operator-humanizer

将人工智能生成的文本转换为真实的人类书写。检测并消除 24 种内容/语言/风格/沟通模式、500 多个 AI 词汇术语和结构性陈词滥调（二元对比、负面列表、虚假代理、戏剧性碎片、远距离叙述者）的 AI 讲述。分析统计信号（突发性、词汇多样性、句子一致性）。通过插入旁白、切线、节奏变化和策略特异性来注入个性。在使内容人性化、检查人工智能告诉、删除机器人模式、使文本听起来不那么优雅或像特定的人一样写作时使用。适用于社交帖子、文章、电子邮件、营销文案、时事通讯、脚本或任何需要听起来像真人编写的文本。

## 概述（中文）

将人工智能生成的文本转换为真实的人类书写。检测并消除 24 种内容/语言/风格/沟通模式、500 多个 AI 词汇术语和结构性陈词滥调（二元对比、负面列表、虚假代理、戏剧性碎片、远距离叙述者）的 AI 讲述。分析统计信号（突发性、词汇多样性、句子一致性）。通过插入旁白、切线、节奏变化和策略特异性来注入个性。在使内容人性化、检查人工智能告诉、删除机器人模式、使文本听起来不那么优雅或像特定的人一样写作时使用。适用于社交帖子、文章、电子邮件、营销文案、时事通讯、脚本或任何需要听起来像真人编写的文本。

## 原文

# Operator Humanizer

Eliminate AI tells. Inject authentic voice. Make it sound like a person wrote it.

## What This Skill Does

Two systems, combined:

1. **Pattern Detection** — 24 AI patterns, 500+ vocabulary terms, statistical signals
2. **Stop-Slop Rules** — structural clichés, phrase bans, sentence-level mechanics

Together they catch what the other misses. Pattern detection handles vocabulary and content signals. Stop-slop handles structure and rhythm.

**Reference files:**
- `references/patterns.md` — The 24 AI patterns with before/after examples
- `references/phrases.md` — Banned phrases and structural clichés
- `references/structures.md` — Structural patterns to avoid
- `references/vocabulary.md` — 500+ AI vocabulary terms by severity tier
- `references/statistical-signals.md` — Burstiness, TTR, sentence variance formulas
- `references/personality-injection.md` — How to add human touches
- `references/examples.md` — Before/after transformations

## Quick Start

1. **Scan content patterns** → Check patterns 1-6 in `references/patterns.md` (inflation, jargon, promotional language, vague attributions)
2. **Flag vocabulary** → Tier 1 = ban completely, Tier 2 = use sparingly, Tier 3 = watch density (`references/vocabulary.md`)
3. **Check phrases** → Remove all throat-clearing openers, emphasis crutches, adverbs (`references/phrases.md`)
4. **Break structures** → Destroy binary contrasts, negative listings, false agency (`references/structures.md`)
5. **Check style patterns** → Em dashes, bold overuse, emoji, passive voice (patterns 13-18)
6. **Remove communication artifacts** → Chatbot openers, sycophancy, cutoff disclaimers (patterns 19-21)
7. **Fix filler and hedging** → Stacked qualifiers, generic conclusions (patterns 22-24)
8. **Add personality** → Parentheticals, tangents, rhythm variation (`references/personality-injection.md`)
9. **Verify** → Read aloud. Does it sound like a human?

## Core Rules (Always On)

### Cut These Immediately

**Throat-clearing openers** — "Here's the thing:", "It turns out", "The uncomfortable truth is", "Let me be clear"

**Emphasis crutches** — "Full stop.", "Let that sink in.", "Make no mistake", "This matters because"

**Chatbot artifacts** — "Great question!", "I hope this helps!", "Let me know if...", "Certainly!", "Of course!"

**Binary contrasts** — "Not X, but Y", "It's not X, it's Y", "The answer isn't X, it's Y" → Just say Y.

**Negative listings** — "Not a X... Not a Y... A Z." → Just say Z.

**Generic conclusions** — "The future looks bright", "Exciting times lie ahead", "This represents a major step"

### Vocabulary Bans

**Tier 1 (dead giveaways — never use):**
delve, tapestry, vibrant, crucial, comprehensive, meticulous, embark, robust, seamless, groundbreaking, leverage, synergy, transformative, paramount, multifaceted, myriad, cornerstone, reimagine, empower, catalyst, invaluable, bustling, nestled, realm, showcasing, underscores, testament, pivotal, enduring, landscape (abstract), journey (metaphorical)

**Tier 2 (suspicious — use sparingly):**
furthermore, moreover, paradigm, holistic, utilize, facilitate, nuanced, illuminate, encompasses, proactive, ubiquitous, quintessential

