface alone, without reaching inside? If not, the seam is in the wrong place: the candidate is unfinished, not ready. This gate is what stops a proposal that merely moves the boundary from passing as a proposal that improves it.
Both verdicts are mandatory on every listed candidate, in the terminal output and in the file. A candidate whose gate verdicts aren't stated is not a weaker candidate; it's an unevidenced one.
The strength scale
Rate each surviving candidate against this table rather than by feel, because an undefined badge is a vibe wearing a label, and strength is the one column a reader trusts without checking.
Strength
What it takes
How it ranks
strong
both gates pass, the files sit in the churn top slice, and the blast radius is named in actual files
crownable
moderate
both gates pass, but the code is cold or the blast radius is wide
crownable only when nothing is strong
weak
gates pass, but the win is stylistic, and nothing changes about what a caller must know
listed, never crowned
Procedure
1. Read the guardrails first
Before looking at any code, read what the repo has already decided:
a repo-root context or glossary file (commonly CONTEXT.md) for the ubiquitous language, which is where the project's own domain terms come from, and where its naming conventions usually surface;
docs/adr/ (or wherever this repo keeps architecture decision records) for decisions already made.
A proposal that reverses an accepted decision record must say so by its number and route to a superseding record rather than quietly re-suggesting a design that was already rejected. Reversing a decision is legitimate; doing it without noticing is not.
Neither file existing is normal. Degrade silently, with no warning and no commentary.
2. Rank by churn, then cluster
git log --format= --name-only --since=<a year or so> gives both signals refactorkit needs in one read, with no tool beyond git: how often each file changes (the churn ranking) and which files keep changing together (the co-change pairs that become the poor-locality finding).
Rank files by change count, then group the hot ones into 3–5 coherent areas, by directory, by feature, by whatever seam the repo actually has. Honour the scope argument when the user gave one: scope narrows the tree, it doesn't change the method.
No git history, meaning a shallow clone or not a repo at all, means no ranking. Scan by structure instead and say plainly that the prioritisation was skipped, so nobody reads the coverage line as more than it is.
3. Fan out, one agent per cluster
Dispatch one subagent per area. Each applies the three read-the-code patterns from The four friction patterns to its own slice, runs both gates itself, and returns a list of candidate records, or an explicit empty result, which is a perfectly good answer.
Give every agent the same fixed return shape, because the main session has to rank findings it never gathered. Freeform prose can't be ranked without re-reading everything the agent read, which throws away the entire reason to fan out:
Poor locality stays in the main session. It's derived from the co-change pairs already in hand, it costs no file reads, and a per-area agent structurally cannot see files that change together while living in different areas. Handing it to an area agent guarantees it is never found.
No subagent tool available → scan the top churn slice inline instead, applying the same patterns and the same gates, and say which path the run took. For a public skill this fallback is a large share of runs rather than an edge case, so it is a real mode with a real coverage claim, not a degraded apology.
4. Synthesize, gate, rank, crown
Drop any record missing a gate verdict. Dropped, not repaired, because an unevidenced gate is exactly what the deletion test exists to catch, and repairing it in the main session means asserting a verdict on code you didn't read.
Dedupe across areas; the same seam often surfaces from both sides of it.
Rank on strength, then leverage, then lowest blast radius.
Crown exactly one. Never a weak candidate, whatever else is on the list. The rest become runners-up, in order.
Crowning one is the work. A list of five equal-looking options is the state you were already in before running this.
5. Report, then write the artifact
Print a compact table first (candidate, pattern, strength, blast radius), then the crowned move with its reasoning, then the runners-up.
Every report carries a coverage line, in the terminal and in the file: "ranked 1,240 files by churn, read the top 40 across 4 areas." A cap nobody can see is a lie about completeness, and it's the difference between "there's nothing structurally wrong here" and "I looked at 3% of it."
The artifact goes to docs/refactor/refactor-<slug>-YYYY-MM-DD.md, built from a lowercase type prefix, a short lowercase kebab-case subject slug, and the ISO creation date at the end. Keep that creation date stable when the file is edited; a re-run on the same subject updates the same file in place rather than spawning a dated copy. When the repo already has an established home or naming scheme for proposal documents, that convention wins.
It contains the coverage line, the ranked table, then one section per candidate covering the problem, the proposed shape, why it is deeper, both gate verdicts, and the blast radius, with a Mermaid before/after diagram for the crowned candidate at minimum. GitHub and most editors render Mermaid natively, which is what makes a generated HTML report unnecessary.
This file is durable and committable, since it's a proposal meant to be reviewed in a PR, not scratch. refactorkit still never commits it.
No writable filesystem (a browser-based agent) → print the artifact as a codeblock under its canonical path and skip the write.
6. Hand off
Write this section in the procedural register: one instruction per sentence, active voice, present tense, no metaphor.
What changed. Report the file written, and explicitly that no source file was touched. Where it landed. Give the path. Next. Take the crowned candidate to a grilling: the write-up is already plan-shaped, and what it lacks is interrogation rather than drafting. Name grillkit when it's installed; otherwise say plainly that the move is to interrogate the proposal yourself, covering the assumptions, the failure paths, and the blast radius, before anybody builds it.
Runners-up: route to a decision-record skill (domainkit when installed) when the proposal supersedes an existing record, and to a build skill (implementkit when installed, otherwise just building it) once the shape is settled.
On an empty result, say there is no next move. See below.
When nothing clears the gates
A clean result is a real outcome, and refactorkit reports it as one: name the coverage, say nothing cleared the gates, and write no file.
This matters more than it looks. A survey obliged to produce findings will manufacture them, and manufactured architecture advice reads exactly like the real thing: same vocabulary, same confidence, same shape. The deletion test is only a genuine gate if "nothing here" is a legal answer.
The terminal line has to carry the coverage, as in "scanned 38 files across src/; nothing cleared the deletion test", or an empty result is indistinguishable from a scan that never ran.
Notes
Zero source mutation, always. One new Markdown file is the entire write surface. Anything else, whether a rename, an extraction, or a "quick fix", is out of bounds even when it's obviously correct.
Read-only shell only. git history, file reads, greps. Never install or run an analysis tool to generate evidence: probing for one and parsing its output couples this skill to a format that changes on a minor release. If the repo already has such an artifact on disk, like a dependency graph or a coverage report, read it. Never generate one.
Route, don't launch. Name the next kit and its one-line invocation; don't invoke it. Whether to act on a proposal is the reader's call, and a survey that chains itself into the build has taken that call away.
Route, don't require. Every recommendation degrades to a plain action when the named kit isn't installed. refactorkit is useful in a bare repo with nothing but git.
Follow the repo over these defaults. An established artifact location, a documented convention, or a stated policy in the repo's agent-guide file (CLAUDE.md or an equivalent) wins, so say that you followed it.
