Prompt Fixer
Soumya7681/prompt-fixer
Limpa e expande prompts brutos automaticamente antes da execução
Ver no GitHubSkill que limpa e melhora prompts brutos antes da execução: corrige typos, expande instruções vagas em pedidos claros, repara placeholders {{template}} quebrados e remove preâmbulos desnecessários como "pode/por favor". Funciona através da própria inteligência do Claude, sem dependências externas além de jq, e pode ser acionado manualmente ou automaticamente.
Instalação
npx convez-ai Soumya7681/prompt-fixer1 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
Casos de uso
- Melhorar prompts mal formados antes de processar
- Expandir instruções vagas em pedidos estruturados
- Reparar placeholders de template incompletos
- Remover ruído de filler language em requisições
Veredito

Ferramenta prática para quem trabalha com prompts no Claude Code ou Antigravity. A abordagem não-intrusiva (sugestão em vez de reescrita silenciosa) respeita a agência do usuário. Implementação leve que degrada gracefully se jq não estiver disponível.
README
A tradução para o português deste README ainda está na fila.
prompt-fixer (Claude Code & Antigravity plugin)
Cleans up rough prompts before the assistant acts on them: fixes typos, expands terse
asks into clear instructions, repairs broken {{template}} placeholders, and
strips filler openers like "can you / please".
Unlike the original CLI, all the actual fixing is done by the AI's own judgment — there's no Node engine to install. The plugin just gives the assistant the methodology and knows when to apply it.
What's inside
| Component | What it does |
|---|---|
Skill prompt-fixer | The fixing methodology. Auto-invoked when a prompt looks rough. |
Command /fix-prompt <prompt> | On-demand. Shows ORIGINAL → FIXED → list of changes, then asks before acting. |
Hook UserPromptSubmit | A lightweight shell pre-filter. Fires only when a prompt looks rough (typos, filler opener, vague one/two-word ask, or unclosed {{) and nudges the assistant to apply the skill. Never blocks or rewrites your prompt. |
Install
For Claude Code
From a Claude Code session, add this repo as a marketplace and install:
/plugin marketplace add Soumya7681/prompt-fixer
/plugin install prompt-fixer@prompt-fixer-marketplace
Installing from a local clone instead? Point
marketplace addat the checkout directory:/plugin marketplace add /path/to/prompt-fixer.
For Antigravity (agy)
Run the following from your terminal:
git clone https://github.com/Soumya7681/prompt-fixer.git
cd prompt-fixer
agy plugin install .
Verification
Then check it loaded:
/reload-plugins # to reload the plugins (Claude Code only)
/help # /fix-prompt should appear
/plugin # prompt-fixer should be listed/enabled
Usage
# On demand — preview the fix, confirm before acting
/fix-prompt can you sumamrize this {{file
# Auto — just type a rough prompt; the hook nudges Claude to clean it up first
genrate tests
How "auto" mode works (and its limit)
A UserPromptSubmit hook is a shell command. It cannot silently
rewrite the text in your input box — it can only add context or block. So this
plugin's hook detects rough prompts and injects a one-line instruction asking
the assistant to apply the prompt-fixer skill before answering. The assistant does the
rewriting in its reasoning and acts on the cleaned version (noting the
interpretation when it meaningfully changed your ask).
This is intentionally non-intrusive: well-formed prompts trigger nothing, and the hook never discards or alters what you typed.
Requirements
jq(used by the hook to parse the prompt payload). Ifjqis missing, the hook degrades silently and your prompts are unaffected — you just lose the auto-nudge. The command and skill still work.
Tuning
- Detection rules (which prompts count as "rough"): edit
hooks/prompt-check.sh. - Fixing behavior (typo list, expansions, guardrails): edit
skills/prompt-fixer/SKILL.md. - Turn off auto mode entirely: disable the hook, or remove
hooks/hooks.json.
License
MIT