Prompt Refiner Skill
Turn a messy request into a self-contained prompt with a clear task, relevant context, constraints and an output format.
What this skill does
Goal diagnosis
Identify the intended task, audience and essential missing inputs.
Prompt construction
Build one prompt with the necessary context, constraints and output contract.
Refinement notes
Explain the edits and suggest concrete inputs that could improve a later revision.
What to provide & what you get
- Input
- A rough prompt or task description, intended audience, desired output and optional target model.
- Expected output
- A copyable revised prompt, a short explanation of the changes and questions for further refinement.
Installation
- Download and extract the ZIP. Keep each complete skill folder, including its supporting files. Do not copy only
SKILL.md. - Copy the skill folder into your agent's skills directory. Use the project path below for this project, or the personal path for all your projects.
Codex
- This project
.agents/skills/prompt-refiner/SKILL.md- All your projects
~/.agents/skills/prompt-refiner/SKILL.md
Claude Code
- This project
.claude/skills/prompt-refiner/SKILL.md- All your projects
~/.claude/skills/prompt-refiner/SKILL.md
- Check the final path. It must match one of the paths above, with one
prompt-refiner/folder directly underskills/. Avoid an extra ZIP-name folder around it. - Start a new task in your project. Use the matching example below to invoke the installed skill. If it is missing, recheck the folder name and reopen your agent.
Before the first run
- An agent that can load local skills
- A draft prompt or a concrete task and desired output
The ZIP installs the skill instructions and bundled files. Any listed software, accounts or external tools need their own setup.
What's in the download
3 files across 1 skill folder. Includes prompt-refiner/SKILL.md.
File integrity (SHA256):
a2ba2db0eb4715f39eef2da7b35fa93935e3e17a7931817857d4008b9c1be666How to use
- Provide the rough prompt and the result you actually need.
- Specify the target tool, audience and hard constraints when they matter.
- Copy the revised prompt into the target tool and compare the result with your original requirements.
Try it in your agent
Install first, then attach or provide the inputs described above. Replace any bracketed placeholders before running.
Codex
Use $prompt-refiner. Refine this prompt for [audience and target tool]. Preserve [constraints] and return one copyable prompt: [rough prompt].
Claude Code
Use /prompt-refiner. Refine this prompt for [audience and target tool]. Preserve [constraints] and return one copyable prompt: [rough prompt].
Check the result: Verify that the revised prompt preserves the original goal and constraints. Token savings and output quality need separate measurement.
Worked example · Editorial fixture
Turn a vague request into a bounded prompt
Fictional inputs and an editorially prepared output for this exact Skill package. No end-to-end agent or model run was performed; the Skill remains structure-checked.
Example input
Raw request: "Make a great meeting summary. Include who does what, keep it short, and fill the gaps so it sounds complete." Known constraints: at most 120 words; use only supplied notes; missing owners stay UNKNOWN; no email or external action. Task: improve the request for another assistant. Do not write the meeting summary.
Prepared output
🎯 Target AI & Mode Any modern LLM — source-bounded meeting summary. ⚡ Optimized Request Summarize the supplied meeting notes in at most 120 words. Include confirmed decisions, action items, explicit owners, and stated dates. Use UNKNOWN for missing owners or dates. Separate proposals from decisions. Do not invent facts or send anything. Before returning the summary, compare every action and owner with the notes. Notes: [paste permitted meeting notes] 📝 Notes Removed the request to fill gaps and made missing fields explicit. This is a prepared prompt; it has not been evaluated against a model. 🔍 Improvement Questions Which audience will read the summary? Should unresolved questions be a separate list?
What to check in this case
- The output is one reusable prompt, not a fabricated meeting summary.
- The 120-word limit, UNKNOWN rule, and no-send boundary survive the rewrite.
- Use an actual note packet and compare results before claiming any improvement in reliability.
References: Community source and author
Keep in mind
- The skill writes a prompt; it does not perform the underlying task.
- The source’s claims about performance and token efficiency were not benchmarked in this review.


