Deslop

Deslop — AI Writing Pattern Removal

Strip AI-generated writing patterns from any content. Works on existing files or inline text.

When to Use

• After drafting any customer-facing content (blog posts, emails, landing pages)
• During content QA, manual or automated
• When reviewing existing content for quality
• Before publishing or committing content

The 22 AI-Tell Patterns

Phrasing Flags

Rhythm Flags

Authenticity Flags

How to Run

On a single file

/deslop content/blog-posts/my-post.mdx

On all blog posts in a project

/deslop --all

Workflow

1. Scan: Read the file(s)
2. Flag: For each pattern, grep or manually identify violations
3. Count: Report violation counts per pattern
4. Fix: Apply fixes inline — preserve meaning, change delivery
5. Verify: Re-scan to confirm clean

Automated Checks (greppable)

These patterns can be detected with grep/regex:

# Em-dashes (count, max 5)
grep -c '—' "$FILE"

# Corrective antithesis
grep -iE "(It'?s not about|Not [A-Z].*\. But |Not [A-Z].*\. It'?s )" "$FILE"

# Dramatic pivots
grep -iE "(But here'?s the (thing|catch|kicker|reality)|Here'?s what (most|many|few))" "$FILE"

# Soft hedging
grep -iE "(It'?s worth (noting|mentioning|remembering)|It'?s important to (note|remember|understand)|Something (we'?ve|I'?ve) observed)" "$FILE"

# AI-tell words
grep -iE "\b(delve|landscape|realm|tapestry|leverage[sd]?|paradigm|robust|utilize[sd]?|synergy|elevate[sd]?|foster|holistic|streamline[sd]?|cutting-edge|game-?changer|deep dive)\b" "$FILE"

# Gift-wrapped endings
grep -iE "^(In (summary|conclusion)|To sum up|The bottom line|All in all|Ultimately,)" "$FILE"

# Throat-clearing
grep -iE "^(Let'?s (explore|unpack|dive|take a closer|break down|look at)|In this (article|post|guide|section))" "$FILE"

# Transition word abuse (at paragraph start)
grep -iE "^(Moreover|Furthermore|Additionally|Consequently|Nevertheless|Nonetheless)" "$FILE"

# Negation framing (#19)
grep -iE "\b(isn'?t just|isn'?t simply|not just about|goes beyond merely|more than just a)\b" "$FILE"

# "From X to Y" coverage spans (#20)
grep -iE "from (startups?|beginners?|small|large|simple|complex|novice|expert)" "$FILE"

# "Whether X or Y" inclusivity (#21)
grep -iE "whether you'?re? (a |an )?\w+" "$FILE"

Manual Checks (require human judgment)

• Staccato repetition (consecutive short sentences)
• Cookie-cutter paragraphs (uniform length)
• Perfect punctuation (no fragments or rule-breaking)
• Copy-paste metaphors (same metaphor repeated)
• Overexplaining (defining obvious terms)
• Generic examples (vague company references)
• Superlative stacking (multiple weak adjectives)
• False balance (hedging every position)
• #22 Tricolon / rule of threes — compulsive 3-part parallel lists ("fast, reliable, and scalable" / "we researched, analyzed, and synthesized"). 2 instances = fine, 4+ = pattern flag.

Severity Levels

Integration

• Automated QA pipelines: Run the greppable checks as a pre-publish gate.
• Content-creator skills: Reference this skill in a final quality checklist.
• Manual review: Use the full 22-pattern checklist before publishing.
• Existing content audit: Run across older posts to find and fix legacy violations.

Known gotchas

• The grep checks are noisy on technical writing. "Robust" is a legitimate engineering word; "leverage" is a legitimate lever-based metaphor. Treat hits as candidates to reconsider, not auto-deletes.
• Em-dash count alone is misleading. A long technical post can earn 8–10 em-dashes legitimately; a 600-word blog post should rarely have more than 3.
• Bulk rewrites flatten voice. Apply the fixes one pattern at a time on a draft, not all at once. Reviewers can spot a globally rewritten draft.

Anti-patterns

• Treating the pattern list as a banned-words list. The patterns are tells, not slurs. The fix is rewriting the sentence, not deleting the word.
• Running deslop on the first draft. Pattern detection works best after the author has already done one human pass.

Validated patterns

• The "negation framing" check (#19) catches the highest-volume tell — most AI-written drafts hit it within the first 100 words.
• Severity levels (BLOCK / WARN / NOTE) work better than a flat list for triage.

Credit

Patterns adapted from Tahi's "12 Red Flags of AI Writing" (mooch.agency, Feb 2026) and Jejomar Contawe's "The Single Most Prevalent AI Writing Tell" (Medium, Feb 2026), plus six additional patterns observed in the wild.

Self-improvement

This skill ships with a lightweight feedback loop. Adopt or ignore — the skill works without it.

Trigger a review when:

• The user explicitly disagrees with a fix (strongest signal — log immediately; 2–3 corrections on the same pattern → promote to body or adjust the pattern).
• A novel AI-writing tell starts appearing across multiple drafts.
• learnings.md crosses ~100 bullets (consolidation time).
• The skill mis-triggers or fails to trigger.
• A new model generation changes the writing tells (e.g., new go-to phrases).

Consolidation pass (5–10 min, weekly or threshold-driven):

• Each entry gets one fate: apply, capture, or dismiss.
• Apply = merge into Known gotchas / Anti-patterns / Validated patterns, or add a new pattern to the list.
• Capture = leave in learnings.md for now.
• Dismiss = delete.
• Bump last-consolidated: in frontmatter.