talk-normal

English | 中文

Make any LLM talk like a normal person. No filler, no fluff, just the answer.

Contributions welcome. See CONTRIBUTING.md and open a rule suggestion in Issues. Recent rule history in CHANGELOG.md.

What it does

A single system prompt that transforms verbose, corporate-sounding LLM output into direct, informative responses. Works with any model (GPT, Claude, Gemini, LLaMA, etc.). Tested on GPT-4o-mini (73% reduction) and GPT-5.4 (72% reduction) while preserving all useful information. See full comparison with all 10 test questions and complete answers.

Question: “What is Python?”

After (GPT-4o-mini, 513 chars):

Python is a high-level, interpreted programming language known for its readability and simplicity. It supports multiple programming paradigms, including procedural, object-oriented, and functional programming. Python has a large standard library and a vibrant ecosystem of third-party packages, making it suitable for web development, data analysis, artificial intelligence, scientific computing, and more. Its versatility and ease of use make it a popular choice among beginners and experienced developers alike.

Real-world example: financial market analysis

This is the actual case that motivated building talk-normal.

Usage

OpenClaw

Three ways, pick whichever fits your workflow:

Option 1: Paste the GitHub link into chat (easiest)

Paste this link into your OpenClaw chat and ask it to install:

https://github.com/hexiecs/talk-normal

Option 2: ClawHub

clawhub install talk-normal && bash skills/talk-normal/install.sh

To pull the latest rules later:

clawhub update talk-normal && bash skills/talk-normal/install.sh

Option 3: Manual git clone

git clone https://github.com/hexiecs/talk-normal.git && cd talk-normal && bash install.sh

All three paths end up running the same install.sh, which auto-detects your AGENTS.md and injects the prompt between # --- talk-normal BEGIN --- and # --- talk-normal END --- markers. The installer is idempotent: re-run it any time to pick up the latest rules without touching the rest of your AGENTS.md.

Uninstall:

bash install.sh --uninstall

Start a new conversation to take effect.

Hermes Agent

Two ways:

Option 1: Install from GitHub

hermes skills install --force hexiecs/talk-normal/skill-hermes

--force is required because this skill modifies your AGENTS.md to inject always-on prompt rules — Hermes’s security scanner flags that as persistent prompt modification. The skill is open-source; review it at skill-hermes/SKILL.md before installing.

Then run the installer (installs globally to ~/AGENTS.md):

cd ~ && bash ~/.hermes/skills/skill-hermes/install.sh

To install for a specific project instead, cd into that project directory before running the installer.

Option 2: Manual git clone

git clone https://github.com/hexiecs/talk-normal.git
cd ~ && bash talk-normal/install.sh

The installer auto-detects your workspace config file (.hermes.md, HERMES.md, or AGENTS.md) and injects the rules. Hermes freezes context files at session start, so start a new session to take effect.

Uninstall:

cd ~ && bash ~/.hermes/skills/skill-hermes/install.sh --uninstall

ChatGPT custom instructions

Copy the contents of prompt-chatgpt.md into ChatGPT’s custom instructions field.

ChatGPT’s custom instructions fields are capped at 1500 characters each. prompt-chatgpt.md is a compressed variant built to fit that limit while preserving every load-bearing rule from prompt.md (negation-frame ban, closing-stamp ban, filler list, conditional-menu ban, few-shot BAD/GOOD examples). Use prompt.md everywhere there is no length cap (OpenClaw, API, Cursor, Continue); use prompt-chatgpt.md only for the ChatGPT custom instructions field.

Any OpenAI API tool

Copy the contents of prompt.md into the system prompt field of whatever tool you use (Cursor, Continue, your own app, etc.)

API calls

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": ""},
      {"role": "user", "content": "What is Python?"}
    ]
  }'

Test results

10 prompts, temperature=0. Measured in characters. Full responses for every question.

GPT-4o-mini — average reduction: 73%

GPT-5.4 — average reduction: 72%

GPT-5.4 is already more concise than 4o-mini out of the box. talk-normal still cuts verbose responses by 20-89% on both models.

Rule iteration

Individual rules are iterated against real LLM output. Each rule that leaks in production gets a file in regressions/ tracking the leak count per version, the specific fix, and the observed LLM excerpts that motivated each round.

Example: the "不是X，而是Y" rhetorical frame went from 6 violations per response to 0 across four iterations on the same stress prompt. The load-bearing change turned out to be removing a specific negative example from the rule text — it was being copied verbatim by the model as a template instead of avoided as an anti-pattern. Full writeup: regressions/rule-17-negation-frame.md.

Star History

License

MIT