**Tier 3 (watch de)
  - Reinvestment assumptions
  - WACC with explicit inputs (risk-free, ERP, beta, debt cost)
  - Terminal value: Gordon or exit multiple (state choice)
- Mandatory sensitivity grid:
  - WACC ±100 bps
  - terminal growth ±50 bps
- Output:
  - base/bull/bear fair value
  - sensitivity table

### 3) Sector-specific adaptation
#### Banks / Insurance / Financials
- Prioritize: `P/B`, `ROE`, asset quality proxies, capital adequacy proxies, funding cost/NIM proxies.
- De-emphasize EV/EBITDA.
- Evaluate sustainability of ROE and provisioning pressure.

#### Cyclicals (steel, chemicals, commodities, shipping)
- Use cycle-aware assumptions:
  - normalized margin, not peak margin
  - conservative terminal assumptions
- Add cycle-risk note as first-class risk item.

## Quality and business resilience checklist
Assess each item as `Strong / Neutral / Weak` with one-line evidence:
- Moat and pricing power
- Governance and capital allocation
- Earnings quality (cash conversion, accrual risk)
- Balance-sheet risk (leverage, maturity risk)
- Cyclicality and external dependency
- Execution track record

## Scenario framework (required)
Always provide three scenarios:
1. `Bull`: better macro + execution upside
2. `Base`: most likely path under current conditions
3. `Bear`: macro/industry shock + execution shortfall

For each scenario include:
- Key assumptions
- Expected fundamental trajectory
- Implied fair value range
- Probability weight (optional but preferred)

## Margin of safety rule
- Define `Fair Value` range from module triangulation.
- Define `Safety Zone` below fair value (default 15-30% depending on confidence and cyclicality).
- Avoid absolute buy/sell commands.
- Use language: "appears undervalued / fairly valued / stretched" and "requires margin-of-safety discipline".

## Decision policy (how to conclude)
Create an integrated view from:
- valuation outputs (multiples + DCF if valid)
- business quality checklist
- macro/news constraints

If the user is managing a watchlist/portfolio, end with **conditional action framing** suitable for `portfolio-risk-manager`:
- `Trigger to add risk` (what would increase conviction)
- `Trigger to reduce risk`
- `Invalidation` (what would make the thesis wrong)
- `Horizon` (ngắn/trung/dài)

Conclusion label:
- `Attractive` (valuation discount + acceptable quality/risk)
- `Watchlist` (mixed signals, wait for trigger)
- `Caution` (valuation unsupported or risk too high)

## Required report output template
Return exactly these sections in this order:

1. `Executive Summary`
- One paragraph: current valuation stance and why.

2. `What Data Was Used`
- Source, as-of date, statement periods, peer set.

3. `Core Thesis (Bull / Base / Bear)`
- Key drivers by scenario.

4. `Valuation Work`
- Multiples table (current vs peer vs implied)
- DCF summary (if run)
- Sensitivity table

5. `Business Quality Assessment`
- Checklist table with evidence lines.

6. `Risk Register`
- Ranked risks with impact, probability, and monitoring trigger.

7. `Fair Value and Safety Zone`
- Fair value range and margin-of-safety zone with rationale.

8. `Confidence and Gaps`
- Confidence level and exact missing data that could change the view.

9. `Disclaimer`
- Educational analysis only, not personalized investment advice.

## Formatting standards
- Use simple language and explain terms briefly.
- State all critical assumptions explicitly.
- Distinguish facts vs assumptions vs inference.
- Do not hide data gaps; surface them early.
- Keep numbers auditable and unit-consistent (VND bn/trn, %, x).

## Minimal scoring rubric (optional but recommended)
If user asks for ranking within this framework:
- `Valuation` 40%
- `Quality` 35%
- `Momentum/Revision` 15%
- `Risk penalty` 10%

Calibrate per sector and confidence.

## Fail-safe behavior
If data quality is low:
- downgrade confidence
- skip fragile modules (e.g., DCF)
- deliver directional valuation only
- list exact data needed for full valuation

## Trigger examples
- "Value HPG with bull/base/bear and margin of safety."
- "Compare VCB vs BID valuation and explain the: revise.

If 5+ of the 24 patterns are present: very likely AI-generated.
If 10+ patterns: almost certainly AI-generated.

## Adding Personality

Use `references/personality-injection.md` for the full guide. Quick version:

- **Parenthetical asides** — (honestly, this part gets me every time) — 1-3 per 500 words max
- **Tangents** — "Speaking of which...", "That reminds me..." — 1-2 per 1000+ word piece
- **Random thought