Section 3.1 Confirmed Issues - MUST use markdown table: | # | File | Line | Issue Summary | Dimension | Fix Type | Group by severity: 🔴 → 🟡 → 🟢, each as separate table.

Section 3.2 Filtered Issues - MUST use markdown table: | # | File | Line | Issue Description | Filter Reason |

Section 4 Fix Proposals - MUST include for EVERY confirmed issue: Location, Problem, Impact, Current (code block), Proposal (code block).

Section 5 Conclusion - PHASE GATE: After outputting, STOP and wait for user input. Do NOT apply fixes until user confirms.

Hello-AuditKit: AI Coding Assistant Audit System

Table of Contents

• Entry Point
• Overview
• Core Principles
• Audit Execution
• Reference Files
• Special Reminders
• External Documentation

Entry Point

On skill invocation, first determine the audit target:

Welcome Message (when no target):

👋 Hello-AuditKit - AI 配置审计工具

支持审计：
• 提示词文本（直接粘贴或任意文件）
• Memory 文件（AGENTS.md, CLAUDE.md, GEMINI.md）
• Skills（含 SKILL.md 的目录）
• Plugins（含 .claude-plugin/ 的目录）

使用方式：
1. 粘贴要审计的提示词文本
2. 提供文件路径：/path/to/file.md
3. 提供目录路径：/path/to/skill/

请提供要审计的内容或路径：

CRITICAL: After showing welcome, STOP and wait for user input. Do NOT proceed with audit until target is provided.

Overview

Comprehensive audit system for AI coding assistant configurations:

Core Principles

Source: Based on Latest GPT Prompting Guide (openai-cookbook/examples/gpt-5)

Principle 0: GPT Prompting Guide Compliance (MANDATORY)

CRITICAL: This is the PRIMARY audit standard. Every audit MUST check these items and report findings.

For ALL content containing AI instructions, verify:

XML Tags Compliance (GPT-5.2 MANDATORY):

CRITICAL: For agentic/multi-phase prompts, XML tags are REQUIRED to prevent format drift. This is a strict audit rule.

Audit output: Report GPT Guide Compliance status with evidence for each check, including XML tags compliance.

Principle 1: GPT-5.2 Specific Checks (For GPT-5.2+ Prompts)

Source: GPT-5.2 Prompting Guide - Key Behavioral Differences

Principle 2: 5-Point Verification

Before marking ANY issue, verify:

1. Concrete Scenario - Can describe specific failure?
2. Design Scope - Within intended boundaries?
3. Functional Capability - Can it actually do what it claims? (Requires domain knowledge first)
4. Flaw vs Choice - Unintentional error or valid choice?
5. Threshold Met - Above quantified threshold?

If ANY fails → Discard the issue

Principle 3: Occam's Razor

"If unnecessary, don't add."

Fix Priority: DELETE > MERGE > RESTRUCTURE > MODIFY > ADD

Principle 4: AI Capability

• AI CAN infer: synonyms, context, standard terms
• AI CANNOT: 3+ step inference, domain-specific variations
• If <30% would misunderstand → exempt from issue

Principle 5: Size Tolerance (SKILL.md body only)

Note: Reference files have no official line limit. Evaluate based on content nature.

Principle 6: Prompt Compliance

For prompts/instructions, verify critical checks (see type-prompt.md → Prompt Compliance Checks):

• Verbosity constraints (Severe)
• Scope boundaries with "do not" list (Severe)
• No fabrication instruction (Severe)
• Output schema for structured tasks (Warning)

Principle 7: Grounding & No Fabrication

• Base all findings on actual content examined
• Never fabricate line numbers, file names, or issue details
• Use hedging language for uncertain assessments: "appears to", "may indicate"

Audit Execution

CRITICAL: Each step below is MANDATORY. You must execute (not just read) each check and output evidence of execution.

Agentic Updates: Send brief updates (1-2 sentences) only at major phase transitions. Avoid narrating routine tool calls.

Tool Parallelization: When scanning multiple files or checking multiple dimensions, parallelize independent read operations for efficiency.

Step 0: Fetch Latest Prompting Guide (MANDATORY STANDARD)

CRITICAL: The GPT Prompting Guide is a primary audit standard, not just a reference. All prompts/instructions MUST be evaluated against these rules.

1. Access this directory page to get file list: https://github.com/openai/openai-cookbook/tree/main/examples/gpt-5
2. Select the latest version prompting guide (e.g., gpt-5-2_prompting_guide.ipynb > gpt-5-1_prompting_guide.ipynb > gpt-5_prompting_guide.ipynb)
3. Extract and apply these mandatory checks from the guide:
   • Verbosity constraints: "≤N sentences/bullets/words" present?
   • Scope discipline: "EXACTLY and ONLY what requested" + "Do NOT" list present?
   • Stop conditions: Explicit completion criteria for multi-phase content?
   • No fabrication: "Never fabricate..." instruction for factual tasks?
   • Long-context handling: Outline + constraint restatement for >10k tokens?
   • Tool preference: Tools over internal knowledge for fresh data?
   • Agentic updates: Brief (1-2 sentences) at major phases only?
   • XML structure enforcement (GPT-5.2+): XML tags wrap critical constraints?
4. XML Tags Compliance Check (STRICT RULE):
   • Identify prompt type: agentic, extraction, factual, tool-using, long-context, high-risk
   • Check for required XML tags per type (see Principle 0 table)
   • If XML tags missing for applicable type → Flag as Severe
   • Required tags: <output_verbosity_spec>, <design_and_scope_constraints>, <user_updates_spec>, <extraction_spec>, <uncertainty_and_ambiguity>, <tool_usage_rules>, <long_context_handling>, <high_risk_self_check>
5. Cross-validate with built-in checks in type-prompt.md
6. Flag as Severe if audited content violates any mandatory check above

Evidence Output: Note guide version fetched, list mandatory checks applied (including XML tags), note any violations found.

