Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

29 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Super-Prompt v2.2: Precision Prompt Engineering — AI That Can't Misunderstand You

License: MIT Standard: Agent Skills GitHub stars

🎯 The prompt engineering tool that prevents AI from misunderstanding you. Anti-Ambiguity Interception × Web Pre-Search × Six-Tier Adaptive Grading × Information Completeness Simulation — turn a single sentence or a product document into a production-grade prompt that AI cannot misinterpret.

Install: npx skills add wlj103/super-prompt  |  Trigger: "write me a prompt" / "optimize my prompt" / "forge a prompt"

Table of Contents

How It Works

When you say "write me a prompt," Super-Prompt doesn't just start writing — it runs a structured pipeline:

Your input → [Step 1: Six-Tier Grading] → picks the right funnel depth
         → [Step 2: Requirements Funnel] → digs into what you actually need
         → [Step 3: Completeness Simulation] → finds gaps before writing
         → [Step 4: Seven Gates] → locks down every ambiguity
         → [Step 5: Multi-Format Delivery] → prompt + visuals + docs

Step 1: Six-Tier Input Grading

Not all users are the same. Super-Prompt auto-classifies your input and adapts:

Tier What You Say What Happens
① One-liner "Write me a prompt" (no context) Full four-layer requirements funnel
② Missing params Has direction, missing key details Streamlined: problem discovery + assumption deconstruction
③ Clear requirement 200-500 word description, logically coherent Confirm → Simulate gaps → "Which gaps to fill?"
④ Diagnostic "Review this prompt for issues" Direct 18-item anti-ambiguity scan
⑤ Existing prompt Brings a prompt to optimize Review mode + completeness simulation
⑥ Complete document Full PRD / spec / structured doc Review mode + completeness simulation

Key principle: For professional users (tiers ③⑤⑥), the deeper funnel is always offered, never forced.

Step 2: Four-Layer Requirements Funnel (Tiers ①②)

For vague or incomplete needs, Super-Prompt uses a requirements mining funnel that fuses five industry methodologies:

Layer Method What It Does Industry Foundation
1. Problem Discovery STAR Model Asks "what problem did you encounter?" not "what do you want?" Customer Discovery (Steve Blank)
2. Assumption Deconstruction 5 Whys Every user statement is a hypothesis, not a fact. "I want X" → "Why?" → dig to root need 5 Whys (Toyota), Socratic Questioning
3. Body Mining Pointer Analysis External pointers (URLs/screenshots) → open and analyze; internal pointers (vague words) → expand via questioning Needfinding (Stanford d.school)
4. Gap Discovery Hypothesis Map Verified vs Assumed vs Inherited → cross-reference → gap questions Socratic Questioning, Funnel Questioning

Core premise: Users don't say what they need — they say what they think the solution is. Every word is a pointer, not the requirement body.

Validated results:

  • Chinese calendar widget: coverage 15% → 100% (16/16 requirements found)
  • Pomodoro timer: 3/8 → 8/8
  • Knowledge management: 3/14 → 14/14
  • v2.1 calendar widget (280 words, tier ③): 4-layer funnel → completeness simulation → 3 gaps surfaced → user filled 1 → full prompt. "Better than I expected."

Step 3: Information Completeness Simulation (Tiers ③⑤⑥)

Before writing a single prompt word, Super-Prompt simulates a dry run and reports gaps with three markers:

Marker Meaning Example
Confirmed User said it, no need to ask "磨砂半透明" (frosted translucent)
⚠️ Guessing Agent can guess but shouldn't; guessing wrong = failure Zoom behavior, weather API, highlight style
Missing Can't run without it; must fill Font licensing, holiday data source

The ⚠️ marker is the innovation: the agent can guess many things, but guessing = gambling. This forces assumptions to surface as explicit questions.

Step 4: Seven Gates

Every prompt passes through seven sequential gates. No skipping, no "good enough."

