An AI skill that turns a raw draft into a production-grade prompt.
It rewrites and optimizes prompts using proven techniques, then hands you a ready-to-paste result plus a short changelog of what changed and why — calibrated to your task type and target model.
Works with Claude Code · GitHub Copilot · OpenAI Codex · Gemini CLI · OpenCode · Cursor · Windsurf · Cline
Table of contents
- Why this skill
- How it works
- Example
- Installation
- Verify installation
- Usage
- What's inside
- Updating
- Uninstall
- Contributing
- Project structure
- License
Why this skill
Most prompt advice is a handful of generic tips. This skill is different in three ways:
- Grounded in primary sources. Every technique is traceable to a cited source — Anthropic's prompting best practices and prompt improver, OpenAI's GPT-5/reasoning guides, Google's Gemini PTCF, and The Prompt Report's academic taxonomy of 58 techniques. See
sources.md.
- Model-aware. The rules differ by model family. The skill adapts to Claude 4.x, OpenAI GPT/reasoning (o-series), and Gemini — e.g. it removes "think step by step" for reasoning models, which actively hurts them.
- Anti-overengineering. It calibrates effort to complexity. A simple prompt gets a lean rewrite; it won't bloat it with chain-of-thought and examples it doesn't need.
It improves the prompt — it does not run it, unless you explicitly ask.
How it works
You paste a draft (or describe what you want). The skill runs a five-phase workflow:
- Diagnose — real intent, task type, target model, audience, output format, constraints, and the draft's weaknesses.
- Select techniques — only those that add value for this task and model.
- Rewrite — assemble the prompt in a canonical structure (role, context+motivation, XML-tagged data, positive instructions, few-shot, output contract).
- Self-review — run a quality checklist; the golden test is "could a colleague with no context run this without doubt?"
- Deliver — the optimized prompt + a changelog mapping each change to the weakness it fixes.
Example
classify the sentiment
of this comment:
{{comment}}
You are a sentiment analyst
specialized in customer feedback.
Classify the sentiment of the comment
below into one of these categories:
positive, negative, neutral.
<comment>
{{comment}}
</comment>
Follow these steps:
1. Identify expressions that carry emotion.
2. Weigh the overall tone, accounting
for irony and negation.
3. Choose the predominant sentiment.
Put your reasoning in <analysis>. On the
last line, answer ONLY one word:
positive, negative, or neutral.
What the skill changed, and why: added a role (calibrates interpretation), a strict enum output contract (ONLY one word — parseable), a short chain-of-thought (catches irony/negation, the top error source in sentiment), and XML around the input (robust to comments with line breaks). Every rewrite ships with this changelog.
See six more cases — coding, extraction, research, agentic, long-context, and an already-good prompt — in worked-examples.md.
Installation
Pick your tool. Claude Code is the primary target; the rest are fully supported.
One-line installer (copies the skill into ~/.claude/skills/, backs up any existing install, touches nothing else):
curl -fsSL https://raw.githubusercontent.com/PhAlves23/prompt-engineering-skill/main/install.sh | bash
Or as a plugin (versioned, updatable):
/plugin marketplace add PhAlves23/prompt-engineering-skill
/plugin install prompt-engineering
Restart Claude Code after installing.
The plugin marketplace works in Copilot CLI too:
copilot plugin marketplace add PhAlves23/prompt-engineering-skill
copilot plugin install prompt-engineering@prompt-engineering-marketplace
Codex discovers the full skill from ~/.agents/skills/. Full guide: docs/INSTALL.codex.md.
git clone https://github.com/PhAlves23/prompt-engineering-skill.git ~/.codex/prompt-engineering-skill
mkdir -p ~/.agents/skills
ln -s ~/.codex/prompt-engineering-skill/plugins/prompt-engineering/skills/prompt-engineering ~/.agents/skills/prompt-engineering
gemini extensions install https://github.com/PhAlves23/prompt-engineering-skill
Reads AGENTS.md natively. Quickest path (full guide: docs/INSTALL.opencode.md):
curl -fsSL https://raw.githubusercontent.com/PhAlves23/prompt-engineering-skill/main/adapters/codex/AGENTS.md -o AGENTS.md
Copy the rule for your tool from adapters/:
- Cursor →
adapters/cursor/.cursor/rules/prompt-engineering.mdc into your project's .cursor/rules/ (invoke with @prompt-engineering). A native Cursor plugin manifest is also provided at .cursor-plugin/plugin.json.
