CRISP

LLM Writing Protocol

CRISP (Concise, Relevant, Intuitive, Simple, Protocol) is a writing protocol for humans and LLMs, packaged as an agent skill. It steers models to say the useful thing once, as clearly as possible, then stop. Use it as a system prompt, a slash command, or a verb: /crisp.

Location

Cluj-Napoca

Oct 2026 - Present

Industry

AI Developer Tools

Visit Website

Challenge

LLM answers often open with flattery, hedge every claim, restate the question and close with “let me know if you have any questions.” Prompts that ask for short answers tend to swing the other way, dropping facts and turning cryptic.

CRISP had to cut the noise without losing precision, and be checkable rather than a vibe: something a person or a model could apply the same way every time.

Results

• 47% fewer tokens than the baseline, while keeping 97% of the required facts.

• 21 of 30 blind-judged pairings won against no prompt at all.

• 30% fewer tokens than ASD-STE100 Simplified Technical English, with the same facts.

• Ten explicit rules, each with a yes/no check, from “answer first” to “stop.”

• Three depth levels: /crisp 1, 2 and 3 change how much is said, never how well.

• Easy install: a plugin or uploaded skill in the Claude app with no terminal, Claude Code, or any repo’s AGENTS.md and CLAUDE.md through install.py.

Stack

Conclusion

CRISP shows that shorter answers can also be better ones. In its own example, a 43-word reply becomes 16 words with nothing lost.

The full protocol, research and benchmark are open on GitHub, with a three-tab dashboard and scripts to rerun the results.

CRISP shows that shorter answers can also be better ones. In its own example, a 43-word reply becomes 16 words with nothing lost.

The full protocol, research and benchmark are open on GitHub, with a three-tab dashboard and scripts to rerun the results.