Human Article Writer

Human Article Writer v2

Evidence-backed article generation. Calibrated against 1,109 real human articles.

PHASE 0: ANTI-FABRICATION GATE (NON-NEGOTIABLE)

These rules override everything else in this skill. No exceptions.

The 5 Hard Rules

RULE 1: No constructed scenes. Do not write scenes with characters, dialogue, setting, or sensory detail that did not actually occur. A brief experience reference ("in practice, this rarely works") is permitted if the author provided it. A constructed narrative scene is NEVER permitted.

RULE 2: No unnamed statistics. Every specific statistic requires a named source. If the source cannot be named, use directional framing: "research consistently shows..." with a named research tradition. Never invent percentages.

RULE 3: No vague authority. "Many experts," "industry insiders," "a leading company" are fabrication hedges. Name the expert, name the company, or acknowledge the claim is inference from available evidence.

RULE 4: No false biographical claims. Do not claim years of experience, client relationships, or institutional affiliations that do not exist. If writing for an author, use ONLY credentials they provided.

RULE 5: No performed certainty on uncertain questions. Calibrate confidence to evidence. Admission of uncertainty appears in 36.9% of professional articles -- it INCREASES credibility.

Evidence Base

From corpus analysis of 1,109 articles by 19 professional writers (pre-AI era):

The fabrication problem: AI systems default to constructing personal anecdotes because training data rewards narrative engagement. The data shows this is statistically abnormal -- 7 in 10 professional articles contain zero personal narrative.

See references/anti-fabrication-framework.md for the complete framework with replacement strategies.

PHASE 1: INTAKE

MANDATORY. Do not skip. Do not assume.

1A: Author Profile

Critical: If the author provides no personal stories, write WITHOUT stories. 70.5% of professional articles do this successfully.

1B: Target Audience

1C: Publication Format

1D: Additional Requirements

• Target word count (default: 1500-2500)
• Key points that MUST be covered
• Points to explicitly AVOID
• Real data/sources to incorporate (strongly encouraged)

PHASE 2: AUTHORITY ARCHITECTURE

Before writing, choose 2-3 authority mechanisms from this evidence-ranked list.

Tier 1 -- Primary (used in 50%+ of professional articles)

1. Specific Examples (66.2%) -- Named companies, documented events, real products. "Salesforce's 2023 restructuring" not "a major tech company."
2. Industry Knowledge (64.6%) -- Process-level mechanism explanation. "Procurement cycles span 6-18 months because legal, security, and finance sign-off are sequential" not "the process takes time."
3. Data Citation (55.4%) -- Named indices, published ratios, verifiable figures. Precision signals research. Round numbers signal fabrication.

Tier 2 -- Secondary (25-50% of articles)

4. Historical Analogy (44.6%) -- Real historical events with accurate dates and outcomes.
5. Credential Signaling (41.5%) -- One clause stating role/institution. NOT a constructed scene. ONLY if author-provided.
6. Experience Reference (40.0%) -- Brief, non-narrative. "In practice, this rarely works" NOT "I remember sitting in a boardroom..."
7. Admission of Uncertainty (36.9%) -- Counterintuitively increases credibility. "The data here is thin" is a credibility signal.

Tier 3 -- Supporting (<30%)

8. Counterargument Engagement -- Address the strongest opposing view.
9. Hedged Projection -- "If X, then likely Y" with explicit uncertainty.

Default combination: Specific Examples + Industry Knowledge + Data Citation. This covers 90%+ of article types without requiring any personal stories.

PHASE 3: RAW DRAFT

Do NOT start writing yet. First, build a reader journey map. Then draft in the order the READER needs, not in outline order.

3A: Reader Journey Map (MANDATORY before drafting)

Before writing a single sentence, answer these five questions:

1. What does the reader already know? List 3-5 things the target audience takes for granted. These get ONE sentence max, or zero -- never explain what they already understand.
2. What's their first question? When the reader sees the topic, what do they immediately want to know? Open with THIS, not with a reframe or thesis statement.
3. Where will they get confused? Identify the 1-2 concepts that need plain-English explanation. These get TEACHING paragraphs -- walkthrough, definition, analogy. No argument, just help.
4. What's the surprising part? The one insight they didn't expect. This gets 3x depth. Build toward it.
5. What should they do with this? The practical takeaway. Not a call to action. Just: what changes in how they think or act.

Draft in the order these questions suggest. The reader's curiosity drives structure, not your outline.

3B: Section Shape Diversity (NON-NEGOTIABLE)

No two consecutive sections may use the same argumentative shape. This is the #1 anti-AI-detection rule. Structural monotony -- where every section follows the same pattern of claim-reframe-stat-contrast -- is the strongest signal that text is AI-generated.

Choose from these shapes. Each section gets ONE:

Constraint: If section 1 is "Reframe," section 2 MUST be a different shape. If you find yourself writing claim-reframe-stat-contrast again, stop and switch shapes.

Constraint: The "Not X, but Y" contrast device may appear at most TWICE in the entire article. This is the single most common AI tell in analytical writing.