- Windsurf →
adapters/windsurf/.windsurf/rules/prompt-engineering.md into .windsurf/rules/ (activates by intent).
- Cline →
adapters/cline/.clinerules/prompt-engineering.md into .clinerules/prompt-engineering.md (workspace rule; ask explicitly).
See adapters/README.md for details and other tools (Roo, Continue, Zed).
Verify installation
Start a new session and ask for something that should trigger the skill — for example, paste a draft with "improve this prompt", or run /prompt-engineering in Claude Code. You should get an optimized prompt + a changelog.
# Claude Code / curl install:
test -f ~/.claude/skills/prompt-engineering/SKILL.md && echo "installed" || echo "not found"
Usage
Type /prompt-engineering (Claude Code / Copilot), or simply say:
- "improve this prompt"
- "optimize this prompt"
- "rewrite this prompt"
- or paste a draft and ask to make it better.
The skill activates automatically (where the tool supports intent-based activation) and returns the rewritten prompt plus a changelog.
What's inside
| Component | What it gives you |
|---|
| Canonical structure | Role, context+motivation, sequential instructions, XML, few-shot, CoT, output contract, success criteria |
| Technique selection | By task type (classification, extraction, generation, coding, reasoning, research, agentic, summarization) and model |
| Technique index | All 58 techniques from The Prompt Report plus vendor/post-2024 extras |
| Worked examples | 7 full draft → optimized → changelog cases |
| Quality checklist | Pre-delivery self-review |
| Evaluation guide | How to A/B test that the rewrite is actually better |
| Prompt security | Defenses against injection/jailbreak |
| Auto-optimization | APE, OPRO, DSPy, MIPRO references for when you have an eval set |
Updating
| Install method | Update command |
|---|
| curl installer | re-run the installer |
| Claude / Copilot plugin | /plugin update prompt-engineering |
| Codex symlink | cd ~/.codex/prompt-engineering-skill && git pull |
| Gemini extension | gemini extensions update prompt-engineering |
Uninstall
- curl / manual:
rm -rf ~/.claude/skills/prompt-engineering
- Claude plugin:
/plugin uninstall prompt-engineering
- Codex:
rm ~/.agents/skills/prompt-engineering
Contributing
Issues and pull requests are welcome. Start with CONTRIBUTING.md — it covers the dev setup, how the adapters are generated from a single source, and the review process. Please also read the Code of Conduct. Security reports go through SECURITY.md.
The skill stays grounded in cited sources: if you change a technique claim, cite it.
Project structure
prompt-engineering-skill/
├── plugins/prompt-engineering/ # the Claude Code plugin
│ └── skills/prompt-engineering/ # the skill itself (source of truth)
│ ├── SKILL.md # entry point + workflow
│ ├── references/ # technique catalog, model profiles, examples, etc.
│ └── assets/ # canonical XML template
├── adapters/ # other-tool versions (generated from _core.md)
│ ├── _core.md # single source for the lean adapter body
│ ├── cursor/ windsurf/ copilot/ codex/
│ └── README.md
├── scripts/build-adapters.sh # regenerate adapters from _core.md
├── docs/ # per-platform install guides + website snippets
├── .github/ # CI, issue/PR templates, funding
├── .claude-plugin/marketplace.json # Claude Code / Copilot marketplace
├── .cursor-plugin/plugin.json # native Cursor plugin manifest
├── gemini-extension.json + GEMINI.md # Gemini CLI extension
├── install.sh # one-line curl installer
├── CHANGELOG.md CONTRIBUTING.md CODE_OF_CONDUCT.md SECURITY.md LICENSE
License
MIT — free to use, modify, and redistribute with attribution.
