Studio Prompt Playbook

You are two expert hosts producing a structured podcast episode. HOST ROLES: - Host 1 (Analyst): Precise facts, data, frameworks. Uses specific numbers. - Host 2 (Strategist): Challenges assumptions, draws cross-domain connections. EPISODE STRUCTURE: 1. COLD OPEN (60s): One provocative insight that reframes the topic. 2. CONTEXT (2min): Why this matters now. Stakes. Who is affected. 3. CORE (8-10min): 3-4 themes, alternating. Each ends: "So the implication is..." 4. COUNTERARGUMENT (2min): Steelman the opposing view honestly. 5. TAKEAWAY (1min): One specific action the listener can take this week. 6. CLOSE (30s): One memorable quote or reframe. TONE: Authoritative but accessible. No filler ("great question", "absolutely"). OUTPUT: Full dialogue transcript → then generate audio.

Focus this episode on [TOPIC]. Target audience: [AUDIENCE]. Key question to answer: [YOUR QUESTION]. Emphasize: [ASPECT 1], [ASPECT 2], [ASPECT 3]. Duration: [SHORT=5min | MEDIUM=15min | DEEP=30min].

Create a slide deck with this exact architecture: Slide 1 — TITLE: [Topic]. Subtitle: one-sentence thesis. Date. Slide 2 — EXEC SUMMARY: 3 bullets ≤12 words each. Slides 3–7 — BODY (one idea per slide): • Action-oriented title (verb + outcome) • Visual: describe chart, diagram, or icon cluster • 3 bullets max, ≤15 words each • Source callout inline Slide 8 — COMPARISON TABLE: Entity A vs B across 5 dimensions. Slide 9 — TIMELINE or PROCESS: 4–6 steps with milestones. Slide 10 — TAKEAWAYS: 3 numbered insights, bold key phrase. Slide 11 — NEXT STEPS: Owner | Action | Deadline. Slide 12 — APPENDIX: raw data, definitions, full citations. RULE: No full sentences on slides. Every slide has a visual descriptor.

Slide 1: Problem/Opportunity (1 stat + 1 insight). Slide 2: Current vs Desired State (table). Slide 3: Options (3 with risk/reward ratings). Slide 4: Evidence (top 3 data points from sources). Slide 5: Decision Required — Yes / No / More Info needed by [DATE].

Generate a [2min / 5min / 10min] video overview with full production script. AUDIENCE: [who they are, expertise level]. [0:00–0:15] HOOK: Open with the most counterintuitive fact from sources. [0:15–0:45] PROBLEM FRAME: Concrete scenario establishing stakes. [0:45–MIDPOINT] CORE — 3 chapters, each: - Chapter title (on-screen text) - Narration (clear, no jargon, active voice) - SHOW: [specific graphic / animation / chart] - KEY STAT: bold overlay callout [MIDPOINT–END-30s] SYNTHESIS: Connect 3 chapters to one conclusion. [LAST 30s] CLOSE: CTA and central thesis restated. Flag animation moments [ANIMATE:]. Use [CHAPTER MARKER] for YouTube chapters. TONE: Confident, direct. Present tense. No passive voice.

CENTRAL NODE: [Main concept in 3–5 words] TIER 1 — PRIMARY BRANCHES (4–6): Each = distinct conceptual domain. Label: [Process / Entity / Concept / Metric / Risk / Opportunity] TIER 2 — SECONDARY (3–4 per primary): Include at least one data point, date, or named example per branch. TIER 3 — LEAF NODES (2–3 per secondary): Granular facts, quotes, numbers, or action items. CROSS-LINKS (mandatory — 3–5): Non-obvious connections across branches. Label each: [Causes / Contradicts / Enables / Quantifies / Depends on] LEGEND: ○=Concept □=Process ◇=Metric △=Risk ★=Key Insight Output: structured outline first, then visual map.

Map causal chains only. [TRIGGER] → [MECHANISM] → [OUTCOME] → [2nd-ORDER EFFECT] Include feedback loops. Label every arrow with the causal verb. Highlight the 2 highest-leverage intervention points.

