wtf is boris-prompts?

boris-prompts is what its README calls an agent skill: a small bundle of instructions you install into a coding assistant like Claude Code, Cursor, ChatGPT, Gemini, or Codex, so that the assistant becomes a tiny expert at writing prompts for other AI tools. The methodology comes from a talk by Boris, who works at Anthropic and is described as the creator of Claude Code, and the project packages his tips into something an agent can apply on demand. When you ask the assistant something like how should I prompt Cursor to add dark mode, or my ChatGPT prompt keeps going off the rails, the skill activates. If your request gives it enough information, it writes the prompt directly. If not, it asks up to three short clarifying questions first, using the agent's question tool, and then produces the prompt. The skill follows five principles, applied in order. The first is that short beats long, and that a two-sentence prompt usually outperforms a screenful of detail. The second is that adding the line make a plan first and run it by me is the highest-value addition you can make to almost any non-trivial prompt. The third is not to over-specify, instead, point the model at where to start, like saying look at how the existing middleware is structured. The fourth is that a feedback loop, tests or lint or screenshots, beats writing detailed step-by-step instructions. The fifth is that persistent context belongs in files like CLAUDE.md or .cursorrules, not stuffed into every prompt. The README walks through an example. If you say you want Claude to add JWT authentication to an Express app and you have tests, the skill produces a compact prompt that asks Claude to look at existing middleware, make a plan, get approval, then implement and run the tests until they pass. Installation uses a separate tool called the skills CLI, run with npx skills add followed by either the GitHub shorthand or the full URL, with flags for installing globally or scoping it to a single agent like Claude Code. The project layout follows that CLI's convention: a top-level skills folder with one subfolder per skill, each containing a SKILL.md file. The license is MIT.