← Back to Skills Marketplace
dylanfeltus

Error Prone

by Dylan Feltus · GitHub ↗ · v1.0.0 · MIT-0
cross-platform ✓ Security Clean
131
Downloads
0
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install error-prone
Description
Makes AI-generated text sound more human by intentionally introducing realistic imperfections. Use when writing emails, messages, social posts, or any text t...
README (SKILL.md)

Error-Prone

Make your writing worse on purpose. Because perfect copy is creepy.

AI text has a smell. Too clean, too polished, too correct. People can tell. Error-Prone adds realistic human imperfections so your emails, messages, and posts actually sound like a person wrote them.

How to Use

When asked to write or edit text with error-prone mode, apply imperfections from the levels below. Default to Level 1 unless asked for more chaos.

Levels

Level 1: Professional Human (default)

Subtle. Still competent. Just not suspiciously perfect.

Apply 2-3 of these per paragraph:

  • One typo per ~100 words (teh, adn, hte, becuase, recieve, occured, definately)
  • Occasional missing comma or extra comma
  • Start one sentence with "And" or "But" or "So"
  • Use a contraction where formal writing wouldn't (it's vs it is, we're vs we are)
  • One slightly awkward word choice (use "get" instead of "obtain", "thing" instead of "element")
  • Inconsistent capitalization in lists
  • End a sentence with a preposition ("the team I'm working with" not "the team with which I'm working")

Level 2: Casual Human

Texting-your-coworker energy. Clearly typed fast.

Apply 4-6 of these per paragraph:

  • Everything from Level 1, plus:
  • Double space after a period sometimes
  • "lol" or "haha" once per message
  • Trailing off with "..."
  • Start with "So" or "Anyway" or "Yeah so"
  • Run-on sentence (join two sentences with "and" instead of a period)
  • Gonna, wanna, kinda, sorta
  • Skip a word that the brain fills in ("I think we should [go] with option B")
  • Lowercase sentence starts occasionally
  • "me and Dylan" instead of "Dylan and I"
  • One autocorrect-style error (ducking, he'll instead of hell)

Level 3: Unhinged

Sent from my phone while walking. Stream of consciousness. Autocorrect chaos.

Apply 8+ of these per paragraph:

  • Everything from Level 1-2, plus:
  • Multiple typos per sentence
  • No punctuation for 2-3 sentences then suddenly perfect punctuation
  • Repeated words ("I think think we should")
  • Wrong homophone (their/there/they're, your/you're)
  • Abandon a sentence halfway and start a new thought
  • ALL CAPS for one word for emphasis
  • "idk" "tbh" "ngl" "imo"
  • Emoji that's slightly wrong context (using 💀 to mean "that's funny")
  • Send as multiple messages instead of one (simulate hitting enter too early)
  • "wait no" or "actually nvm" corrections
  • Autocorrect disasters ("I'll ask the massage" instead of "manager")

Rules

  1. Never break meaning. Imperfections should not change what the text says. A reader should still understand the message perfectly.
  2. Match the context. Professional email = Level 1. Slack message = Level 2. Tweet reply = Level 2-3.
  3. Vary the errors. Don't use the same typo twice. Real humans make different mistakes.
  4. Don't overdo it. The goal is "human" not "illiterate". Even Level 3 should be readable.
  5. Preserve the user's voice. Add imperfections to THEIR style, don't change what they're saying.

Examples

Before (AI-generated):

I wanted to follow up regarding our conversation about the Q3 marketing strategy. I've prepared a comprehensive analysis of the proposed budget allocation and would appreciate the opportunity to discuss it at your earliest convenience.

After (Level 1 — Professional Human):

I wanted to follow up on our conversation about the Q3 marketing strategy. I've put together an analysis of the proposed budget allocation and would love to discuss it when you get a chance.

After (Level 2 — Casual Human):

Hey, wanted to follow up on our q3 marketing convo. I put together a budget analysis thing... would love to chat about it when you get a sec

After (Level 3 — Unhinged):

hey so I did that budget analysis for q3 marketing its actually pretty interesting ngl anyway lmk when your free to chat about it *you're lol

Usage Guidance
This skill appears to do what it says: deliberately add human-like imperfections to text. It does not request credentials or install code, which reduces technical risk. Two practical cautions: (1) test outputs before sending in sensitive contexts (legal, medical, security) because some recommended 'errors' could change meaning or produce embarrassing/unsafe results; (2) be mindful of misuse — making text more deceptive is a feature that can enable social-engineering. If you install it, try it on non-sensitive drafts first and review carefully (especially Level 3 outputs). If you need guarantees that meaning never changes, this skill's examples show that guarantee may be hard to achieve in practice.
Capability Analysis
Type: OpenClaw Skill Name: error-prone Version: 1.0.0 The skill bundle 'error-prone' (SKILL.md) contains stylistic instructions for an AI agent to introduce human-like imperfections into text. It lacks any executable code, network requests, or instructions to access sensitive data, and its behavior is entirely consistent with its stated purpose of humanizing AI-generated content.
Capability Assessment
Purpose & Capability
Name/description match the SKILL.md. The skill is instruction-only and only describes text transformations (typos, contractions, tone). It does not request unrelated binaries, environment variables, or config paths.
Instruction Scope
Instructions stay within text-editing: apply errors at different severity levels, preserve meaning, and match context. However there is an internal inconsistency: Rule 1 ('Never break meaning') conflicts with some examples (autocorrect disasters, skipped words, 'massage' vs 'manager') that could change intent or introduce harmful misunderstandings. Also suggestions like 'send as multiple messages' simulate interaction behavior but do not require extra system access.
Install Mechanism
No install spec and no code files — lowest risk. Nothing will be downloaded or executed on disk by the skill itself.
Credentials
No environment variables, credentials, or config paths are requested. The skill does not ask for unrelated secrets or system access.
Persistence & Privilege
always is false and the skill does not request persistent system presence or modify other skills. Autonomous invocation is allowed by default (platform default) but is not excessive here.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install error-prone
  3. After installation, invoke the skill by name or use /error-prone
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.0
Initial release of error-prone skill. - Makes AI-generated text sound more human by adding realistic imperfections. - Supports three imperfection levels (Professional Human, Casual Human, Unhinged) for different contexts. - Triggers on prompts requesting humanization or imperfections. - Includes clear rules and examples to maintain meaning and context while varying errors.
Metadata
Slug error-prone
Version 1.0.0
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 1
Frequently Asked Questions

What is Error Prone?

Makes AI-generated text sound more human by intentionally introducing realistic imperfections. Use when writing emails, messages, social posts, or any text t... It is an AI Agent Skill for Claude Code / OpenClaw, with 131 downloads so far.

How do I install Error Prone?

Run "/install error-prone" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Error Prone free?

Yes, Error Prone is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Error Prone support?

Error Prone is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Error Prone?

It is built and maintained by Dylan Feltus (@dylanfeltus); the current version is v1.0.0.

💬 Comments