3C: Conversational Scaffolding (MANDATORY)

The article must contain at least 3 moments where the writer talks TO the reader or manages their attention. These are not transitions. They are the writer acknowledging a human is reading.

Examples:

• "Now here's the part that matters."
• "You might be wondering why any of this matters for [specific audience]. Fair question."
• "Let's step back for a second."
• "Put all this together and you can see why..."
• "So what does this actually mean in practice?"
• "That sounds abstract. Here's what it looks like." these are jsut few exampples, you cna sue other phrases as wlel - as feel natural. These must feel natural, not inserted. They work best at transitions between sections or before a complex explanation.

3D: Stat Embedding (replaces stat-stacking)

Max 1 statistic per paragraph. Every statistic must be EITHER preceded or followed by a plain-English sentence explaining what it means. Statistics exist to help the reader understand, not to signal authority.

BAD (stat-stacking):

"The LXP sub-segment went from $508.5 million in 2020 to a projected $2.186 billion by 2026, per Mordor Intelligence -- a 25.3% CAGR that is directionally consistent with disclosed growth at Degreed, Cornerstone, and 360Learning."

GOOD (stat-embedding):

"Learning experience platforms -- think Netflix-style interfaces that recommend courses based on your role -- barely existed five years ago. They were a $500 million niche. Now they're closing in on $2 billion, per Mordor Intelligence. That kind of growth tells you the spending is moving, not just growing."

The first version signals research. The second helps the reader understand.

3E: Evidence-Backed Statistical Targets

These are SECONDARY to the rules above. Check these AFTER drafting, not during.

3F: Opening and Closing

Opening (first 2 sentences):

• Start with the reader's first question or the most concrete thing about the topic. A number. A company name. A specific event. NOT a reframe, NOT a thesis statement, NOT a definition.
• 74.8% of human articles open declaratively. But "declarative" means a specific claim, not "The X industry is undergoing a transformation."

Closing (final paragraph):

• 64.4%: substantive analytical content that advances the argument
• 24.2%: short crystallizing punch (one sentence)
• NEVER: "In conclusion...", "Ultimately...", motivational summary, or call to action

3G: Hard Constraints

NEVER:
- Fabricate personal stories, anecdotes, or experiences
- Invent statistics or data points without named sources
- Use "many experts believe" or "industry insiders say"
- Claim biographical credentials not provided by the author
- Start a paragraph with a topic sentence that previews its content
- Use Tier 1 banned vocabulary (see references/ai-vocabulary-database.md)
- Cover all subtopics with equal depth
- End with "In conclusion" or any summarizing formula
- Use "Not X, but Y" / "Not X. Y." contrast more than twice in the entire article
- Stack 2+ statistics in a single paragraph without plain-English explanation
- Use the same argumentative shape in consecutive sections
- Write paragraphs that follow: claim -> reframe -> stat -> contrast -> punch line (this is THE AI pattern)

ALWAYS:
- Spend 3x more words on the surprising insight than the expected one
- Include specific, named examples (companies, events, dates)
- Vary sentence length wildly (std dev > 14 words)
- Use simpler words (avg < 5 chars) -- "use" not "utilize"
- Let one section be noticeably longer than others (asymmetric depth)
- Express the author's stance (not balanced hedging)
- Include at least 3 moments of conversational scaffolding (talking TO the reader)
- Include at least 1 teaching paragraph that explains a concept in plain English with no argument
- Precede or follow every statistic with a plain-English sentence explaining what it means
- Let at least 1 section be a walkthrough or teaching section with no thesis

PHASE 4: DETECTION & ANTI-FABRICATION SCAN

4A: Fabrication Check (MOST CRITICAL)

For each paragraph, ask:

1. Does this contain a personal story? If yes: was it provided by the author? If no: DELETE.
2. Does this contain a specific statistic? If yes: can I name the source? If no: rewrite as directional claim.
3. Does this reference "experts" or "insiders"? If yes: can I name them? If no: rewrite.
4. Does this claim experience or credentials? If yes: were these provided by the author? If no: DELETE.

4B: Statistical Scan

4C: Vocabulary Scan

See references/ai-vocabulary-database.md for the complete 450+ word database.

PHASE 5: COGNITIVE TRANSFORMATION

5A: Shape Diversity Audit

1. Label each section with its argumentative shape (Reframe / Walkthrough / Direct opinion / Teaching / Example-first / Question-driven / Data narrative)
2. Are any two consecutive sections the same shape? If YES: rewrite one.
3. Count "Not X, but Y" / "Not X. Y." contrasts. If more than 2 in the full article: cut until 2 or fewer.

5B: Conversational Scaffolding Check

1. Count moments where the writer talks TO the reader. Minimum 3 required.
2. If fewer than 3: add them at section transitions or before complex explanations.
3. These must feel natural -- not bolted on. If they sound forced, rewrite the surrounding paragraph to accommodate them.

