talk-normal-llm-prompt

name	talk-normal-llm-prompt
description	System prompt toolkit that removes AI slop and makes any LLM respond like a normal person — concise, direct, no filler.
triggers	["make the LLM stop being verbose","remove AI slop from responses","use talk-normal system prompt","make ChatGPT talk normally","reduce LLM output verbosity","install talk-normal prompt","strip filler from AI responses","configure LLM to be concise"]

talk-normal

Skill by ara.so — Daily 2026 Skills collection.

talk-normal is a system prompt (plus a shell-script helper) that strips AI slop — bullet-point padding, hollow affirmations, corporate filler — from any LLM while preserving all useful information. Tested at ~73% character reduction on GPT-4o-mini and GPT-5.4 with no information loss.

---

How it works

The project is a single prompt.md file (the system prompt) plus optional shell helpers. You copy the prompt text into the "System" field of any LLM interface or API call.

repo layout
├── prompt.md          ← the system prompt (main artifact)
├── CHANGELOG.md       ← rule history
├── CONTRIBUTING.md    ← how to add rules
└── TEST_RESULTS.md    ← before/after comparisons

---

Installation

1 — Clone the repo

git clone https://github.com/hexiecs/talk-normal.git
cd talk-normal

2 — Read the prompt

cat prompt.md

3 — Copy into your tool

Paste the contents of prompt.md into:

• ChatGPT → Settings → Customize ChatGPT → Custom Instructions → "How should ChatGPT respond?"
• Claude.ai → Project Instructions
• Cursor / Windsurf → .cursorrules or global AI rules
• API calls → system parameter (see examples below)

---

Using the prompt via API

OpenAI (Python)

import os
from pathlib import Path
from openai import OpenAI

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

system_prompt = Path("prompt.md").read_text()

response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[
        {"role": "system", "content": system_prompt},
        {"role": "user",   "content": "What is Python?"},
    ],
)
print(response.choices[0].message.content)

OpenAI (curl)

SYSTEM=$(cat prompt.md | jq -Rs .)

curl https://api.openai.com/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{
    \"model\": \"gpt-4o-mini\",
    \"messages\": [
      {\"role\": \"system\", \"content\": $SYSTEM},
      {\"role\": \"user\",   \"content\": \"What is Python?\"}
    ]
  }"

Anthropic Claude (Python)

import os
from pathlib import Path
import anthropic

client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])

system_prompt = Path("prompt.md").read_text()

message = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=1024,
    system=system_prompt,
    messages=[{"role": "user", "content": "Explain Docker in one paragraph."}],
)
print(message.content[0].text)

Google Gemini (Python)

import os
from pathlib import Path
import google.generativeai as genai

genai.configure(api_key=os.environ["GEMINI_API_KEY"])

system_prompt = Path("prompt.md").read_text()

model = genai.GenerativeModel(
    model_name="gemini-1.5-flash",
    system_instruction=system_prompt,
)

response = model.generate_content("What is a neural network?")
print(response.text)

Ollama (local models)

SYSTEM=$(cat prompt.md)

ollama run llama3 \
  --system "$SYSTEM" \
  "What is a REST API?"

Or via the Ollama Python SDK:

import subprocess, json
from pathlib import Path

system_prompt = Path("prompt.md").read_text()

result = subprocess.run(
    ["ollama", "run", "llama3"],
    input=f"SYSTEM: {system_prompt}\nUSER: What is a REST API?",
    capture_output=True, text=True,
)
print(result.stdout)

---

Shell helper: one-liner wrapper

A reusable shell function that injects the prompt automatically:

# Add to ~/.bashrc or ~/.zshrc
export TALK_NORMAL_PROMPT="$HOME/talk-normal/prompt.md"

asknormal() {
  local question="$*"
  local system
  system=$(cat "$TALK_NORMAL_PROMPT")

  curl -s https://api.openai.com/v1/chat/completions \
    -H "Authorization: Bearer $OPENAI_API_KEY" \
    -H "Content-Type: application/json" \
    -d "$(jq -n \
      --arg sys "$system" \
      --arg q   "$question" \
      '{model:"gpt-4o-mini",messages:[{role:"system",content:$sys},{role:"user",content:$q}]}'
    )" | jq -r '.choices[0].message.content'
}

Usage:

source ~/.bashrc
asknormal "What is the CAP theorem?"

