Deslop

Deslop

Use this skill for broad code-quality cleanup where the goal is to make the codebase smaller, clearer, more strongly typed, and less defensive or legacy-laden.

Operating Model

Treat this as coordinated research plus implementation, not a blind cleanup sprint.

1. Inspect the repo first: language, package manager, test commands, lint/typecheck commands, dependency graph tooling, and current git status.
2. If a subagent or multi-agent tool is not already visible, search for one before falling back. If subagents are available, dispatch exactly 8 independent subagents, one for each lane below. If subagents are unavailable, run the same lanes sequentially and say so.
3. Give each lane a narrow, evidence-based prompt. Require detailed research, a critical assessment of current code, recommendations with confidence levels, and implementation of only high-confidence recommendations.
4. Integrate the results yourself. Review every diff, reconcile conflicts, and reject speculative or risky changes.
5. Verify with the strongest practical checks: typecheck, lint, targeted tests, dependency analysis, and any affected app build.

The 8 Lanes

1. Deduplication and DRY Find duplicated logic, repeated utilities, copied UI, repeated config, and near-identical workflows. Consolidate only when it reduces complexity or clarifies ownership.

2. Shared Type Definitions Find local type/interface/model/schema definitions that should be shared. Consolidate types into existing shared modules or a minimal new shared home that matches repo conventions.

3. Unused Code Use tools such as knip, language-native analyzers, compiler diagnostics, static search, and test/build evidence to find unused exports, files, dependencies, scripts, and dead routes. Remove only after proving no real references remain.

4. Circular Dependencies Use tools such as madge, compiler/module diagnostics, dependency graph commands, and import tracing to find cycles. Break cycles by moving shared contracts, inverting dependencies, or splitting mixed-responsibility modules.

5. Weak Types Find any, unknown, casts, unchecked dictionaries, untyped errors, nullable escape hatches, and language equivalents. Research call sites, package types, runtime schemas, and domain models before replacing with strong types.

6. Defensive Error Handling Find try/catch blocks and equivalent fallback/error-hiding patterns. Remove defensive handling that has no clear boundary role. Keep and clarify handling for unknown input, external systems, parsing, IO, network, user data, and intentional recovery paths.

7. Deprecated, Legacy, and Fallback Paths Find deprecated APIs, legacy compatibility branches, duplicate old/new flows, migration leftovers, feature-flag fallbacks, and stale adapter paths. Remove when the current code path is clear and tests/builds confirm the old path is no longer needed.

8. AI Slop, Stubs, LARP, and Comments Find stubs, placeholder logic, fake implementations, TODO theater, unnecessary comments, comments about in-progress replacement work, and verbose AI-style explanation. Remove them or replace with concise comments that help a new maintainer understand non-obvious code.

Subagent Prompt Requirements

Each lane prompt must include:

• the exact lane scope and files/tools likely to matter
• instructions to research before editing
• instructions to write a critical assessment before implementation
• permission to implement only high-confidence recommendations
• a warning not to perform broad rewrites, speculative removals, or style churn
• required output: findings, evidence, recommendations not taken, files changed, and verification run

Decision Standards

• Prefer deletion over abstraction when code is genuinely unused or obsolete.
• Prefer existing repo patterns over new architecture.
• Do not consolidate code just because it looks similar; shared abstractions must reduce cognitive load.
• Do not replace weak types with fake precision. Strong types must be backed by actual data shape evidence.
• Do not remove error handling at trust boundaries.
• Do not hide uncertainty. Leave lower-confidence recommendations as assessment notes rather than code changes.
• Preserve user work in a dirty git tree. Never revert unrelated changes.

Verification

Before finalizing, run the repo's relevant checks. Common examples:

pnpm typecheck
pnpm lint
pnpm test
pnpm exec knip
pnpm exec madge --circular .

Use the repo's actual package manager and commands. If a tool is missing, install only with user approval or use the nearest existing analyzer.

Final Response

Report:

• the eight lane outcomes
• the highest-impact changes implemented
• notable recommendations intentionally left unimplemented
• verification commands and results
• any remaining risk or follow-up that needs human judgment