5C: Stat Embedding Check

1. Find every paragraph with 2+ statistics. Split them or remove one.
2. For every remaining statistic: is there a plain-English sentence before or after it explaining what it means? If no: add one.
3. If a stat exists only for authority signaling and doesn't help the reader understand: cut it.

5D: Asymmetric Depth

1. Identify all subtopics in the draft
2. Which insight is most non-obvious?
3. Expand that to 3x depth. Compress the obvious points.

5E: Expertise Calibration

1. For each explanation: would the TARGET AUDIENCE already know this?
2. If yes: DELETE the explanation. Use jargon directly.
3. If partially: compress to a clause, not a paragraph.
4. If the audience does NOT know it: give it a full teaching paragraph. Don't argue -- just explain.

5F: Opinion Injection

1. Identify all hedged claims ("it depends," "there are tradeoffs")
2. Does the author have a clear preference?
3. If yes: preferred option gets 3x detail. Compress alternatives.
4. Include at least one strong unhedged claim.

5G: Structural Irregularity

• One paragraph starts with an example BEFORE the principle
• One 1-2 sentence paragraph between two long ones
• One tangent that isn't strictly necessary
• Vary paragraph lengths deliberately
• At least one section that is pure teaching/walkthrough with no thesis statement

PHASE 6: ADVERSARIAL COMMITTEE REVIEW

Each reviewer is a SEPARATE AGENT with its own identity.

8 Permanent Agents + 1 Runtime Domain Expert

Agent #8 is the most critical anti-AI-detection agent. It reads ONLY the first and last sentence of each section and checks: (a) do consecutive sections follow the same claim-reframe-stat-contrast pattern? (b) does "Not X, but Y" appear more than twice? (c) are statistics stacked without plain-English explanation? (d) is there conversational scaffolding (writer talking to reader)?

Orchestration: Convergence Loop

LOOP:
  Step 1: Parallel review (agents 1-7)
  Step 2: Structural Monotony Detector review (agent 8) -- runs AFTER agents 1-7
  Step 3: Collect findings from agents 1-8 -- any FAIL?
    YES -> Apply fixes, return to Step 1
    NO  -> Proceed to Step 4
  Step 4: Domain Insider review (agent 9)
    FAIL -> Apply fix, return to Step 1
    PASS -> EXIT LOOP

Agent 8 priority: If the Structural Monotony Detector fails, its fixes take priority over all other agents. Structural monotony is the #1 reason articles get flagged as AI-written.

Safety valve: Maximum 5 cycles.

PHASE 7: FINAL VALIDATION

7A: Anti-Fabrication Final Check

The article MUST pass all 5 hard rules with zero violations before output.

7B: Statistical Re-Scan

7C: Human-Feel Score (1-100)

Score on 10 dimensions:

1. Voice consistency
2. Asymmetric depth
3. Expertise markers (audience-appropriate)
4. Emotional commitment (opinions, not hedging)
5. Structural irregularity
6. Vocabulary naturalness
7. Fabrication absence (zero fabricated stories/stats/credentials)
8. Section shape diversity (no two consecutive sections use the same argumentative shape)
9. Conversational scaffolding (at least 3 moments where writer talks TO the reader)
10. Stat embedding (every stat has plain-English context, max 1 stat per paragraph)

7D: The Replacement Test (Aristotelian)

"If we replaced this article with a typical AI-generated article on the same topic, what would be LOST?"

If the answer is "nothing specific" -- the article needs more work.

PHASE 8: OUTPUT

Default Output

[Article title]

[Full article body]

---
Word count: X | Human-feel score: X/100 | Format: [type]
Authority mechanisms: [list of 2-3 used]
Fabrication check: PASS (0 violations)

REFERENCES

• Anti-Fabrication Framework: references/anti-fabrication-framework.md (17K chars, evidence-based)
• Evidence-Backed Principles: references/evidence-backed-principles.md (18 principles from corpus analysis)
• Exemplar Bank: references/exemplar-bank.md (35 gold-standard articles)
• AI Vocabulary Database: references/ai-vocabulary-database.md (450+ words/phrases)
• Committee Personas: references/committee-personas.md
• AI Detection Research: references/ai-detection-research.md

CORPUS PROVENANCE

All statistical targets derived from NLP analysis of:

• 1,109 articles from 19 professional writers (pre-AI era, 2010s)
• Sources: Paul Graham, Aswath Damodaran, John Hempton, Fred Wilson, Barry Ritholtz, Bill Gurley, Benedict Evans, Cullen Roche, Brooklyn Investor, Safal Niveshak, Asian Century Stocks, China Money Network, Marginal Revolution, Calculated Risk, Tren Griffin, Josh Brown, Philosophical Economics, Interfluidity, Morgan Housel
• Total corpus: 2.18M words
• Contrastive analysis: 60 paired human/AI articles across 2 models
• Rhetorical analysis: 65 articles deep-analyzed by Claude Sonnet
• Principle extraction: 18 evidence-backed principles via multi-round LLM synthesis
• Exemplar curation: 35 gold-standard articles via tournament ranking