Gate What It Does What It Prevents
G0: Intent Anchoring Classifies input into six tiers, picks the right funnel depth Building a prompt for the wrong problem
G1: Role & Boundaries Defines AI role, input/output format, hard/soft constraints Role ambiguity, scope creep, format mismatch
G2: Structure Skeleton Recommends prompt section structure; user confirms Disorganized prompts with overlapping responsibilities
G3: Section-by-Section Refinement Drafts each section, runs anti-ambiguity check before moving to next Vague terms like "appropriate," "good," "reasonable"
G4: Anti-Ambiguity Audit 18-item checklist, item by item; no empty "pass" Ambiguity residues that survived earlier gates
G5: Adversarial Test Simulates 3 ways AI could misinterpret key instructions Prompts that work in testing but fail on edge cases
G6: Delivery Final prompt + Quick Reference Card + maintenance obligations Deploying a prompt without knowing how to maintain it

Web search runs throughout all gates — every question is backed by domain knowledge, not assumptions.

Step 5: Multi-Format Output

Not just text. When the prompt involves workflows, architectures, or data flows, Super-Prompt generates supplemental outputs:

Scenario Output
Decision trees / branching logic Flowcharts (Draw.io / Mermaid)
System architecture Architecture diagrams
Data flows / processing pipelines Data flow diagrams
Concept relationships Mind maps
Data comparison / statistics Charts (bar, pie, line)
Demo / presentation needs Podcast audio or video
Formal delivery .docx / .pptx / PDF

Principle: If plain text isn't enough, supplement with visuals, documents, audio, or links.


The Problem with Prompt Engineering

Prompt engineering has a fundamental flaw: models learn to circumvent prompts. You say "enforce strict verification," and the model outputs four words: "Verification enforced successfully." You say "be detailed," and it writes 3 lines of fluff. You say "use JSON format," and it freely invents field names. You say "don't hallucinate," and it hallucinates with extra confidence.

This isn't a model intelligence problem — it's a prompt discipline problem. Traditional templates give you the what but not the how. They tell you to "be specific" without telling you what specific means, or how to verify it before the model runs.

Super-Prompt is the missing layer. It doesn't just help you write prompts — it forces you through a structured, evidence-backed process that leaves no ambiguity for the AI to exploit.

What is Super-Prompt?

Super-Prompt is a prompt engineering framework and optimization tool designed as an Agent Skills skill. It's not a template, not a fill-in-the-blank sheet — it's a complete prompt discipline system that:

  • 🔍 Pre-searches the web before you write a single word, so your prompt is backed by domain knowledge
  • 🧠 Simulates information completeness before drafting — finds gaps with ✅⚠️❌ markers so you choose what to fill
  • 🛡️ Runs 18 anti-ambiguity checks that systematically lock down every path the AI could misinterpret
  • ⚔️ Adversarially tests key instructions by simulating 3 ways AI could misinterpret them
  • 📋 Produces a Quick Reference Card with maintenance obligations, degradation handling, and exception protocols

The result: prompts that work the first time — whether you're writing a one-line classifier, a 500-line agent system prompt, or a multi-agent SOP.

Who Is This For?

You are... Super-Prompt helps you...
AI/LLM Engineer Ship production-grade system prompts with built-in validation and anti-hallucination guards
Claude Code / Cursor / Copilot user Stop wasting tokens on vague instructions; get precise, verifiable agent instructions
Prompt Engineer Replace guesswork with structured anti-ambiguity checks and adversarial testing
Solo Developer Get the quality of a team code review through automated anti-ambiguity checks
Technical Writer Convert product requirements into structured, unambiguous AI instructions
AI Agent Builder Design agent system prompts that handle edge cases, degradation, and exceptions

Quick Start

npx skills add wlj103/super-prompt

Then trigger with any of these:

  • "write me a prompt"
  • "optimize my prompt"
  • "forge a prompt"

Features

🎯 Six-Tier Adaptive Grading

Super-Prompt classifies your input into six tiers — from one-liners to complete PRDs — and adapts the funnel depth accordingly. Professional users get professional treatment: the four-layer funnel is offered, never forced.

🧠 Information Completeness Simulation

Before writing a single prompt word, Super-Prompt simulates a dry run of your requirement and reports gaps with three markers: ✅ Confirmed, ⚠️ Guessing (agent can guess but shouldn't), ❌ Missing. No more "does this look right?" — you get a structured gap report and choose what to fill.

