An agent skill that writes high-quality prompts for any LLM (Claude, Claude Code, ChatGPT, Gemini, Cursor, Windsurf, etc.) using Boris's methodology — the patterns Boris (Anthropic, Claude Code's creator) shared in his "Pro Tips & Tricks" talk.

If your request is unclear, the skill asks 1–3 targeted clarifying questions before writing. The output is short, copy-pasteable, and follows the five principles below.

https://x.com/Etudecn/status/2057238154701426726

This skill is distributed via the skills CLI (npx skills). Works with Claude Code, Codex, Cursor, OpenCode, and 50+ other agents.

# GitHub shorthand
npx skills add LingyiChen-AI/boris-prompts --skill boris-prompts

# Or full URL
npx skills add https://github.com/LingyiChen-AI/boris-prompts --skill boris-prompts

# Or point at the skill folder directly
npx skills add https://github.com/LingyiChen-AI/boris-prompts/tree/main/skills/boris-prompts

# Install globally (available across all projects)
npx skills add LingyiChen-AI/boris-prompts --skill boris-prompts -g

# Install only for Claude Code
npx skills add LingyiChen-AI/boris-prompts --skill boris-prompts -a claude-code

After install, restart your agent (or start a new session) and the skill loads automatically.

When you say things like:

• write a prompt that makes Claude convert all our logs to JSON
• how should I prompt Cursor to add dark mode
• this ChatGPT prompt isn't working — keeps going off the rails
• give me a good way to ask Gemini to refactor this module

…the skill triggers. If it has enough info, it writes the prompt directly. If not, it asks at most three questions (via AskUserQuestion) and then writes it.

The skill applies these in order:

1. Short beats long. A two-sentence prompt usually beats a screen-filling one. If you find yourself writing more than five lines, half of it probably belongs in a persistent-context file.
2. "Make a plan first" is the single highest-ROI addition. Appending "Before you write code, make a plan and run it by me for approval" upgrades almost any non-trivial prompt.
3. Don't spec every detail. Point the model at the right starting place ("look at how X is implemented") instead of re-describing the codebase.
4. A feedback loop beats detailed instructions. Give the model a way to verify (tests, screenshots, lint) and let it iterate, rather than spelling out every step.
5. Persistent context lives in files / settings, not the prompt. CLAUDE.md, .cursorrules, Custom Instructions, Projects, Gems — whichever the target supports.

You: "I want Claude to add JWT auth to my Express app. There are tests."

Skill output:

Add JWT auth to the Express app. Look at how existing middleware is structured first, then make a plan and run it by me. After implementing, run the test suite and iterate until it passes.

Plan-first + feedback loop applied. Anchors Claude in existing patterns rather than re-describing them.

This repo follows the convention the skills CLI expects: a top-level skills/ directory containing one folder per skill, each with a SKILL.md.

boris-prompts/                  (this repo)
└── skills/
    └── boris-prompts/          (the skill)
        └── SKILL.md

To add more skills later, drop them in as siblings under skills/.

Methodology from Boris's "Claude Code Pro Tips" talk (Anthropic). This repo just packages it as an agent skill.

Community: https://linux.do

MIT