PHASE 1: Tri-Stream Foraging (search across three streams)
├─ Ecological stream → context (partners, competitors, supply chains)
├─ Quantitative stream → facts (financial data, reports, metrics)
└─ Narrative stream → story (events, conflicts, statements)

↓ Structured into two levels:

PHASE 2: Dual-Granularity Scrubbing
├─ Atomic Facts (micro) → specific numbers, dates, names
└─ Context Blocks (macro) → trends, connections, frameworks

↓ Plus an expert schema:

PHASE 3: Schema-Guided Planning
└─ Activate a ready-made template from the library (e.g., "Tech breakthrough":
    Market context → Core of the innovation → Players' reaction → Consequences)

↓ Break into atomic blocks:

PHASE 4: Atomic Blocks Sequencing
└─ Each block = one function:
    • Data Anchor → numbers and facts only
    • Narrative Cut-in → a vivid scene
    • Deep Insight → a logical conclusion
    • Stylistic Anchor → an emotional anchor

↓ Generation under constraints:

PHASE 5: Adversarial Prompting
├─ Rhythm Break → alternate very long and very short sentences
├─ Logic Fog → avoid explicit connectors, let the reader complete the logic
└─ Lexical Hedge → mix jargon with conversational language

Important: The full DeepNews system is a parallel multi-agent architecture (Map-Reduce). But the principles work in a regular chat too: saturated context loading, a two-level data structure, an explicit structural schema, adversarial constraints.

⚠️ Where the method shines: Long, deep texts that demand factual accuracy and logical depth (analysis, investigations, breakdowns). Not for short pieces or creative writing.

Task: Write a deep breakdown of a (hypothetical) food delivery market after a major player, FastBite, was forced out in 2024 — why the runner-up, DashCart, didn't capture the whole market, where users went, and what happens next.

Prompt:

I've gathered material for a deep breakdown (final text: 12,000 words).

STEP 1 — DATA STRUCTURE (Dual-Granularity):

MICRO-FACTS (Atomic Facts):
• DashCart market share before/after FastBite's exit: 28% → 41%
• Average order value up 15% (from $28 to $32.20)
• Delivery time rose from 32 to 47 minutes
• 23% of FastBite users switched to pickup
• Logistics investment: DashCart +$210M, GrocerGo +$180M
• Number of couriers: 45k before, 38k after (-15%)

MACRO-CONTEXT (Context Blocks):
• The food delivery market is an oligopoly with high barriers to entry
• FastBite's exit created a shortage of logistics capacity
• Consumers split: premium moved to DashCart, mass market moved to pickup
• Restaurants gained more leverage in commission negotiations
• Background trend: the rise of dark kitchens

STEP 2 — STRUCTURAL SCHEMA:

Use the "Power vacuum" schema:
1. CONTEXT BEFORE THE SHOCK (the old balance of power)
2. THE SHOCK EVENT (what happened)
3. THE BATTLE FOR THE VACUUM (who tried to grab share and why it didn't work)
4. THE NEW BALANCE (who won, who lost, unexpected consequences)
5. WEAK SIGNALS (what comes next)

STEP 3 — ATOMIC BLOCKS:

Break each section into blocks with ONE function:
• Data Anchor Block → numbers and facts only (no interpretation)
• Narrative Cut-in → a vivid scene (courier, user, restaurant)
• Deep Insight Block → a logical conclusion or counterintuitive idea
• Stylistic Anchor → an emotional note or provocation

STEP 4 — ADVERSARIAL CONSTRAINTS:

1. RHYTHM BREAK: Alternate sentences sharply.
   Short (5-8 words). Then a long one, 2-3 lines, with details and facts.
   Short again. Don't settle on an "average length of 15 words" — it's deadly dull.

2. LOGIC FOG: Don't write "therefore", "however", "thus".
   Let the reader make the connection. Put facts side by side — a smart reader
   will see cause and effect on their own.

