ai-slop

name: ai-slop description: Multi-agent AI-slop and revision-language sweep for academic manuscripts. Chunks the paper, dispatches parallel agents to flag reviewer-response framing, revision-tracking language, AI-slop tells, and padding (with confidence ratings), consolidates into Tier 1 / Tier 2 fixes, and applies them in batches. Use before submission or resubmission, especially after AI-assisted revision cycles. invocation: user

/ai-slop — Multi-Agent AI-Slop & Revision-Language Sweep

A deeper pass than /review-paper item 13. Used for revision cycles or AI-drafted manuscripts where reviewer-response framing and revision-tracking language need to be scrubbed in addition to the standard banned-words list. Agents flag and report; the user (with Claude's help) triages and applies.

When to Use

• Manuscript is approaching submission or resubmission.
• A revision cycle has added new content (often in response to reviewer comments) and you want to confirm the new prose doesn't read as reviewer-response.
• The draft used AI assistance and you want a defensible "we cleaned the prose" pass before shipping.
• /review-paper item 13 found surface slop, but you want the deeper multi-agent treatment with reviewer-response and revision-tracking categories.

Don't use for: line-by-line copy-editing of grammar/typos, or technical-correctness review (different tasks).

Workflow Overview

Phase 1. Context discovery       ← read CLAUDE.md, locate main file
Phase 2. Define chunks            ← split manuscript by section
Phase 3. Launch parallel agents   ← one per chunk, read-only, write reports
Phase 4. Consolidate findings     ← read reports, identify cross-chunk patterns
Phase 5. Triage into tiers        ← Tier 1 + recurring, Tier 2 grouped
Phase 6. Apply Tier 1 batch       ← unambiguous fixes, no further discussion
Phase 7. Apply Tier 2 group-by-group ← user approves each thematic group
Phase 8. Final verification grep  ← confirm zero hits in body text

Key rule: Agents do NOT edit the manuscript. They only write reports. The author decides which fixes to apply.

Phase 1 — Context Discovery

1. Read the project's CLAUDE.md to find the manuscript file path, paper topic, structure, and any revision context (revision cycle? AI-drafted?).
2. Locate the main .tex file and any \input / \include structure.
3. If the project uses a revision-tracking macro (e.g., \rev{...} for blue-coloring revisions), note it — patterns inside such wraps are typically the highest-density slop targets.
4. Check notes/ for prior ai_slop_* reports — if a sweep already ran, focus on newly added content rather than re-flagging.

Phase 2 — Define Chunks

Split the manuscript by major sections. Three chunks is the sweet spot — fewer means agents run out of attention; more means consolidation gets messy.

For each chunk, record: line range (or chapter ID), descriptive section list (used in the agent prompt).

Phase 3 — Launch Parallel Agents

Use the agent prompt template at references/agent-prompt-template.md. For each chunk:

1. Replace every <PLACEHOLDER> with the project-specific value (range, file path, section list, output report path).
2. Add the relevant section-specific note (math-heavy / results-heavy / fixed-in-prior-pass) — see template.
3. Save report to notes/ai_slop_<chunk-id>.md (e.g., notes/ai_slop_intro_background.md).

Launch all chunk agents in parallel (single message, multiple Agent tool calls). They run concurrently and each writes its own report. Use subagent_type: general-purpose unless the project has a more specific agent type.

The agent prompt enforces:

• Read-only (no edits).
• Four categories with confidence ratings (skip "low").
• Markdown table output grouped by category.
• Skip rules for citations, equation labels, math mode, LaTeX commands, etc.

Phase 4 — Consolidate Findings

Read each report. Look for patterns across reports:

• A word/phrase used in multiple sections is a "recurring pattern" (highest leverage — fix once, propagate).
• Reports often agree on the worst offenders → those go in Tier 1.
• Reports with concentrated flags in one subsection signal a draft hot-spot (often the most recently rewritten section).

Build a consolidated view organized by theme, not by chunk.

Phase 5 — Triage into Priority Tiers

Out of scope / skip:

• LaTeX % comments (invisible to readers).
• Conventional math-prose ("It follows that", "Hence", "Suppose") in proofs.
• Standard back-matter declarations (formulaic).
• Citations, labels, table cell values, model/author names.
• Established technical terms (e.g., "robust optimization" is a term of art).