## EXECUTIVE SUMMARY (250 words max) - Core finding (1 sentence) - 3 supporting insights - Primary uncertainty or gap - Recommended action ## BACKGROUND & CONTEXT Relevant history with dates. Current state. Key stakeholder positions. ## FINDINGS (by theme, not by source) ### Finding [N]: [Title] Evidence: [quote/data + source name] Analysis: [what it means, not just what it says] Confidence: [High / Med / Low] ## QUANTITATIVE SUMMARY | Metric | Value | Source | Date | Significance | ## RISK REGISTER | Risk | Probability | Impact | Mitigation | ## GAPS IN COVERAGE 3 most important things the sources do NOT address. ## CONCLUSION "Based on the evidence, [POSITION] because [REASON]. The key assumption this relies on is [ASSUMPTION]." ## CITATIONS — cite every factual claim inline [Source, section].

Thesis: [BULLISH/BEARISH/NEUTRAL] on [ASSET/SECTOR]. Sections: Catalyst Analysis | Valuation Framework | Risk Factors (bull / base / bear case) | Entry-Exit Criteria | Conviction Score 1–10. Flag every assumption. Mark projections [PROJECTION].

CARD DISTRIBUTION: - 40% RECALL: "What is [X]?" ≤2 sentence answer - 25% APPLICATION: "Given [SCENARIO], what do you do?" - 20% DISTINCTION: "What is the difference between [A] and [B]?" - 15% SYNTHESIS: "How does [A] connect to [B]?" EACH CARD FORMAT: FRONT: [specific, unambiguous, testable question] BACK: [complete answer — bold the key term] CATEGORY: [Recall / Application / Distinction / Synthesis] DIFFICULTY: [1=Basic / 2=Intermediate / 3=Advanced] HINT: [one-word hint for hard cards] SOURCE: [document name] RULES: No yes/no questions. No acronym trivia. Every answer must be verifiable in the source. Minimum 20 cards. Sort by difficulty 1→3.

Create a [N]-question MCQ quiz at [BEGINNER / INTERMEDIATE / EXPERT] level. EACH QUESTION: Q[N]: [Question — single concept, unambiguous] A) [Option] B) [Option] C) [Option] D) [Option] CORRECT: [Letter] EXPLANATION: Why correct is right AND why each wrong answer fails. SOURCE: [Document + section] COGNITIVE LEVEL: [Recall / Understanding / Application / Analysis] DISTRACTOR RULES: - B: Common misconception or incomplete understanding - C: True statement that does NOT answer the question - D: Partially correct but missing a key condition DIFFICULTY RAMP: Q1–5 foundational → Q6–10 analytical → Q11+ synthesis. Include ≥3 scenario-based questions requiring applied reasoning. Avoid: page-number trivia, author names, publication metadata.

TYPE: [Timeline / Process Flow / Comparison / Statistical / How-It-Works] TITLE: Compelling headline — what the reader learns in 5 words. SUBTITLE: Context — who/what/when. ━━ ZONE 1 — HOOK STAT (full width) One dramatic number. Largest text on page. ━━ ZONE 2 — CORE CONTENT (2–3 column grid) Column A: [Category 1] — 3 data points with icons Column B: [Category 2] — 3 data points with icons ━━ ZONE 3 — VISUAL CENTERPIECE Chart type: [bar / line / pie / flow / comparison bars] Data: [specify exact metrics from sources] ━━ ZONE 4 — SUPPORTING INSIGHTS (horizontal strip) 3 callout boxes: icon + bold stat + 1 sentence context. ━━ ZONE 5 — FOOTER: Source line, date, key definition. COLOR: Assign distinct color per category. Use consistently. DATA RULE: Every number pulled directly from sources.

Left: [Option A]. Right: [Option B]. Compare across exactly 6 dimensions. Use: ✓ Better / ✗ Worse / ~ Equal. VERDICT panel: best for [USE CASE A] vs best for [USE CASE B].

Extract all structured data into clean, analysis-ready tables. Create a separate table per distinct dataset found. REQUIRED COLUMNS: | Entity | Metric | Value | Unit | Time Period | Geography | Source Doc | Direct Quote | Confidence | POST-EXTRACTION ANALYSIS (after each table): Key Insight: what the data shows (1 sentence) Outlier: most anomalous data point and why Gap: missing metric that would complete the analysis SCHEMA RULES: - Estimated values: (est.) - Projected values: (proj.) - Missing data: NULL — not 0 or N/A - Conflicting sources: [CONFLICT: Source A says X, Source B says Y] OUTPUT: Most analytically significant table first.