🔍 Web Pre-Search Engine

Before asking you a single question, Super-Prompt searches the web for domain-specific knowledge, industry best practices, known edge cases, and common pitfalls. Your prompt is backed by evidence, not assumptions.

🛡️ 18-Point Anti-Ambiguity Checklist

The most comprehensive anti-ambiguity system in any prompt tool. Four categories: semantic vagueness, structural gaps, boundary ambiguity, and hidden assumptions. Each of the 18 items requires evidence — no checkbox theater.

⚔️ Adversarial Testing

Before delivery, Super-Prompt simulates 3 ways the AI could misinterpret your key instructions. Each weak point is hardened before you ship.

📋 Quick Reference Card

Every prompt ships with a maintenance card that includes degradation scenarios, exception handling protocols, update frequency recommendations, and known failure modes.

🌐 Bilingual Support

Full Chinese and English support. SKILL.md, README, CHANGELOG, and CONTRIBUTING are all available in both languages.

🔌 Agent Skills Compatible

Built as an Agent Skills skill. Works with any Agent Skills-compatible AI coding assistant, including Claude Code, Cursor, and Coze.

Design Principles

  1. Pre-search, don't guess — Every question backed by web search; results converted into options for injection
  2. Layered guidance, not interrogation — ≤4 questions per round, starting from the simplest mandatory questions
  3. Anti-ambiguity, not assumption — 18-item checklist, systematically lock every ambiguity path
  4. Evidence-driven, not opinion-driven — Design decisions backed by search evidence or industry standards
  5. Adversarially tested, not "looks good" — Simulate AI misinterpretations before shipping
  6. Deliver with insurance — Quick Reference Card + maintenance obligations + degradation and exception handling
  7. Output on demand, not over-deliver — Never skimp, never bloat; safety margin only increases, never decreases

How It Compares

Feature Super-Prompt Promptfoo dair-ai Guide orbit-prompt
Approach Structured prompt discipline Prompt testing framework Educational guide Claude Code skill
Anti-ambiguity ✅ 18-item checklist
Web pre-search ✅ Built-in
Adversarial testing ✅ Simulated misinterpretations ✅ CI-integrated
Production output ✅ Prompt + QRC + maintenance ❌ Test results only
Best for Writing precise prompts Testing existing prompts Learning prompt engineering Quick Claude Code prompts

super-spec × super-prompt

Super-Prompt is the upstream of super-spec. The flow is:

User need → [super-prompt] → precise prompt document → [super-spec] → 5 engineering docs → delivery
super-prompt super-spec
Comes first ✅ Defines what to do and how to say it Takes the prompt and breaks it into engineering docs
Problem AI expresses vaguely, lacks structure, leaves ambiguity AI skips steps, claims success with zero evidence
Solution Prompt discipline — 7 gates, 18 anti-ambiguity items, adversarial testing Engineering discipline — 7 gates, 5 documents, anti-corner-cutting
Domain How agents should express How agents should execute

Together: prompt defines the contract, spec enforces the execution.

FAQ

How is this different from a prompt template?

Templates give you a structure to fill in. Super-Prompt gives you a process — it actively searches, checks, tests, and validates. A template won't tell you your prompt is ambiguous; Super-Prompt's 18-item checklist will.

Does this work with Claude? GPT? Other models?

Yes. Super-Prompt is model-agnostic. The prompt discipline principles apply to any LLM — Claude, GPT, Gemini, Llama, or local models.

Can I use this for agent system prompts?

Yes — that's the Heavy mode. It's designed for complex agent instructions, multi-agent SOPs, and production-grade system prompts.

What's the Quick Reference Card?

A one-page maintenance document that ships with every prompt. It tells you when the prompt might degrade, how to detect it, and what to do about it.

Is this free?

Yes. MIT licensed, open source, free forever.

Can I contribute?

Absolutely. See CONTRIBUTING.md for guidelines. We welcome prompt patterns, anti-ambiguity item suggestions, and adversarial testing improvements.

License

MIT © 2026