Prompt Refine skill
Skill silently rewrites your request for the model family that answers it, then answers without showing the rewrite.
- 54 ★ on the repository
- Last commit
- MIT
Minimal access
Why this level
- The skill quietly rewrites the request, so an answer can differ from a literal reading.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/prompt-refineDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skill quietly rewrites the request, so an answer can differ from a literal reading.
This is third-party code. Review the repository files before installing.
What it does
Prompt Refine checks which model is answering and applies that family's prompting strategy. Claude gets the Anthropic strategy, GPT the OpenAI one, and an unknown model a universal fallback. You just ask: the agent reshapes the request for itself, preserves your language and answers without showing the rewrite. It reads conversation context, so follow-ups inherit the goal, constraints and terminology, while the latest user instruction still wins. Only a short skill file and one selected strategy from `data/strategies.json` enter the context, with no dependencies, server or extra optimizer call. Compatibility is stated for Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot, Windsurf and CodeBuddy.
Who it is for. People who switch between Claude, GPT and other models and do not want to adjust request wording for each by hand.
Good fit when
- The same request goes to different models and comparable quality is expected.
- A vague phrasing must become a structured task without a dialogue.
- A long conversation needs corrections combined with the earlier goal.
- The request is in Russian and the answer must stay in Russian with no prompt tweaking.
Not a fit when
- You need to see and edit the rewritten prompt, since the skill shows only the answer.
- The host sends no signal about the current model, so only the universal strategy applies.
- The request is already tuned by hand and should not change.
Example request
Explain how best to organize frontend caching. First reshape my question for yourself.Limitations
The route is chosen from the host's signal about the current model. With no signal or an unknown model, the universal strategy applies rather than one tuned to the family. The rewrite is hidden from the user, so inspecting it means asking the skill to show it separately. The authors give their own evaluation on a set of 120 prompts, which does not replace testing on your tasks. The npm version and the repository install into the project skills folder, and updating means reinstalling.
How to disable. Delete the `.claude/skills/prompt-refine` folder or wherever you installed the skill. There are no leftover settings or background processes.
FAQ
How does the skill choose a strategy?
From an explicit signal about the current model: Anthropic for Claude, OpenAI for GPT. An unknown model gets the universal strategy.
Does it need an extra model call?
No, there is no extra optimizer call. A short skill file and one strategy enter the context.
Is the request language preserved?
Yes, the skill preserves the user language, so a Russian request stays Russian.
Which platforms are listed?
Claude Code, Cursor, OpenAI Codex, Gemini CLI, GitHub Copilot, Windsurf, CodeBuddy and others that support agent skills.
How do I see the rewritten request?
It is hidden by default. Ask the skill to show the rewritten version explicitly and compare it with the original.
Discussion
No questions or reviews yet.
Related
SkillsA library of prompts in one XML-tagged template, Claude Code skills for development and DevOps, and a set of hooks
Skill that reads your prompt history, finds the asks you repeat and turns each into a short command such as /sent
A catalog of 28 mental-model skills for agents: diagnosis, decisions, systems, risk, strategy and idea generation
The prompthub CLI manages prompts, skills, MCP servers and plugins and distributes one skill to Claude Code, Cursor and Codex.
A skill that turns a vague idea into a structured requirements list and asks precise questions until you confirm understanding
A library of 229 skills for work and life: sales, marketing, finance, engineering, design and parallel agent runs
README in short
The README explains routing by model, shows examples, built-in strategies, an evaluation on 120 prompts and a platform list. English and Chinese versions are available.