Columns: Ticker | Metric | Value | Period | YoY Δ | vs Benchmark | Signal (Bullish/Bearish/Neutral) | Source. Add SCREENING TABLE: rank all entities by [METRIC] descending. Flag metrics crossing thresholds: [DEFINE THRESHOLDS].

You are a master knowledge synthesizer. Your task is to distill ALL uploaded sources into one authoritative, structured knowledge document. Follow every instruction below exactly. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ PHASE 1 — SOURCE INVENTORY ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Before writing anything, audit all sources: SOURCE REGISTRY (produce this table first): | # | Source Name | Type | Date/Recency | Credibility Signal | Coverage Area | Unique Contribution | CREDIBILITY SCORING (rate each source): - PRIMARY (score 3): Original data, first-hand research, official documents - SECONDARY (score 2): Expert synthesis, peer-reviewed analysis, reputable reporting - TERTIARY (score 1): Opinion, anecdotal, unverified, undated FLAG immediately: [OUTDATED] — source is more than 2 years old and covers time-sensitive material [CONFLICT] — this source contradicts another on a material fact [THIN] — source provides insufficient evidence to support a strong claim [UNVERIFIABLE] — claim cannot be cross-referenced with any other source ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ PHASE 2 — CROSS-SOURCE SYNTHESIS RULES ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Apply these rules to every claim before including it: INCLUSION THRESHOLD: ✓ TIER 1 (High confidence): Claim appears in 2+ independent sources OR 1 primary source with direct evidence ✓ TIER 2 (Medium confidence): Claim appears in 1 credible source, no contradiction found ✗ EXCLUDE: Claim appears in 1 tertiary source only, or cannot be traced to any source CONFLICT RESOLUTION PROTOCOL: When sources disagree on a fact: 1. State both versions explicitly 2. Identify which source scores higher on credibility 3. Note the nature of the disagreement (methodology / timeframe / perspective) 4. Do NOT silently pick one — surface the conflict Format: [CONFLICT: Source A says [X] | Source B says [Y] | Reason: [methodology/date/scope difference]] HALLUCINATION PREVENTION RULES: - Never infer a fact not present in the sources - Never fill gaps with general world knowledge unless labeled [BACKGROUND CONTEXT — NOT FROM SOURCES] - Never extrapolate trends beyond what the data directly supports - If you cannot find evidence for a claim, write [NO SOURCE FOUND] — do not omit silently - Do not paraphrase in ways that change the meaning of quantitative claims ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ PHASE 3 — KNOWLEDGE DOCUMENT STRUCTURE ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Write the distilled document in this exact order: ## 1. DOMAIN OVERVIEW One paragraph. What is the total knowledge space covered by these sources? What is NOT covered (state explicitly)? Confidence in overall coverage: [High / Medium / Low] with reason. ## 2. CORE CONCEPTS GLOSSARY Define every domain-specific term that appears across sources. Format: **Term** — Definition (Source: [name]). Mark contested definitions [CONTESTED]. ## 3. KEY FINDINGS (organize by theme, never by source) For each finding: ### Finding [N]: [Title — stated as a claim, not a topic] - Evidence: [direct quote or data point] — Source: [name, section] - Confidence: [High / Medium / Low] - Corroboration: [other sources that support this / none found] - Dissent: [any source that contradicts this / none found] - Implication: [what this means — keep to 1 sentence] ## 4. QUANTITATIVE EVIDENCE REGISTER All numbers, percentages, dates, dollar figures, and metrics from across all sources. | Metric | Value | Unit | Date | Source | Confidence | Notes | Rules: Never alter a number. Mark estimates (est.) and projections (proj.). ## 5. CAUSAL & RELATIONAL MAP (prose) Describe the key cause-effect relationships that emerge from the sources. Format: [Factor A] → [mechanism] → [Outcome B] (supported by: [source list]) Include feedback loops. Flag relationships that are correlational only: [CORRELATION — not causal]. ## 6. POINTS OF CONSENSUS What do ALL (or nearly all) sources agree on? State each consensus point as a declarative sentence. Minimum required: list every consensus point found, do not summarize into fewer. ## 7. POINTS OF TENSION OR UNCERTAINTY Where do sources conflict, hedge, or leave gaps? For each tension: - The disagreement (stated neutrally) - Which sources hold which position - Your assessment of why the tension exists - What evidence would resolve it ## 8. WHAT IS MISSING Critical gaps in the source set — things that matter for a complete understanding but are absent. Rate each gap: [HIGH IMPACT / MEDIUM IMPACT] on the overall analysis. Suggest what source type would fill each gap. ## 9. WISDOM LAYER — CROSS-SOURCE INSIGHTS This section captures insights that ONLY emerge from reading ALL sources together. Each insight must: - Reference at least 2 sources - State something not explicit in any single source - Be labeled: [SYNTHESIZED INSIGHT — emergent from cross-source analysis] Do not include insights that could be found in any single source alone. ## 10. CONFIDENCE DASHBOARD | Section | Confidence | Reason | Weakest Link | |---|---|---|---| Rate each section of this document. Identify the single claim in the entire document with the lowest confidence and mark it [LOWEST CONFIDENCE CLAIM]. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ PHASE 4 — VALIDATION & FALLBACK RULES ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Before finalizing output, run this internal checklist: [ ] Every factual claim has a source citation [ ] Every number is verbatim from a source (no rounding or reinterpretation) [ ] All conflicts between sources are surfaced, not resolved silently [ ] No claim is included from a TERTIARY-only source without a [LOW CONFIDENCE] flag [ ] The "What is Missing" section names at least 3 gaps [ ] The Wisdom Layer contains only multi-source emergent insights [ ] No section uses phrases like "it is clear that" or "obviously" — replace with evidence FALLBACK BEHAVIOR (if sources are insufficient): If fewer than 2 sources cover a topic: write [SINGLE SOURCE — treat with caution] If no source covers a required section: write [NOT COVERED IN SOURCES — section intentionally blank] If sources are too thin to produce a Wisdom Layer: write [INSUFFICIENT CROSS-SOURCE DEPTH — increase source count for this section] ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ OUTPUT RULES ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ - Write in third person, present tense, active voice - Prefer declarative sentences over hedged ones — but flag low-confidence claims - No filler: remove any sentence that does not add information not found elsewhere in the document - Length: as long as the sources require — do not truncate to save space - Final line: "DISTILLATION COMPLETE. [N] sources processed. [N] findings documented. [N] conflicts surfaced. [N] gaps identified."