3. LEXICAL HEDGE: Mix styles.
   "Dark kitchens are grabbing share" and "cloud kitchens are reshaping the market"
   in the same paragraph. Term + conversational equivalent. Don't stay in one register.

NOW WRITE the 12,000-word article.

Result:

The model will produce a text structured around the five sections of the "Power vacuum" schema. Within each section the blocks will alternate: first a Data Anchor packed with numbers (market share, average order value, delivery time), then a Narrative Cut-in with a vivid scene (say, a DashCart courier at rush hour), then a Deep Insight with a counterintuitive conclusion (why DashCart's growing share doesn't mean it won). The rhythm will be jagged: punchy short sentences alternating with long analytical stretches. Explicit connectors ("therefore", "however") will be rare — the logic is completed through the density of facts. The language will mix expert and conversational registers.

RLHF trains LLMs to maximize the probability of a "safe, averaged answer". That produces smoothness — text with no rough edges, no risk, no errors. For short tasks that's fine. For long, deep writing it's a trap. You get "correct nonsense": stylistically even, logically consistent, but with no information density and no insights. Plus hidden hallucinations — the model fills data gaps with plausible fabrications that aren't immediately obvious.

Deep writing requires high-entropy input and low-entropy output. An expert journalist doesn't write linearly. They gather 10x more information than will make it into the piece, activate a ready-made cognitive schema from long-term memory (a narrative template for this type of story), lay the material out along that structure, constantly check themselves with adversarial thinking (the critic/skeptic role), and vary the rhythm and density of the text. DeepNews codifies these implicit processes into explicit instructions.

The Knowledge Cliff (10:1 threshold) is explained through Shannon's information theory. For high-precision output in a high-uncertainty domain (financial markets, for example), the system needs to bring in enough redundancy for triangulation. One fact = unreliable. Three independent sources confirming a fact = reliable. 30,000+ characters of context physically give the model the ability to check every claim against multiple sources. Below that threshold the model is forced to "fill in" — and it fills in with plausible fiction.

Dual-granularity structuring (micro-facts and macro-context kept separate) replicates cognitive psychologist Kintsch's Construction-Integration Model. People store text in two ways: local propositions (specific numbers, dates, names) and the global gist (trends, cause-and-effect links, frameworks). Typical RAG mixes everything into uniform chunks — the model "can't see the forest for the trees" or gets "bones without meat". Separating the two levels gives the model both detail and coherence.

Adversarial constraints (Rhythm Break, Logic Fog, Lexical Hedge) physically break statistical smoothness. The model can't produce an "average 15-word sentence" — the instruction requires alternating short (5-8 words) and long (2-3 lines). That creates burstiness — the statistical roughness characteristic of expert writing. Logic Fog removes explicit connectors ("therefore") — the reader completes the logic themselves, which keeps their attention engaged. Lexical Hedge mixes registers — this creates an estrangement effect and breaks the sterility of AI text.

Prompt control knobs:

1. Input/output ratio (10:1) — drop it to 5:1 for less critical tasks (saves tokens), raise it to 15:1 for maximum accuracy. Below 5:1 it's physically impossible to avoid hallucinations.

2. Schema structure — swap "Power vacuum" for another schema that fits your task ("Tech breakthrough", "Regulatory shock", "Bubble burst"). A schema = a ready-made cognitive template; it lowers the task's difficulty from "open-ended composition" to "filling in slots".

3. Adversarial constraints (intensity) — for formal writing, ease off Logic Fog and Lexical Hedge (keep explicit connectors and a single register). For opinion journalism, crank them up (remove "therefore" entirely, mix styles as much as possible). Rhythm Break always works — monotonous sentence length = losing the reader's attention.

4. Atomic Blocks (types) — add or remove block types for your task. For technical writing, drop the Narrative Cut-in (vivid scenes) and beef up the Data Anchor. For storytelling, add more Narrative and Stylistic Anchors.

