Deslop

Deslop

Aggressive whole-codebase cleanup for AI-generated code. Eight specialized passes, each a fresh-eyes subagent with a narrow mandate. Based on the pattern from Shaw (@shawmakesmagic). The framing: "the quality of your vibecoded slop is horrible… fortunately, there is a fix."

What It Does (the 8 passes)

1. DRY / dedupe — consolidate duplicated code where it reduces complexity
2. Consolidate shared types — unify type definitions that should be shared
3. Remove dead code — find unreferenced code with stack-appropriate tools
4. Untangle circular dependencies — surface and break cycles
5. Remove weak types — replace any/unknown/T.untyped/interface{} with real types
6. Remove defensive try/catch — delete error handlers that swallow or hide errors without a real purpose
7. Remove legacy/deprecated/fallback code — consolidate to a single current code path
8. Remove AI-slop comments — strip in-motion commentary, stubs, larp, and obvious-code comments

When to Use

• After a heavy AI-coding session (you've generated a lot of code quickly)
• Before merging a long-running feature branch
• On a codebase inherited from another AI session or another developer
• As a periodic deep-clean (quarterly, per-release)
• Position in pipeline: heavy AI coding → /deslop → /simplify → /adversarial-review → /finalize → commit

When NOT to Use

• Trivial changes or recent edits — use /simplify instead
• A codebase you don't have time to review carefully — this will move files and kill code
• Mid-feature — finish the feature first
• Without a clean git tree (see Safety below)

Safety — Non-Negotiable

This skill deletes code. Before running anything:

1. Verify clean git tree — git status must be clean. Refuse to proceed otherwise. If uncommitted changes exist, tell the user to stash or commit.
2. Create a dedicated branch — git checkout -b deslop/YYYY-MM-DD before any pass writes. Never run on main/master.
3. Commit between passes — each pass gets its own commit so individual passes can be reverted without redoing the rest.
4. High-confidence only — each subagent implements only changes it is confident about. Borderline calls go in its report for human review, not in the code.
5. Install the tools. If a pass's canonical tool isn't installed, install it — don't fall back to grep/ripgrep/manual scanning for passes that have a dedicated tool. The whole point of the skill is to use the right analyzer; grep doesn't see dynamic references, doesn't do reachability analysis, and doesn't detect cycles. See "Tool Installation Policy" below.

Tool Installation Policy

Each pass has a canonical tool for its stack (see references/passes.md and the per-language references). When a subagent starts a pass, it must ensure the tool is available — and install it if not.

JavaScript / TypeScript — zero-install via npx

All canonical JS/TS tools run through npx --yes without modifying package.json:

npx --yes knip        # dead code
npx --yes madge --circular src/   # circular deps
npx --yes ts-prune    # alt: unused exports

npx downloads to the global npm cache on first run. No project changes. This is the default — don't add anything to package.json for a one-off deslop run.

Ruby — gem install to user gems

gem install debride
bundle exec srb tc     # only if sorbet/config exists in the project

Installs to the user's gem path. Don't add to the project's Gemfile for a one-off deslop run.

Go — go install to $GOPATH/bin

go install honnef.co/go/tools/cmd/staticcheck@latest
go install golang.org/x/tools/cmd/deadcode@latest

Installs a binary in $(go env GOPATH)/bin. Standard Go tooling practice.

If install fails (no network, unsupported Go version, etc.): report the failure, name the tool, and skip that pass. Do not substitute grep. A skipped pass is an honest result; a grep-based result is misleading — it looks like an analysis but misses most of what the dedicated tool would catch.

When a pass legitimately has no tool (e.g. circular deps in Ruby without packwerk, type consolidation without a type system): say so explicitly in the pass report. "Skipped — no type system in use" is a fine result. "Skipped — ran grep instead" is not.

Input Parsing

Parse $ARGUMENTS:

• No args → run all 8 passes on the repo root
• Path (e.g. app/services or src/) → scope all passes to that path
• --dry-run → run the research phase for each pass, print a critical assessment, but make zero code changes
• --passes 1,3,8 → run only the listed pass numbers (comma-separated)
• --stack js|ruby|go → force the stack (skip auto-detection)

Step 1: Detect the Stack

Look for lockfiles in the target path (or repo root if no path given):

A repo may have multiple stacks (e.g. Rails app with a Vue frontend). Record all stacks found. When dispatching each pass, tell the subagent which stacks are in scope for the files it will touch. For a nested frontend (e.g. app/frontend/), pass 3 and 4 dispatch a JS/TS-aware subagent for that subtree and a Ruby-aware one for the rest.

If --stack is given, skip detection and use the forced stack.

For stack-specific tooling details (commands, install, quirks), read the relevant reference file only when a pass that uses that tool is about to run:

• references/javascript.md — knip, madge, ts-prune, TypeScript type checking
• references/ruby.md — debride, packwerk, Sorbet notes
• references/go.md — staticcheck, go mod graph, deadcode

Step 2: Research Phase (parallel, read-only)

For each pass that is in scope (not skipped by --passes), dispatch a subagent in the same turn so all 8 research tasks run in parallel.

Use the Agent tool with subagent_type: general-purpose. Each research subagent:

• Reads code only — makes no edits
• Runs read-only analysis tools (linters, knip --dry-run, madge --circular, etc.)
• Produces a critical assessment: what's wrong, where, how confident, how to fix
• Categorizes each finding as high-confidence (safe to auto-apply) or needs-review
• Returns a structured report

Research subagent prompt template (fill in {pass_name}, {pass_instructions}, {stack_info}, {scope_path}):

You are a senior engineer performing a cold read of this codebase. You are
scoped to a single concern: {pass_name}.

