convez.ai
Skill
chendren

crisp

chendren/crisp

Ver no GitHub
InicialPluginScore · de 0 a 100

Ultra-clear prompts for non-native English speakers and vibe coders. Hook + /crisp skill + drop-in steering for Grok Build, Claude Code, and similar agents.

Instalação

$npx convez-ai chendren/crisp

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 4 meses
Versão
v0.1.0
Issues abertas
1
Contribuidores
2

README

A tradução para o português deste README ainda está na fila.

crisp

Ultra-clear prompts for non-native English speakers and vibe coders.

GitHub stars License: MIT

One lightweight, always-on hook + one powerful /crisp skill + one drop-in steering file.

  • Frugal: deterministic core does 80-90% of the work with zero model tokens and <150ms latency.
  • Fast: hook never blocks the UI meaningfully.
  • Extremely accurate & valuable: bakes in the exact prompt craft that makes Grok Build, Claude Code, Kiro, and similar agents ship complete, correct work on the first serious pass — even when the human input is casual, translated, or underspecified.

No regex in the implementation (string methods + simple scanners only). Respects and amplifies your existing CLAUDE.md / AGENTS.md / "vibe coding" rules.

GitHub: https://github.com/chendren/crisp

Install (Grok Build / Claude Code)

# From GitHub (recommended)
grok plugin install chendren/crisp --trust

# Or during local development
grok plugin install /absolute/path/to/crisp --trust

Then trust the plugin in the hooks/plugins modal (/plugins or Ctrl+L).

The UserPromptSubmit hook activates immediately for auto-enhancement.

The UserPromptSubmit hook activates immediately for auto-enhancement.

Usage

Automatic (hook)

Just type normally. When crisp detects vibe language, short requests, or non-native patterns, it injects a systemMessage with the enhanced version at the very start of the turn. The agent sees the crisp intent first.

You will see the enhancement in the scrollback (annotated as crisp).

Toggle with config if you want it silent or off.

On-demand (best quality)

/crisp make a nice dashboard with filters and row detail for the customer data
/crisp --profile=vibe the login flow should feel premium
/crisp i want that you create the api that return only active vendors with total value

The skill produces a full structured prompt (and can continue straight into implementation if the context is "build this").

Drop-in steering (works everywhere)

Copy steering/CRISP.md into any project's CLAUDE.md, AGENTS.md, .cursor/rules, or equivalent. Gives passive improvement even without the plugin.

Why this exists

Non-native speakers and fast "vibe" thinkers lose huge amounts of leverage with agentic coding tools. The raw prompt has low signal. The model spends early turns doing disambiguation that a better prompt would have made unnecessary.

crisp normalizes at the exact right moment (submit) or on explicit request, using both fast local logic and the full power of your preferred model + your personal steering.

Result: fewer turns, higher quality first outputs, and the non-native user sees the "correct English shape" of a great prompt — learning happens as a side effect.

See real transformations in examples/before-after.md.

Configuration

See crisp.config.example.json. Place a crisp.config.json in your project or user plugin data dir.

Key toggles:

  • autoEnhance — control the UserPromptSubmit hook
  • defaultProfile
  • includePersonalStandards — pull extra rules from your global Claude.md when enhancing

Development & Verification

After any change:

  1. node src/rewriter.mjs (or import it) on sample strings — must produce clean output, zero regex in the module.
  2. Test the hook script with a small JSON stdin blob containing user_prompt.
  3. Install locally: grok plugin install . --trust
  4. Submit real messy prompts. Check latency in the hooks modal and that the agent acts on the enhanced version.
  5. /crisp on a non-trivial vibe request → agent ships complete correct work with no follow-up questions.
  6. Grep the entire tree for actual regex usage in code (comments explaining "no regex" are fine).
  7. Drop steering/CRISP.md into a temp project and observe improved baseline behavior.

See examples/before-after.md for real examples and the living corpus of transformations.

Contributing & Feedback

  • Open issues with real (sanitized) messy prompts + what crisp produced.
  • PRs for new profiles, better detection, more languages, or steering improvements welcome.
  • For non-GitHub feedback, reply on any launch post or LinkedIn.

Philosophy (matches high-signal users)

  • Treat "vibe", "make it nice", short descriptions as full specs.
  • Build the complete thing. No stubs.
  • Plan mode for anything >3 files or architectural.
  • Short functions. Functional utilities. No regex. Validate at boundaries.
  • The enhancement itself must be high-signal and low-token-overhead.

Works everywhere (even without the plugin)

Copy steering/CRISP.md into any project's CLAUDE.md, AGENTS.md, .cursor/rules, Windsurf rules, or equivalent. Passive improvement for Cursor, Cline, Continue, and other agents.

License

MIT

Status

v0.1.0 — public.

Core deterministic rewriter + hook + skill + steering are solid.

Feedback wanted: especially real messy prompts from non-native English speakers and "vibe" describers. Drop them in GitHub Issues or reply with before/after results.

Star the repo if crisp helps you ship faster with fewer turns.

Ver o original no GitHub

Bom para

Relacionados