Deslop

deslop

Purpose

Strip AI-writing tells (negative parallelism, significance inflation, hedge stacks, formatting reflexes, era vocabulary, uniform rhythm) from a draft so it reads human — with the edits bounded by deterministic tooling. It fixes the two observed failures of prompt-only humanizers: silent damage (facts, quotes, and code drift during "improvement") and over-correction (real writers, especially non-native ones, get sanded flat; absolute rules like "zero em dashes" produce a new tell instead of removing one).

When to use

• "deslop / de-AI this", "strip the AI slop / tells", "this reads machine-written — clean it up", "make it read human without changing meaning"
• Cleaning AI-assisted drafts before publishing, with proof nothing broke

When NOT to use

• Score or report only, no edit wanted → slop-lint.
• Checking an edit that already happened → deslop-verify.
• Grammar/typo fixing, tone shifts, shortening, translation — different jobs; never smuggle them into a de-slop pass.
• "Beat the AI detector" requests: decline the detector-evasion framing; offer editorial cleanup instead (this skill improves writing, not evasion).

Workflow

Sibling scripts (installed together with this skill): LINT="${CLAUDE_SKILL_DIR}/../slop-lint/scripts/slop_lint.py", VERIFY="${CLAUDE_SKILL_DIR}/../deslop-verify/scripts/verify_edit.py". If a sibling is missing (partial install), do not edit blind — the no-op gate, span bounds, and verification all depend on the tools. Say which tool is missing, deliver playbook-based suggestions (a list of flagged spans and proposed transforms, not an applied rewrite), and point to the full install.

1. Lint first (no-op gate). Save the original to a temp file; run python3 "$LINT" original.md --genre <genre> --json > lint.json. If the band is clean (score < 25), return the text untouched and say why — editing clean text is over-correction, not service. If confidence is low (under 150 words), the score gates don't apply: make bounded playbook edits to clearly flagged spans only, note the low confidence, and still verify in step 5.
2. Inventory before editing. Note: protected spans (quotes, code, numbers, URLs, names), the writer's own habits (their dashes/triads are their voice), and any consistent dialect/L2 features. Read references/tells-playbook.md for the per-family transforms and the dialect guard — it is the edit contract, not advice.
3. Bounded edits. Edit only lint-flagged spans plus their sentence; copy every unflagged sentence verbatim. Apply the playbook transform for each active family; prefer deleting padding over swapping synonyms; never invent facts, anecdotes, typos, or slang to "add humanity".
4. Re-lint the result. Accept the pass only if the score dropped ≥15 points or the band is now clean. If not, do one more constrained pass on remaining flagged spans. Two passes maximum — then stop and report what remains rather than thrash.
5. Verify, fail closed. python3 "$VERIFY" original.md edited.md --lint-report lint.json. On any non-zero exit: exit 2 (hard invariant broken) → restore the lost content or revert to the original; exit 1 (verify could not run) → treat the rewrite as unverified and do not present it as done. Never present a rewrite that failed or skipped verification.
6. Report. Deliver the edited text plus: before/after scores, families fixed, verify result (state it as surface integrity, not proven meaning), and anything deliberately left (dialect features, writer habits, meaning-bearing hedges, remaining warn-band items).

Output spec

The edited text, changed as little as possible: all hard invariants pass in deslop-verify, edit ratio ≤ 0.30 (near 0 for clean input), length within ±25%, lint score reduced ≥15 points or banded clean, unflagged prose byte-equal, and a report of before/after scores + verify status. On clean input: the original text, unchanged, with the no-op explanation.

Gotchas

• The dialect/ESL guard is non-negotiable. Never convert nonstandard, regional, or L2 English toward Standard American English; if it could be dialect or error, leave it. Evidence and the full rule: the playbook's "Dialect and ESL guard" section (26% vs 92% marker retention; 61.2% detector false-positive rate on non-native writing).
• Meaning-bearing hedges are content. "May cause drowsiness" and "does not establish causation" must survive; only stacked hedges are tells.
• Zero em dashes is itself a tell. Reduce flagged clusters; never purge.
• Don't chase score 0. Below warn is done. Looping to a perfect score Goodharts the linter and flattens voice — two passes, then stop.
• A second deslop of your own output should change < 2%. If it doesn't, the first pass was over-editing; revert and report.
• Drafts produced late in a long chat session tend to carry more tells than fresh-session drafts — lint and deslop the final text in a fresh pass rather than polishing mid-thread.
• Rewriting wholesale because "everything sounds AI" is refused by design; the edit budget exists to protect the writer.

Pointers

• references/tells-playbook.md — per-family transforms, what NOT to flag, dialect guard, evidence with primary sources.
• Sibling tools: slop-lint (scoring, thresholds provenance) and deslop-verify (invariants, over-correction gates).