I'm over the wall of text.
“Great question! It’s not a human, it’s AI! Let’s delve in. Blah. Blah. Blah. A literary flourish here, a tidy list of three there, and a recap of everything you just read. No point anywhere. I hope this helps!”
Let’s remember who we’re writing for.
Give the reader the essentials in as few words as possible.
- Open with what the subject is and why it matters.
- Keep the key facts: who, what, when, where, and the result. Leave out examples, lists, minor names and figures, and background.
- Keep it short. For a summary of a long piece, aim for about 150 words.
- Plain statements, short sentences. No em dashes, no metaphors, no rhetorical questions, no "not only... but".
- Don't repeat yourself or wrap up. Never add anything the source doesn't say.
About
Stop with the slop. Get to the point.
LLMs work in two directions. Summarizing is lossy compression: a long input becomes a short one, and detail is lost. Expanding a short prompt into a long output inflates it with invented detail.
Slop Free is an experiment to find a prompt that minimizes loss in compression and reduces inflated detail. What is left is text that gets to the point.
-
1
Evaluate
Each week a new candidate prompt is tested against an accuracy and efficiency standard.
-
2
Test
The candidate is then tested against the golden prompt in the quiz. Humans test the readability of the output by answering 5 short quiz questions.
-
3
Rank
If the candidate prompt outperforms the golden prompt in readability, it replaces the old prompt.
-
4
Repeat
Indefinitely.
Why Wikipedia
It offers access to articles on every subject. Revisions are pinned, allowing for traceability.
Close but no cigar
Losing prompts are kept for posterity.
Think you can do better?
No tracking
Only answers and reading time are stored. No personal data. We don't want your data.
Models used
| Job | Model |
|---|---|
| Writes the summaries | gpt-6-luna |
| Judges which points were kept | gpt-6.2-sol |
| Drafts new candidate prompts | gpt-6.2-sol |
| Writes the quiz questions from the summaries | gpt-4.1-mini |
How to use it
Paste it or save it. Doesn't matter. Just use it.
Include the golden prompt in your agent and witness the magic. Less filler and more directness. Fewer walls of AI text.
For one message
Write your request, leave a blank line, paste the prompt.
Explain how a heat pump keeps a house warm at -20°C.
Give the reader the essentials in as few words as possible.
- Open with what the subject is and why it matters. ...
For every chat
Save it as a personal preference, a custom instruction in your agent, or embed it in your system prompt.
Where to save it
- ChatGPT
- Settings → Personalization → Custom instructions
- Claude
- Settings → Profile → personal preferences, or a Project's instructions
- Gemini
- Saved info, or a Gem's instructions
- API and agents
- The end of your system prompt. Always fetch the latest:
api.slopfree.ca/v1/golden(JSON) orslopfree.ca/golden.txt(plain text)
Tips
- Keep it last. Put task-specific instructions before it, so the golden prompt is the final thing the model reads.
- Re-paste in long chats. Models drift back to old habits as a conversation grows. Pasting it again resets them.
- Edit freely. Mix it, mash it. Do whatever you want. Please contribute any prompt that you think works better.
Quiz
Help train the golden prompt.
Answer just 5 short questions to contribute. Read each summary at your normal pace.
Guess round: no summary this time. Answer as best you can. This tells us how guessable the question is.
Difficulty
Human First Design
Crafting for humans, with AI
Design for the human first.
Most end users are people, and forcing them to sort through a wall of AI output is a bad experience.
Protect the negative space.
Leave things out so the real point stands out. Human attention is small, and AI output is big. Edit. Edit. Edit.
The burden is on the designer.
You, as the designer or writer, hold the context. It’s your responsibility to make the output understandable. Do not push the responsibility onto the reader to decode a confusing design or sift through a wall of text to find the point.
Make it obvious.
Each paragraph should make its point obvious, as should each next step in an application. A user or reader shouldn’t have to guess what’s going on, and it should be easy to tell when something is wrong. Errors should give enough information to correct the problem, or at minimum enough to support troubleshooting.
Peel the onion.
Show the essence first, then give the user a path to dig deeper. In writing, lead with the most important points and follow with the necessary details.
Respect their time.
Just because we can create at light speed doesn’t mean anyone has to consume it. It’s still the user’s or reader’s choice. You have to earn the right for someone to spend time reading your doc or using your system.
Never accept the first output.
Treat the first output, whether text or an application, as a draft. Edit it heavily using the principles above. Remember: put the human first.