---

Embedding in a project's AI config

Cursor (.cursorrules)

# Prepend talk-normal to your existing rules
cat talk-normal/prompt.md > .cursorrules
echo "" >> .cursorrules
echo "# Project-specific rules below" >> .cursorrules
cat your-existing-rules.md >> .cursorrules

OpenAI Assistants API

import os
from pathlib import Path
from openai import OpenAI

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
system_prompt = Path("talk-normal/prompt.md").read_text()

assistant = client.beta.assistants.create(
    name="Normal Assistant",
    instructions=system_prompt,
    model="gpt-4o-mini",
)
print(f"Assistant ID: {assistant.id}")

---

Combining with your own system prompt

talk-normal rules are additive — prepend them before your domain instructions:

from pathlib import Path

talk_normal = Path("talk-normal/prompt.md").read_text()

your_rules = """
You are a senior backend engineer. Answer questions about Python, Go, and distributed systems.
"""

combined_system = f"{talk_normal}\n\n---\n\n{your_rules}"

---

Common patterns

Pattern 1: Measure verbosity reduction

def verbosity_ratio(before: str, after: str) -> float:
    """Returns fraction of original length kept (lower = more concise)."""
    return len(after) / len(before)

before = "Python is a high-level, interpreted programming language known for its readability..."  # 1583 chars
after  = "Python is a high-level, interpreted language known for readability..."                  #  513 chars
print(f"{verbosity_ratio(before, after):.0%} of original length")  # → 32%

Pattern 2: A/B test with and without the prompt

import os
from pathlib import Path
from openai import OpenAI

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
system_prompt = Path("talk-normal/prompt.md").read_text()

question = "What is Kubernetes?"

def ask(system: str | None, user: str) -> str:
    messages = []
    if system:
        messages.append({"role": "system", "content": system})
    messages.append({"role": "user", "content": user})
    resp = client.chat.completions.create(model="gpt-4o-mini", messages=messages)
    return resp.choices[0].message.content

without = ask(None, question)
with_prompt = ask(system_prompt, question)

print(f"Without: {len(without)} chars")
print(f"With:    {len(with_prompt)} chars")
print(f"Reduction: {(1 - len(with_prompt)/len(without)):.0%}")

Pattern 3: Keep the prompt up to date

# Pull latest rules from upstream
cd talk-normal
git pull origin main

# Check what changed
git log --oneline -10
cat CHANGELOG.md | head -50

---

Contributing a new rule

1. Fork the repo and create a branch: git checkout -b rule/no-em-dashes
2. Edit prompt.md — add your rule in plain imperative English
3. Add an entry to CHANGELOG.md
4. Open an Issue or PR describing: what slop the rule targets, a before/after example

# Quick before/after test for your new rule
SYSTEM=$(cat prompt.md)
echo "Test question" | asknormal   # uses your modified prompt

---

Troubleshooting

Symptom	Fix
Model still uses bullet points	Ensure the system prompt is in the system role, not prepended to user
Prompt too long for context window	Use a smaller model or trim older messages; prompt.md is intentionally compact
Ollama ignores system prompt	Some quantized models have weak instruction-following; try mistral or llama3
Rules conflict with your own system prompt	Put talk-normal rules first; add # Override: comment before conflicting rules
Response is too terse / lost information	The prompt reduces filler, not facts — file an issue with a reproduction case

---

Key facts for agents

• Primary artifact: prompt.md — copy its text verbatim as the system message
• No code to run: this is a prompt, not a library; no pip install, no build step
• Model-agnostic: works with GPT, Claude, Gemini, LLaMA, Mistral, etc.
• Tested reduction: ~72–73% character reduction, zero information loss on 10-question benchmark
• License: MIT — use freely in commercial products