Sanyam
Back to Agentic AI

Agentic AI / Reusable artifact

Humanize and Unslop Skill

A Claude Code Skill that edits AI-drafted writing back into something a person would actually write: banned-word lists, banned punctuation, and a self-audit pass, or a detect-only mode that flags the tells without touching the text.

Skills

Most "make it sound less AI" prompts are one line, something like "rewrite this so it doesn't sound like ChatGPT." Claude does something for a paragraph, then drifts back to the same tics by the third page. This skill exists because "sound human" isn't one instruction, it's a checklist, and checklists are exactly what skills are for.

What it does

The skill runs two distinct jobs depending on what you ask for:

  1. Edit (default). Hand it a draft. It scans for banned vocabulary (delve, leverage, robust, testament, and about twenty more), banned punctuation (em dashes, mid-sentence colons, curly quotes), and a list of named patterns like weasel attribution ("experts agree"), synonym cycling, and summary-recap endings ("in conclusion..."). It strips those, then does a second pass to add back rhythm and opinion, since sterile writing without slop-words is still obviously machine-made. Before it hands the draft back, it asks itself what still reads as AI-generated and fixes that too.
  2. Detect. Same scan, no rewrite. It returns a bullet list: pattern name, the exact offending quote, a short fix. Useful when you want to know how bad a draft is before deciding whether it's worth editing at all.
<!-- diagram: flow-humanize-unslop -->
graph TD
    A[Draft comes in] --> B{Edit or detect?}
    B -- Edit --> C[Scan for banned words, punctuation, patterns]
    C --> D[Strip and replace with concrete language]
    D --> E[Add rhythm and opinion back in]
    E --> F["Self-audit: what still reads as AI?"]
    F --> G[Return edited draft + changelog]
    B -- Detect --> H[Scan only, no rewrite]
    H --> I["Return: pattern, quote, fix"]

Why this earns a place here

The rules are specific enough to be checkable rather than vibes. "No em dashes" is a rule you can grep for. "Faux-insight setups like 'what most people get wrong'" is a pattern you can name and catch on sight. That specificity is what makes it a skill instead of a one-line prompt. Reminding Claude that AI writing has tells doesn't do much on its own; enumerating them precisely enough that the same tics get caught the same way every time, on page one and on page ten, does.

This portfolio runs it on every blog post and every write-up in this section, including the one you're reading, before publishing.

How to get it

This skill is maintained publicly in Agentry Labs, a framework-agnostic collection of reusable AI-agent components. Install it into any project with one command:

curl -sSL https://raw.githubusercontent.com/ersanyamarya/agentry-labs/main/scripts/install.sh | bash -s -- skills/humanize-and-unslop

Then invoke it inside Claude Code:

claude -p "/skill humanize-and-unslop"

The skill is a single SKILL.md, no bundled scripts or eval fixtures, just the persona, the process, and the banned-pattern lists that do the actual work.