• Why this skill
• How it works
• Example
• Installation
• Verify installation
• Usage
• What's inside
• Updating
• Uninstall
• Contributing
• Project structure
• License

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.

You paste a draft (or describe what you want). The skill runs a five-phase workflow:

1. Diagnose — real intent, task type, target model, audience, output format, constraints, and the draft's weaknesses.
2. Select techniques — only those that add value for this task and model.
3. Rewrite — assemble the prompt in a canonical structure (role, context+motivation, XML-tagged data, positive instructions, few-shot, output contract).
4. Self-review — run a quality checklist; the golden test is "could a colleague with no context run this without doubt?"
5. Deliver — the optimized prompt + a changelog mapping each change to the weakness it fixes.

Input (your draft):

classify the sentiment of this comment: {{comment}}

Output (what the skill returns):

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 the expressions that carry emotion (positive or negative).
2. Weigh the overall tone, accounting for irony and negation.
3. Choose the category that best represents the predominant sentiment.

Put your reasoning in <analysis>. On the last line, answer ONLY one word: positive, negative, or neutral.

Plus a changelog explaining each change (role, enum output contract, short CoT for irony/negation, XML input). See six more cases in worked-examples.md.

Pick your tool. Claude Code is the primary target; the rest are fully supported.

curl -fsSL https://raw.githubusercontent.com/PhAlves23/prompt-engineering-skill/main/install.sh | bash

/plugin marketplace add PhAlves23/prompt-engineering-skill
/plugin install prompt-engineering

copilot plugin marketplace add PhAlves23/prompt-engineering-skill
copilot plugin install prompt-engineering@prompt-engineering-marketplace

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

curl -fsSL https://raw.githubusercontent.com/PhAlves23/prompt-engineering-skill/main/adapters/codex/AGENTS.md -o AGENTS.md

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"

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.

• curl / manual: rm -rf ~/.claude/skills/prompt-engineering
• Claude plugin: /plugin uninstall prompt-engineering
• Codex: rm ~/.agents/skills/prompt-engineering

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.

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

MIT — free to use, modify, and redistribute with attribution.