For the canonical banned-words list and wordy-phrase replacement table, see ../../guides/paper-review-checklist.md item 13. The skill extends that list with the two revision-cycle categories (reviewer-response framing and revision-tracking language).

Phase 6 — Apply Tier 1 + Recurring (Batch)

Apply all Tier 1 fixes and recurring-pattern fixes in one pass without further discussion (this tier is pre-approved). Use parallel Edit calls where changes are on different paragraphs.

Watch for:

• Cross-references — verify any new \ref{...} resolves to an existing label.
• Revision-tracking macro wraps (e.g., \rev{...}) — surgical edits inside such a wrap keep the wrap; replacements that delete the wrapped text need new wrapping.
• Math-mode boundaries — don't accidentally edit text inside $...$ or display environments.
• Recurring patterns — when varying a repeated word/phrase, vary the closest pair first (large gaps tolerate repetition).

Phase 7 — Apply Tier 2 Group-by-Group

Present Tier 2 in thematic groups (e.g., "wordy/intensifiers" group, "transition phrases" group, "soft openers" group). For each group:

1. List all proposed changes (current → proposed) in a table.
2. Ask the user to approve, veto, or modify each.
3. Apply the approved subset in parallel.

Tier 2 is where author voice lives, so author approval matters more here than in Tier 1.

Phase 8 — Final Verification Grep

Run a single Grep across the manuscript for the canonical reviewer-talk and AI-slop patterns. The grep should return zero hits in body text (LaTeX comments and macro definitions in the preamble are OK).

Baseline pattern (extend with project-specific tokens — see "Adapting the verification grep" below):

[Tt]o address (concerns|the (concern|practical))|[Aa]s (suggested|recommended) by|[Ff]ollowing the reviewer|[Nn]ewly added|[Nn]ewly introduced|the revised manuscript|in this revision|now includes|[Nn]otably,|particularly notable|underscores that|delve into|leverage(d|s)? (the|our|a|to)|valuable insights|play a (crucial|vital|key) role|shed light on|in the realm of|navigate (the|a)|comprehensive review

If hits remain, decide each: false positive (technical context, leave) or residual fix.

Notes Convention

• During execution: reports go to notes/ai_slop_<chunk-id>.md.
• After all fixes are applied and the user is satisfied: the user moves the chunk reports plus a session summary to notes/done/ai_slop_<YYYYMMDD>/ (manual step — the skill does not auto-archive).

Common Pitfalls

1. Over-flagging in math-heavy sections. Use the math-heavy section-specific note in the agent prompt.
2. Technical-term collisions. "Robust" in robust optimization is a term of art; "comprehensive examination" is sometimes literal. Only flag in non-technical contexts.
3. Confidence calibration drift. Agents over-flag at first. Require high/medium ratings and skip low automatically.
4. Cross-reference breakage. When dropping a sentence, check that you don't break a \ref{...} or remove a label used elsewhere.
5. Recurring-pattern blindness. Single agents can't see across chunks — the user must look at all reports together.
6. Counting mismatches. Lists that say "Three Xs" and then have four items are a common AI tell.
7. Over-aggressive stripping. Some hedging is legitimate. Preserve author voice; don't strip everything.
8. Agent context limits. Don't give a single agent more than ~700 lines.

Adapting the Verification Grep

The baseline grep covers a generic taxonomy. Extend the right-hand side of each alternation with patterns specific to the manuscript's domain:

• ML/AI papers: tokens that paired with non-technical "robust" in the draft (e.g., "robust to noise", "robust to perturbations").
• Optimization papers: tokens that paired with "comprehensive" or "extensive" (e.g., "comprehensive analysis").
• Survey papers: tokens that paired with "various" or "myriad" (e.g., "various approaches", "myriad of methods").
• Empirical papers: tokens that paired with "leverage" (verb), "delve into", "shed light on".

Run the grep, eyeball the hits, decide each.

Reference Files

• references/agent-prompt-template.md — parameterized prompt for the chunk agents (Phase 3).
• ../../guides/paper-review-checklist.md — canonical banned-words list, wordy-phrase table, structural patterns (item 13). This skill extends item 13 with the revision-cycle categories.
• ../../skills/review-paper/SKILL.md — single-pass review skill. /review-paper references this skill as the deeper option for revision cycles.