Task: {description of the long-form text: topic, length, purpose}

STEP 1 — INPUT DATA STRUCTURE (Dual-Granularity):

MICRO-FACTS (Atomic Facts):
{List the specific data: numbers, dates, names, events, quotes.
One fact per line. At least 20-30 facts.}

MACRO-CONTEXT (Context Blocks):
{Describe the general frameworks, trends, cause-and-effect links,
historical context. 5-7 blocks of 2-3 sentences each.}

STEP 2 — STRUCTURAL SCHEMA:

Use the "{schema name}" schema:
{List the sections of the structure. For example:
1. CONTEXT BEFORE
2. THE SHOCK EVENT
3. PLAYERS' REACTIONS
4. THE NEW BALANCE
5. WHAT'S NEXT}

STEP 3 — ATOMIC BLOCKS:

Break each section into blocks with ONE function:
• Data Anchor Block → facts and numbers only, no interpretation
• Narrative Cut-in → a vivid scene with details
• Deep Insight Block → a logical conclusion or counterintuitive idea
• Stylistic Anchor → an emotional note

Alternate the blocks for rhythm: Data → Narrative → Insight → Data → Stylistic.

STEP 4 — ADVERSARIAL CONSTRAINTS:

1. RHYTHM BREAK: Alternate sentences sharply.
   Short (5-8 words). Long, 2-3 lines. Short again.
   No "average length" — it's deadly dull.

2. LOGIC FOG: {pick a level: "Minimal explicit connectors" or "Remove 'therefore/however' entirely"}
   Let the reader complete the logic through the density of facts.

3. LEXICAL HEDGE: {pick: "Strict single register" or "Mix jargon with conversational language"}
   {If mixing — give examples of terms and their conversational equivalents}

NOW WRITE a {length}-word text.

Placeholders: - {description of the long-form text} — topic, target length (in words/characters), purpose of the text - {micro-facts} — specific data: numbers, dates, names, quotes (20-30+ items) - {macro-context} — general frameworks, trends, connections (5-7 blocks) - {schema name} — a structural template (e.g., "Power vacuum", "Tech breakthrough", "Two-company game") - {structure sections} — list the 4-6 main sections of the schema - {Logic Fog level} — how much to strip explicit connectors ("minimal" or "entirely") - {Lexical Hedge level} — whether to mix registers ("single" or "mixed")

Critically important: Stick to the 10:1 ratio — the input context (micro-facts + macro-context) must be at least 10x longer than the final text. If you're writing 3,000 words, load 30,000+ characters of context. Below that threshold it's physically impossible to avoid hallucinations.

🚀 Quick start — paste into a chat:

Here's a DeepNews template for deep long-form writing. Adapt it to my task: [your task].
Ask me questions to fill in every section.

[paste the template above]

The LLM will ask: what's the topic, how long, which facts you have, which structural schema to use (or it'll offer a few to choose from), and what level of adversarial constraints you need (formal or opinion-style). It'll take the template's structure (dual-granularity, atomic blocks, adversarial constraints) and fit it to your task. You get a working prompt without assembling it by hand.

⚠️ High token cost: The 10:1 ratio means ~200,000 tokens per article (30k characters of context + generation + iterations). On DeepSeek-V3 that's about $0.60-0.90 per article; on GPT-4, about $4-6. For mass content production it may be out of reach. But that's not "overspending" — it's the necessary cognitive price of high-accuracy content.

⚠️ Not for short texts: The method is overkill for posts, short news items, announcements. The Knowledge Cliff only applies to long, complex texts (3,000+ words). For a 300-word post, the "10:1 context" principle is unnecessary and uneconomical.

⚠️ Requires context to exist: If you don't have access to 30k+ characters of relevant context (proprietary data, no materials, a niche topic with no sources), the method doesn't work. This isn't "generating from thin air" — it's "compressing facts". No facts, nothing to compress.

