prompt-master: Stop Hand-Tuning Prompts for Every AI Tool

Every new AI tool wants a different prompt shape. Midjourney wants comma-separated descriptors with --ar flags. Cursor wants ordered sections with explicit file paths. Claude Code wants a brief plus a "Done When" checklist. Most people write these by hand, re-derive the conventions each time, and burn tokens on verbose framings. nidhinjs/prompt-master collapses that loop into a single slash command.

Why This Skill Matters

The skill encodes a fixed audit: nine dimensions of intent (task, input, output, constraints, context, audience, memory, success criteria, examples), a hard cap of three clarifying questions, a tool-specific profile picker for 20+ named systems, and a token-efficiency pass that strips non-load-bearing words before delivery. It refuses to ask more than three questions even when context is thin, and it explicitly avoids techniques that read well but degrade output — Tree of Thought, Graph of Thought, Universal Self-Consistency, and prompt chaining are out of scope.

For anyone who writes prompts as part of their job — designers using Midjourney, engineers using Cursor, PMs using Claude or GPT — this skill removes the rewrite loop.

Installation

The README recommends two paths. The browser path is the one most users want.

# Browser (claude.ai) — recommended path
# 1. Download https://github.com/nidhinjs/prompt-master as ZIP
# 2. claude.ai → Sidebar → Customize → Skills → Upload a Skill
# 3. Select the unzipped folder

If you are on Claude Code CLI and prefer a local install:

mkdir -p ~/.claude/skills
git clone https://github.com/nidhinjs/prompt-master.git ~/.claude/skills/prompt-master

The README marks the CLI path as "Not Suggested" because the browser upload gets the memory-block continuity between sessions for free. Pick CLI only if you do not use claude.ai.

Real Workflow: Write a Midjourney Prompt From a One-Line Idea

You want a Midjourney image of "a realistic samurai standing in the rain at night" but you do not want to remember the convention for aspect ratios, version flags, or negative prompts.

Step 1. In a Claude session with prompt-master installed, type:

Write me a midjourney prompt for a realistic samurai standing in the rain at night.

Step 2. The skill detects the target tool, picks its Midjourney profile, runs the nine-dimension audit, decides the brief is already specific enough (no clarifying question needed), and returns one copy-ready block:

Realistic samurai standing alone in the rain at night, kimono soaked through,
katana lowered at his side, neon kanji sign reflection on wet asphalt behind
him, cinematic lighting, shallow depth of field, moody color grade --ar 16:9
--v 6 --style raw

Negative prompt: blurry, low quality, watermark, cartoon, anime, extra limbs,
deformed hands

Strategy: comma-separated descriptors over prose, lighting and mood anchored
early, aspect ratio and version locked, negative prompt prevents style drift.

Step 3. Paste it into Midjourney. The strategy note is the audit trail — if the output is off, you know which lever to tune (lighting, framing, negative list).

Real Workflow: Hand Cursor a Refactor Brief With No Ambiguity

You want Cursor to refactor your auth module but past experience shows vague asks produce vague diffs. Prompt-master is good at the opposite — output that leaves no room for guesswork.

Step 1. Invoke with /prompt-master:

/prompt-master I want to ask Cursor to refactor my Express auth middleware
to use async/await consistently and split the JWT verify source into its own
module.

Step 2. The skill asks at most one question (for example: "Target Node version?") and emits a brief with: an explicit list of files to touch (src/middleware/auth.ts, src/lib/jwt.ts), the desired signature of the new helper, the constraint that no behavior changes, the test file to update, and a "Done When" list.

Step 3. Paste the brief into Cursor. The "Done When" section is the contract — Cursor's diff should make every item true, and you review against it instead of guessing what "done" means.

• Save the output of one good prompt run as your personal template — the Memory Block carries stack and convention choices forward across sessions.
• When the brief is thin, answer the three clarifying questions precisely; do not volunteer extra context the skill did not ask for, because it intentionally refuses to consume more than that.
• For o1/o3-style reasoning models, do not add CoT or step-by-step instructions to the output — those models think internally and explicit CoT degrades their output.
• After a few uses, check the strategies the skill emits at the bottom of each prompt; they document the lever to tune when the result drifts.

When Not to Use This

If your task is short and well-defined — "rewrite this function to use Promise.all" — skip the skill and just type the ask. The nine-dimension audit and three-question gate exist for prompts where the brief is fuzzy or the target tool is opinionated; for one-line rewrites they add overhead without value.

See the leaderboard for more prompt-craft skills.