Unslop

Unslop

One skill, two modes. Both remove AI-generated slop; the artifact decides the mode. Writing goes through the prose pattern catalog; code goes through a bounded, behavior-preserving cleanup.

Choose the mode

• Prose (replies, docs, PR descriptions, commit messages, log text) -> prose mode below
• Code (an explicit file list or a branch diff) -> code mode below

Prose mode

• NEVER use U+2013 (en dash) or U+2014 (em dash) in prose. Replace them with a period, comma, colon, or semicolon. Use ASCII - only for syntax, flags, identifiers, and true compound words, not as dash punctuation.

1. Scan for the patterns in references/slop-patterns.md.
2. Rewrite. Preserve meaning, match intended tone.
3. Add soul (below).
4. Self-audit: "What makes this obviously AI generated?" Fix remaining tells.

Adding soul

• Have opinions. React to facts instead of neutrally listing pros and cons.
• Vary rhythm. Short sentences. Then longer ones that take their time.
• Acknowledge complexity. "Impressive but also kind of unsettling" beats "impressive."
• Use "I" when it fits. First person isn't unprofessional.
• Let some mess in. Perfect structure looks machine-made.
• Be specific. Not "this is concerning" but "there's something unsettling about agents churning away at 3am."

Code mode

Use for a bounded cleanup of AI-generated code, not a general refactor. Keep behavior, public APIs, type hints, dependencies, and project conventions intact.

Scope and safety

1. Use the caller's explicit file list. Otherwise inspect the branch diff against its merge-base. Exclude deleted, binary, generated, vendored, and lock files.
2. Read references/cleanup-playbook.md before making a cleanup plan or judging a diff.
3. Lock observable behavior with existing or narrow regression tests before edits. A prose-only file has no behavioral seam: do not add wording-pinning tests.
4. Establish a green relevant baseline. If it cannot be established, stop and report it; cleanup on unverified ground is reckless garbage.
5. Prefer deletion, reuse, or a platform/standard-library capability before simplifying code. Make only behavior-obvious changes. When uncertain, keep the code.

Execution

• Write a per-file plan: deletion-ladder result, applicable categories, order, and risk.
• Work safest to riskiest: comments, dead code, defensive code, duplication, complexity, abstraction/boundaries, performance, coverage, then module size.
• For a multi-file pass, delegate independent files only through the runtime's currently exposed agent tools. Give each worker the exact file, scope, constraints, and required category review; batch only within available concurrency. The owner integrates and verifies every change.
• Use project-native tests, lint, type checks, and scanners. Do not invent unavailable tools or silently skip a gate; report genuinely inapplicable gates as N/A with the reason.
• If a change fails validation, revert only the faulty hunk, fix it if safety is evident, and rerun the affected validation. Escalate after three failed attempts on one file.

Review output

Report scope, baseline/behavior lock, cleanup plan, per-file results and deliberate skips, each quality gate, safety/behavior/quality review, issues fixed, net impact, and deferred risks. Never claim a passing gate without its observed output.

Required follow-up reads

Merged from remove-ai-slops (GPL-3.0-or-later, anntnzrb) and pstack unslop (MIT, Lauren Tan).