⚠️ Adversarial constraints can hurt formal writing: Logic Fog (removing explicit connectors) and Lexical Hedge (mixing registers) are great for opinion pieces and analysis, but they can wreck academic or legal text, where you need crystal-clear logical links and a single strict register. For formal writing, turn these tactics off or dial them way down.

⚠️ Narrow validation scope: The experimental validation was done only on financial journalism (Chinese tech media). It's unknown how the method performs in other deep-content domains (legal analysis, popular science, investigative journalism in other fields). The principles are universal, but the specific schemas (DNFO-v5) are fintech-specific.

The researchers wanted to understand why LLMs produce shallow content and what to do about it. They took 20 financial topics and generated 5 versions of each, with different amounts of input context: 5k, 10k, 15k, 30k, and 40k characters. 100 articles total. Each was checked for its Hallucination-Free Rate (HFR) — the share of text without factual fabrications (wrong data, nonexistent events, misattribution). It turned out that the curve of HFR versus context volume isn't a straight line but a sigmoid with a cliff.

Up to 10k characters, HFR is ~20% — the model fabricates wholesale. At 15k characters (5:1 relative to article length), HFR is only 40% — half the content is hallucinated! But at 30k characters (10:1), HFR jumps to 85% — a phase transition. At 40k it rises to 90%, but slowly — diminishing returns. Insight: There's a physical context threshold (10:1) below which no amount of prompt engineering will save you from hallucinations. Above the threshold, the task changes its nature — from "composition" to "fact compression".

Then they ran a blind field test for 40 days in the newsroom of a top Chinese tech media outlet. Two teams: the Red Team used GPT-5 (a SOTA model) with an optimized prompt (not plain zero-shot but Persona-based Prompting, to avoid a Straw Man argument). The Blue Team used DeepNews built on DeepSeek-V3-0324 (a previous-generation model, weaker than GPT-5 in parameters). The editors didn't know the articles were written by AI. The metric was the real editorial acceptance rate.

The result was shocking: Red Team (GPT-5) — 10 articles submitted, 0 accepted (0%). Editor feedback: "shallow", "sounds like a press release", "no depth", "too smooth". Blue Team (DeepNews) — 12 articles submitted, 3 accepted (25%). Feedback: "data-rich", "tight logic", "unique insights". Takeaway: Workflow design outweighs model scale in vertical domains. A smaller model with the right architecture beat the SOTA model head-on.

They ran an ROI analysis. DeepNews is more expensive ($0.7 per article vs $0.5 for GPT-5), but GPT-5 with a 0% acceptance rate has an infinite effective cost (any input = total loss). DeepNews with 25% acceptance has an effective cost of $2.8 per accepted article (0.7/0.25). High token consumption isn't "overspending" — it's the necessary cognitive price of expert-level content.

They ran an ablation study, removing Schema Planning and Adversarial Tactics one at a time to understand each module's contribution. They tested on 20 topics × 4 conditions = 80 articles, scoring Logical Coherence, Stylistic Burstiness (rhythmic variability), and Information Density. They found that Schema Planning gives +40% to Logical Coherence — an explicit structural schema sharply improves coherence. Adversarial Tactics give +300% to Burstiness — without them the text is statistically monotonous (a flat distribution of sentence lengths); with them, sharp variability (like a human expert). Information Density rises only when both components are present — they're synergistic.

Workflow is All You Need: Escaping the "Statistical Smoothing Trap" via High-Entropy Information Foraging and Adversarial Pacing Zhongjie Jiang (independent researcher with 10+ years of experience at top-tier financial media)

Key references from the study: - Information Foraging Theory (Pirolli & Card) — information-seeking behavior as biological foraging - Construction-Integration Model (Kintsch, 1988) — two-level representation of text in memory - Schema Theory (Bartlett, 1932) — abstract cognitive structures for organizing knowledge - Agentic Workflow Theory (Andrew Ng) — workflow design > model scale for specialized tasks - "Lost in the Middle" (Liu et al., 2023) — LLMs losing attention in the middle of long context