Skip to main content
Lab Grimoire
TW EN
☕ Coffee
Spellbook / talk-normal ↗ Visit project
ai-tools

talk-normal

A single system prompt file that forces any LLM to respond directly and without filler. Bans all opener phrases and the 'not X, but Y' sentence pattern. Tested to reduce word count by 72–73% while retaining all substantive information. 1,375 stars in 13 days, MIT licensed, includes a Claude Code skill version.

Best For

Anyone who finds AI responses bloated with filler, hedging, or tonal artificiality. Specifically:

  • Engineers and writers who need LLM output that reads more like direct human prose
  • Developers who want output quality control in Claude, GPT, or Gemini system prompts
  • Claude Code users who want agent output with fewer hollow acknowledgment phrases
  • Anyone tuning an AI assistant's tone who is tired of "That's a great question"

If your main frustration with AI responses is "a lot of words, none of them necessary," talk-normal addresses this directly. If you need AI to maintain warm, empathetic tone for customer-facing contexts, these rules are the wrong fit.

How I Actually Use It

talk-normal formalizes something that should already be a principle: unnecessary words are a failure mode.

The repo centers on a prompt rule set targeting three specific noise patterns in LLM output:

  1. Filler openers are all banned: "I'd be happy to help," "Great question," "First, we need to understand," "In conclusion"
  2. The "not X, but Y" sentence pattern is prohibited in any language, any position. This structure is the single most identifiable marker of AI-toned prose, and removing it has an immediate effect.
  3. When asked for a recommendation, "both sides have merit" non-answers are prohibited. The rule requires a choice and a reason.

TEST_RESULTS.md provides empirical data: 72–73% word count reduction with human evaluation confirming all substantive information was retained. This is a prompt engineering project with testable results, not a style preference document.

1,375 stars in 13 days. The author also ships skill-hermes, a ready-to-install Claude Code skill version.

image

Where It Is Strong

  • Copy prompt.md content into any system prompt and it takes effect immediately. No installation.
  • The contrastive negation ban is the most valuable single rule. This sentence pattern is the most persistent source of AI tonal artificiality, and few other tools explicitly target it.
  • TEST_RESULTS.md provides 72-73% word count reduction data. Not subjective.
  • skill-hermes enables /talk-normal invocation with minimal integration cost
  • The rules apply to Claude, GPT-4, Gemini, LLaMA -- any model that takes a system prompt

Where It Fails

  • In contexts requiring empathy or emotional support, these rules produce responses that read as blunt rather than helpful.
  • Fewer words is not the same as better answers. A model that does not know the answer will now say "I don't know" more concisely.
  • The contrastive negation ban is absolute. Some comparative analysis genuinely requires the pattern; blanket prohibition may constrain expression in legitimate comparison tasks.
  • The repository is 13 days old. Whether the maintainer iterates this prompt over time remains to be seen.

Pricing, Difficulty, and Risk

Free and open source under MIT. Lowest possible difficulty — the core use case is pasting text into a system prompt field, no technical background required. The Claude Code skill version installs in one command. No privacy risk, no API costs, no licensing restrictions.

Verdict

A prompt engineering artifact with empirical backing and testable results.

The contrastive negation ban is the rule worth borrowing directly — it targets the most persistent marker of AI-toned prose. If your LLM responses consistently make you think "this could have been one sentence," this is worth three minutes to test.

Source