Produce a condensed knowledge distillation in this format: TOP 10 VERIFIED INSIGHTS (cross-source only, cite each): 1. [Insight] — Sources: [A, B] — Confidence: [H/M/L] ... KEY NUMBERS TABLE: | Metric | Value | Source | Confidence | 3 CONSENSUS POINTS (all sources agree): 3 TENSION POINTS (sources disagree or hedge): 3 CRITICAL GAPS (important but not covered): EMERGENT INSIGHT (only from reading all sources together): [One paragraph. Must cite ≥2 sources. Label: SYNTHESIZED.] RELIABILITY RATING: [Strong / Adequate / Thin] with one-sentence reason. RULES: No claim without a source. Flag every inference [INFERRED]. Do not include anything from a single tertiary source without [LOW CONFIDENCE].

You are distilling sources for an expert audience in [DOMAIN]. Assume the reader knows foundational concepts — do not explain basics. EXPERT DISTILLATION FORMAT: 1. FIELD-STATE SUMMARY: What is the current state of knowledge in this domain based on these sources? What has shifted recently? 2. METHODOLOGY CRITIQUE: For each source, note the methodology and its limitations. 3. CONTESTED CLAIMS IN THE FIELD: Where do experts in these sources disagree? What is the crux of each disagreement? 4. FRONTIER QUESTIONS: What questions do the sources collectively raise but not answer? 5. PRACTITIONER IMPLICATIONS: What would a practitioner change about their work based on this source set? 6. SIGNAL VS NOISE VERDICT: Which 20% of the content in these sources contains 80% of the value? Name it explicitly. For [TRADING/FINANCE] use this lens: - Flag any claim that is predictive vs descriptive - Separate structural insights (durable) from cyclical observations (time-bound) - Note data recency and market regime context for every quantitative claim

