explaingit

shubhamv123/crisp

★ 17ShellAudience · vibe coderComplexity · 1/5ActiveSetup · easy

TLDR

An Agent Skill that makes AI coding assistants drop filler and reply more tersely, while keeping code, errors, and safety warnings intact.

Mindmap

mindmap
  root((crisp))
    Inputs
      Trigger phrase
      Skill install command
      Agent conversation
    Outputs
      Shorter replies
      Preserved code blocks
      Safety warnings intact
    Use Cases
      Cut filler in chats
      Save output tokens
      Toggle terse mode
    Tech Stack
      Markdown
      Agent Skills spec
      Shell installer

Things people build with this

USE CASE 1

Install crisp into a Claude or Codex agent with npx skills add shubhamV123/crisp to get shorter answers in code chats.

USE CASE 2

Paste the SKILL.md contents into a browser-based chat that does not support the skills installer.

USE CASE 3

Toggle terse mode mid-conversation with crisp mode and switch back with stop crisp.

USE CASE 4

Use the included benchmark numbers to justify enabling crisp on cost-sensitive Claude API workloads.

Tech stack

MarkdownShellAgent Skills

Getting it running

Difficulty · easy Time to first run · 5min

Needs an agent that supports the Agent Skills spec, or you can paste SKILL.md at the top of any chat as a fallback.

In plain English

crisp is a small instruction pack, called a skill, that you can install into an AI coding agent to make its replies shorter. The README's before-and-after example shows the typical effect: where the default reply might open with "Sure, I'd be happy to help, the issue you're seeing is probably caused by...", the crisp version says "Bug in auth middleware. Token expiry check uses < not <=." and then shows the fix. The point is to keep the same technical content but cut the filler around it. The author is careful to say what the skill does not do. crisp does not make the model more correct, it only changes the style. It also does not compress the parts of a reply where shortening could make the answer unsafe or ambiguous. For destructive operations, security warnings, and sequences where step order matters, the skill instructs the agent to switch to clear full-sentence prose, give the warning, and then return to concise output. Code blocks, exact error messages, commands, flags, and numbers are kept untouched. The skill follows the open Agent Skills specification, so installation is the same pattern as other skills in that ecosystem: npx skills add shubhamv123/crisp. For agents or browser interfaces that do not understand that command, the README suggests pasting the contents of the included SKILL.md file at the top of a conversation. Several trigger phrases turn the mode on, including crisp mode, go crisp, /crisp, be brief, and cut the fluff. Phrases like stop crisp, normal mode, or explain in detail turn it off. Once enabled the mode stays on for the rest of the conversation. The README includes a small benchmark, measured against Claude API output tokens averaged over three runs per prompt. With Haiku 4.5 the average output token count dropped by about 29 percent, with Opus 4.7 by about 61 percent, and with Sonnet 4.6 by about 70 percent. The author notes that input token usage does not change, that the benchmark measures length only and not correctness, and that risky-operation prompts are intentionally less compressed.

Copy-paste prompts

Prompt 1
Install shubhamV123/crisp into Claude Code and confirm it is active. Show the exact npx command and how I verify with a small test prompt.
Prompt 2
Read SKILL.md from shubhamV123/crisp and list every rule that prevents over-compression for destructive or security-sensitive operations.
Prompt 3
Run my own benchmark of shubhamV123/crisp on Sonnet 4.6 with 10 coding prompts. Provide a Python script that calls the Claude API and tallies output tokens.
Prompt 4
Compare shubhamV123/crisp against simply adding be terse to my system prompt. Which approach is more robust and why.
Prompt 5
Adapt shubhamV123/crisp into a Cursor rule file so the same terse style applies in Cursor. Show the .cursorrules content.
Open on GitHub → Explain another repo

Generated 2026-05-22 · Model: sonnet-4-6 · Verify against the repo before relying on details.