CRISP

Say the useful thing once, as clearly as possible, then stop.

CRISP (Concise · Relevant · Intuitive · Simple · Protocol) is a writing protocol for humans and LLMs. This plugin adds it to Claude: one skill, /crisp, that makes replies read like one competent engineer talking to another instead of a chatbot.

What it does

Once added, Claude applies ten checkable rules to its answers whenever you say /crisp, "use CRISP", "crispify this", or "make this crispier", and whenever a project's CLAUDE.md or AGENTS.md asks for CRISP:

1. Answer first. The first sentence is the answer, decision, result, or ask.
2. Say it once. No preview, recap, restated question, or restated context.
3. Keep what changes action. Every sentence changes what you know, do, or decide.
4. Plain words, same words. Common verbs; one term per concept; no intensifiers without a measurement.
5. Name the actor, state the condition. "If X, do Y." "X failed because Y."
6. Replace vague with checkable. Soon, appropriate, usually become a number, a name, a condition, or "unknown".
7. Structure is earned. Prose by default; lists, tables, headings only when they aid scanning.
8. Short is not cryptic. Code, corrections, risks, and real uncertainty stay.
9. Sound like a colleague. No flattery, ceremony, apology, or ritual hedges.
10. Stop. No closing summary, no "let me know".

Three depths: /crisp 1 (just the answer), /crisp 2 (default), /crisp 3 (plus rationale, alternatives, risks). Levels change how much is said, never how well.

Example

Prompt: "My React component re-renders forever when I fetch data in useEffect. What's wrong?"

Without CRISP, a typical answer opens with "Great question!", lists every possible cause, adds unrelated advice about cleanup functions, and closes with "I hope this helps!" (154 words).

With CRISP (80 words):

Most likely your effect has no dependency array, so it runs after every render, and its setState triggers the next render.

1. Check the useEffect call. If it has no second argument, add one: useEffect(() => {...}, []).
2. If it already has an array and still loops, look for an object or function in it that is created inline in the component body. It gets a new identity on every render.
3. Wrap that value in useMemo or useCallback.

Measured

Blind pairwise judging against the same model with no instructions, 10 prompts × 3 samples (Claude Sonnet 5.5, judged by Claude Opus 5.5): 47% fewer tokens, 21 of 30 pairings preferred, 97% of required facts kept. Rewriting an existing answer with CRISP ("crispify this") scored 27 of 30. Full method, raw data, and a second lane on GPT-6: andreiverdes.github.io/crisp/benchmark.

For projects

skills/crisp/scripts/install.py writes a CRISP block into a project's AGENTS.md and CLAUDE.md so every agent in that repo replies in CRISP, with or without this plugin. Run it from the project root; --level crisp|minimal|full picks the prompt size, --command also creates a /crisp slash command, --check and --remove do what they say. The script only touches those files in the current directory and reads nothing else.

What this plugin runs, sends, or fetches

Nothing. It is one SKILL.md with instructions, four prompt files, two reference files, and the optional install.py above. No hooks, no MCP servers, no network calls, no credentials.

Source

The protocol, the research behind each rule, and the benchmark harness live at github.com/andreiverdes/crisp. This repository is the plugin bundle built from it, published by HorizonLoop. Version 1.0.0, MIT.