If WebFetch fails: Retry before falling back to offline mode. If still fails, use built-in checks in type-prompt.md, note "offline mode - [error reason]" in report.

Step 1: Detection & Classification

Scan path → identify type → load appropriate rules:

Any text/file   → type-prompt.md (default for unrecognized types)
Memory file     → type-memory.md (AGENTS.md, CLAUDE.md, GEMINI.md)
Skill           → type-skill.md (directory with SKILL.md)
Plugin          → type-plugin.md (directory with .claude-plugin/)
Composite       → Apply all + cross-*.md

Step 2: Execute Universal Checks (ALL TYPES)

FIRST: Execute Principle 0 (GPT Guide Compliance) checks before proceeding.

GPT Guide Compliance Check (MANDATORY FIRST):

Execute each check from Principle 0 table, record status and evidence (line numbers, quotes).

Every audit MUST execute these checks from rules-universal.md:

Numbering Check Execution (commonly missed):

1. Find all numbered lists (1. 2. 3. or Step 0, Step 1, etc.)
2. Verify: sequential? no duplicates? no gaps?
3. Find all TOC entries → verify each has matching heading
4. Cross-section: if steps span sections (Step 0 here, Step 3 there), verify continuity

Step 3: Execute Type-Specific Checks

Based on content type, execute ALL checks in the relevant file:

For Prompts (type-prompt.md):

For Memory Files (type-memory.md):

For Skills (type-skill.md):

For Plugins (type-plugin.md):

Step 4: Execute Cross-Cutting Checks (Multi-file Systems)

For Skills, Plugins, Composites, execute ALL checks from:

From cross-design-coherence.md:

From cross-progressive-loading.md:

From cross-composite.md:

Step 5: Issue Verification (5-Point Check)

For each suspected issue, verify ALL points:

1. Concrete scenario - Can describe specific failure?
2. Design scope - Within intended boundaries?
3. Functional capability - Does implementation match claimed capability?
4. Flaw vs choice - Unintentional error or valid design?
5. Threshold met - Above quantified threshold?

If ANY fails → Discard the issue (move to Filtered)

For "missing/incomplete" issues: Re-read the source content fully before confirming. ASCII diagrams are prone to parsing errors on first scan.

Step 6: Fix Proposal Verification (Principle Check)

CRITICAL: Before outputting ANY fix proposal, verify it against core principles:

Verification Process:

1. For each proposed fix, ask: "If I remove this fix, would AI still produce correct output based on existing content?"
2. If YES → The fix is unnecessary, discard it
3. If NO → Verify the fix uses minimal intervention (prefer MODIFY over ADD)

If ANY check fails → Revise or discard the fix proposal

Step 7: Generate Report

Follow references/ref-output-format.md for structure.

Section 2 Cross-Cutting Analysis MUST include:

• Naming & Numbering: actual check results with specific findings
• TOC-Content Match: comparison results
• Reference Integrity: broken refs listed
• (For multi-file) Design Coherence, Progressive Loading results

Section 3 Issue Inventory MUST include:

• Verification Statistics: "Scanned X → Verified Y → Filtered Z"
• Both Confirmed and Filtered issues with filter reasons

Step 8: Wait for User Confirmation (PHASE GATE)

CRITICAL: After generating the report, STOP and wait for user input. Do NOT apply any fixes automatically.

User interaction flow:

1. Output complete audit report (Sections 0-5)
2. STOP - Wait for user to select which fixes to apply
3. Only after user confirms (e.g., "1", "1,2", "all") → Apply selected fixes
4. If user provides no selection → Do nothing, wait

Reference Files

Layer 0: Core Methodology (Immutable Principles)

Read references/methodology-core.md when:

• Need to verify if something is truly an issue
• Deciding fix priority
• Understanding AI capability boundaries

Layer 1: Universal Rules (Common Rules)

Read references/rules-universal.md when:

• Starting any audit
• Need Should Flag / Should NOT Flag patterns
• Checking size thresholds

Layer 2: Type-Specific Rules (Type Rules)

Layer 3: Cross-Cutting Rules (Cross-Cutting Rules)

Layer 4: Reference Materials (Reference Materials)

Special Reminders

Key References by Topic

Quick Filtering Rules

External Documentation

Version Policy: Always use the latest version of GPT prompting guides as authoritative source. When multiple versions exist in the gpt-5 directory, prefer the highest version number (e.g., gpt-5.2 over gpt-5.1 over gpt-5). The directory contains prompting guides, troubleshooting guides, and optimization cookbooks.