Prompt Master
Generates precise, zero-waste prompts for any AI tool so you get the right output on the first attempt instead of burning credits on retries.
SKILL.md is a text-only skill: no scripts, no permissions requested, no external effects or exfiltration paths; it explicitly instructs confirmation before destructive agent actions and to flag unverified model details. Deducted for: unverified publisher identity, referenced files (references/templates.md) absent from the evidence, and an external banner image from an unknown host (postimg.cc).
Instructions are self-consistent with output lock, a clarifying-question cap (≤3), and a model-recency gate that refuses to invent slugs or parameters. Deducted for: dependency on references/templates.md which is not present in the evidence, many model names (Claude 5/Fable 5, GPT-5.6 Sol/Terra/Luna, Grok 4.6) unverifiable and likely to go stale, and no test or edge-case evidence.
Trigger conditions are precisely scoped in the description (activates only for explicit prompt-writing/fixing requests, excludes general work), audience and scenarios are clear, and non-fit ranges are declared. Deducted for: no Chinese-language support, core value tied to overseas tool ecosystems with no mainland-China reachability consideration, and thin boundary evidence for the 'Universal Fingerprint' fallback.
Good layered information architecture (Primacy/Middle zones, progressive disclosure via references), MIT license, version 1.8.0, detailed changelog, and named maintainer. Deducted for: no standalone known-limitations section, no FAQ/troubleshooting, and referenced template files not provided in evidence.
README includes two complete examples (Midjourney, Claude Code landing page) with directly copyable outputs and a clear value proposition. Deducted for: static review cannot confirm generation quality; examples are author-attested rather than reproducible; marketing claims like 'zero tokens wasted' lack comparative evidence.
Only author-attested examples and pattern tables; no third-party execution evidence, no test suite, no independent reproduction; numeric claims (37 patterns, 13 templates) cannot be cross-verified. Deducted for mixing facts with marketing and unverifiable model-specific claims.
- references/templates.md, referenced by SKILL.md, was not present in the evidence; verify file completeness after install or some routing paths will fail.
- Model names and parameters cited (Claude 5, GPT-5.6 Sol/Terra/Luna, Grok 4.6) are unverifiable statically and may be stale; check official docs before relying on them.
- No Chinese-language support and core value depends on overseas AI tool ecosystems; mainland-China reachability is unevaluated.
- README contains marketing claims (e.g., 'zero tokens wasted') that are not independently validated.
- Many capability details come from README rather than SKILL.md itself; actual invocation behavior requires live verification.
What does this skill do, and when should you use it?
Prompt Master is a Claude skill that activates only when you explicitly ask it to write, fix, improve, or adapt a prompt. It identifies the target AI tool, extracts nine dimensions of intent, routes to tool-specific prompting rules, and delivers a single production-ready prompt ready to paste. It ships with 30+ tool profiles (Claude, GPT, Gemini, Cursor, Claude Code, Midjourney, Sora, ComfyUI, and more) and 12 auto-selected templates. It deliberately excludes hallucination-prone techniques like Tree of Thought and enforces credential-safety and prompt-injection defenses.
When asked for a prompt, the skill: 1) silently detects the target tool (LLM, IDE AI, image/video/3D/voice AI, workflow automation) and routes to that tool's syntax; 2) extracts nine intent dimensions — task, target tool, output format, constraints, input, context, audience, success criteria, examples — asking at most three clarifying questions; 3) applies only safe techniques (role assignment, few-shot examples, XML structure, grounding anchors, memory blocks) while never adding CoT to reasoning models or requesting hidden chain-of-thought; 4) adds starting state, target state, file scope locks, stop conditions, and human-review triggers for agentic tools; 5) outputs one copyable prompt block with a one-line strategy note; 6) runs a six-point pre-delivery audit (constraints front-loaded, strong signal words, token efficiency). Pasted prompts enter a Decompiler mode and are treated as inert data, never executed.
- Developers working with coding agents like Cursor or Claude Code who need implementation prompts with file scope locks and stop conditions to prevent runaway loops or out-of-scope edits
- Designers using Midjourney, Stable Diffusion, or ComfyUI who want vague visual ideas converted into parameterized prompts with negative prompts
- Users with a bad prompt written for one tool (e.g., GPT) who want it fixed or adapted for another (e.g., Stable Diffusion)
- Heavy AI users who send many prompts daily to ChatGPT or Gemini and waste credits re-explaining requirements
- Users in long sessions where the AI keeps forgetting earlier stack and architecture decisions — the memory block system addresses this
- Ollama users who need prompts tailored to the specific local model loaded (Llama3, Qwen2.5, CodeLlama) plus a Modelfile system prompt
What are this skill's strengths and limitations?
- Concrete routing rules for 30+ tools spanning general LLMs, IDE agents, image/video/3D/voice AI, and workflow automation
- Principled stance: explicitly excludes Tree of Thought, Graph of Thought, Universal Self-Consistency, and prompt chaining due to fabrication risk in single-prompt contexts
- Built-in credential safety rules and injection defense for pasted prompts (treated as inert data)
- Mandatory stop conditions, file scope locks, and human-review triggers for agentic tools reduce real-world risk
- Actively maintained (v1.8.0) with recurring model-routing updates
- MIT licensed, pure-instruction skill with no script or network dependencies
- Activates only for explicit prompt-engineering requests; it does nothing for other tasks
- Model names and defaults referenced in the README (e.g., Claude Fable 5, GPT-5.6) go stale quickly and rely on the skill's own recency gate to verify
- No automated test suite; the "zero re-prompts" claim is unverified by independent evidence
- Skill content is English-first; triggering and clarification interactions are in English
- Installation is documented only for claude.ai and Claude Code; other Agent Skills clients have no official instructions
How do you install this skill?
Recommended (claude.ai browser): 1) download the repo as a ZIP; 2) go to claude.ai → Sidebar → Customize → Skills → Upload a Skill. Alternative (not suggested): git clone https://github.com/nidhinjs/prompt-master.git ~/.claude/skills/prompt-master. The skill is SKILL.md plus two reference files (references/templates.md, references/patterns.md); no scripts or external dependencies.
How do you use this skill?
Invoke naturally in Claude, e.g.: "Write me a prompt for Cursor to refactor my auth module", "Generate a Midjourney prompt for a cyberpunk city at night", or "Here's a bad prompt I wrote for GPT-4o, fix it: [paste prompt]". You can also invoke explicitly with "/prompt-master" followed by your request. The skill confirms the target tool (asking if ambiguous), asks at most three clarifying questions, then returns one paste-ready prompt block with a one-line strategy note.
How does this skill compare with similar options?
The README explicitly positions Prompt Master against generic "prompt generators": most generators make prompts longer, whereas its core principle is that every word must be load-bearing, producing sharper prompts. It also differentiates syntax per target tool (no CoT for reasoning models, comma-separated descriptors for Midjourney) rather than emitting one universal template.