## Stack Info
{stack_info}

## Scope
Target path: {scope_path}

## Your Mandate
{pass_instructions}

## Output Format

Return two sections:

### Critical Assessment
A prose summary of the state of the code for your concern. What patterns do
you see? How pervasive is the problem? Any notable surprises?

### Findings
A list of concrete, actionable findings. For each finding, provide:

- **Location:** file path(s) and line range(s)
- **Issue:** what's wrong
- **Fix:** specific, concrete change
- **Confidence:** high | needs-review
- **Rationale:** why this is safe (for high-confidence) or what the risk is (for needs-review)

Be honest about confidence. If you can't verify a symbol is unused across the
whole codebase, it's needs-review. If removing a try/catch might swallow a real
error condition, it's needs-review. When in doubt, err toward needs-review.

## Do Not
- Make any file edits
- Run stateful commands (git commits, migrations, package installs)
- Go beyond your concern's scope

See references/passes.md for the full per-pass instructions, stack-aware tool invocations, and what counts as high-confidence vs needs-review for each pass.

Step 3: Apply Phase (sequential, one commit per pass)

If --dry-run, skip this step and jump to the report.

For each pass with high-confidence findings, in this order:

1. Dead code (pass 3) first — removing unused code shrinks the surface area for later passes
2. AI slop comments (pass 8) — comments are low-risk
3. Legacy/deprecated (pass 7)
4. Defensive try/catch (pass 6)
5. Weak types (pass 5)
6. Consolidate shared types (pass 2)
7. DRY / dedupe (pass 1) — most invasive, runs after others have cleaned the field
8. Circular deps (pass 4) — structural changes, run last

For each pass:

1. Dispatch an apply subagent. Give it the findings from its research pass and instruct it to implement only the high-confidence items. Provide file contents fresh (don't rely on research-phase cached state).
2. After the subagent returns, run the project's type checker / linter / test suite if available — record pass/fail
3. Commit with message deslop: <pass_name> (e.g. deslop: remove dead code)
4. If tests or type-check fail, do not proceed to the next pass. Report the failure and hand back to the user.

Apply subagent prompt template:

You are implementing a single cleanup pass: {pass_name}.

## Stack Info
{stack_info}

## Findings to Apply
{high_confidence_findings}

## Rules
- Implement ONLY the findings above. Do not expand scope.
- If a finding looks wrong on closer inspection, skip it and note why in your
  report. Safety over completeness.
- Make minimal edits. Don't reformat, rename, or touch surrounding code.
- If a finding would break a public API, skip it and flag it.

## Output
Return a summary:
- Applied: N findings
- Skipped: N findings (with brief reason for each)
- Files touched: list

Step 4: Report

After all passes (or after --dry-run research), produce a single report:

## Deslop Report

**Stack:** <detected stacks>
**Scope:** <path>
**Mode:** <full | dry-run | --passes 1,3,8>
**Branch:** deslop/YYYY-MM-DD (or "no changes — dry-run")

### Pass Summary
| Pass | Findings | High-Conf | Applied | Needs-Review | Skipped |
|------|----------|-----------|---------|--------------|---------|
| 1. DRY/dedupe | ... | ... | ... | ... | ... |
| 2. Shared types | ... | ... | ... | ... | ... |
| 3. Dead code | ... | ... | ... | ... | ... |
| 4. Circular deps | ... | ... | ... | ... | ... |
| 5. Weak types | ... | ... | ... | ... | ... |
| 6. Defensive try/catch | ... | ... | ... | ... | ... |
| 7. Legacy code | ... | ... | ... | ... | ... |
| 8. AI slop comments | ... | ... | ... | ... | ... |

### Needs-Review Items
Grouped by pass. Each item includes location, issue, suggested fix, and why
it was deferred. Review these manually.

### Commits
List of commits created on the branch.

### Next Steps
- Review needs-review items
- Run the full test suite
- Run `/simplify` + `/adversarial-review` before merging

Common Mistakes

• Running on a dirty tree. The whole point of commits between passes is clean reverts. If you skip the clean-tree check, a bad pass can mix with the user's in-flight work and become a nightmare to untangle.
• Skipping the research phase. Going straight to edits produces the cargo-cult behavior Shaw was railing against. Research first, commit to findings, then apply.
• Applying needs-review items. If the subagent wasn't confident, neither should the apply phase be. Surface them in the report and let the user decide.
• Running pass 1 (DRY) before pass 3 (dead code). You'd waste effort deduping code that's about to be deleted.
• Running all passes in parallel during apply. The passes can conflict (e.g. type consolidation moves a type the dead-code pass wants to remove). Research in parallel, apply sequentially.
• Treating missing tools as blocking. If knip isn't installed, the JS/TS dead-code pass reports "skipped — install knip" and the rest of the run continues. Never halt the whole flow because one tool is missing.

References

• Shaw's original prompt: https://x.com/shawmakesmagic/status/2044269097647779990
• references/passes.md — full per-pass instructions
• references/javascript.md — JS/TS tooling (knip, madge, ts-prune)
• references/ruby.md — Ruby/Rails tooling (debride, packwerk)
• references/go.md — Go tooling (staticcheck, go mod graph)