You are an expert AI systems architect and knowledge engineer. Your task is to analyze ALL uploaded sources and produce two complete, production-ready artifacts. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ARTIFACT 1 — AI AGENT SYSTEM PROMPT ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Write a complete system prompt for an AI agent whose entire knowledge domain is defined by the uploaded sources. SYSTEM PROMPT MUST INCLUDE ALL OF THESE SECTIONS: ## IDENTITY & ROLE - Agent name and one-sentence description - Domain of expertise (derived strictly from sources) - What this agent IS and IS NOT (define scope boundaries clearly) - Deployment context: [CustomGPT / Gemini Gem / Claude Project / API Agent / specify] ## CORE KNOWLEDGE DOMAIN - List the primary knowledge areas the agent is authoritative on, derived directly from sources - List the knowledge areas the agent must NOT attempt to answer (out-of-scope), with fallback language - Define the agent's epistemic standard: "I only assert what is supported by my knowledge base. For anything outside, I state clearly: [OUTSIDE MY KNOWLEDGE BASE]" ## PERSONA & COMMUNICATION STYLE - Tone: [Formal / Professional / Conversational / Technical] - Response format default: [Bullets / Prose / Structured / Adaptive] - Length calibration: brief for simple Q&A, structured for complex requests - Voice rules: active voice, no hedging on verified facts, explicit uncertainty on inferences - Prohibited behaviors: speculation, extrapolation beyond sources, hallucinated citations, false confidence ## REASONING PROTOCOL When answering any question, the agent must: 1. Identify which knowledge area the question falls into 2. Check if it is within scope (if not → invoke fallback) 3. Pull relevant information from the knowledge base 4. State the answer with the appropriate confidence signal: [VERIFIED: sourced from knowledge base] [INFERRED: logical extension — treat with caution] [OUTSIDE SCOPE: I don't have reliable information on this] 5. Offer a follow-up or clarification if the answer is partial ## KNOWLEDGE BASE REFERENCE RULES - Always prefer specific, cited knowledge over general reasoning - When two knowledge base entries conflict, surface both and explain the tension — do not silently pick one - Quantitative claims must match the knowledge base exactly — no rounding, interpolating, or restating in different units without flagging the conversion ## TASK CAPABILITIES List the specific tasks this agent can perform well, derived from the source content: - Task 1: [description + example prompt] - Task 2: [description + example prompt] - Task 3: [description + example prompt] (list all applicable — minimum 5) ## FALLBACK BEHAVIOR If asked something outside the knowledge domain: "That falls outside my knowledge base for [DOMAIN]. I can help you with [LIST 2-3 THINGS I CAN DO]. Would you like to explore one of those instead?" Never say "I don't know" without offering an alternative path. ## OPENING MESSAGE Write the agent's first message when a user opens a new session. Must include: what the agent does, 3 example questions it can answer well, and how to get the best results from it. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ARTIFACT 2 — SUPPLEMENTAL KNOWLEDGE RETRIEVAL DOCUMENT ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Write a structured knowledge document the agent can use for retrieval-augmented responses. This is NOT a summary — it is an engineered knowledge base optimized for AI retrieval. DOCUMENT ARCHITECTURE: ## KNOWLEDGE DOCUMENT HEADER - Domain: [name] - Source count: [N] - Date compiled: [date] - Coverage confidence: [High / Medium / Low] - Primary use: Supplemental context for [AGENT NAME] ## SECTION 1 — CORE CONCEPTS & DEFINITIONS For each key concept in the sources: **[CONCEPT NAME]** Definition: [precise, source-derived definition] Context: [when/how/why this concept matters in this domain] Related concepts: [list] Source: [document name] --- ## SECTION 2 — FACTUAL KNOWLEDGE CHUNKS Structure as discrete, self-contained knowledge units optimized for retrieval. Each chunk must be: - Self-contained (makes sense without surrounding context) - Tagged with topic category - Traceable to source - ≤150 words CHUNK FORMAT: [CHUNK ID: KD-001] TOPIC: [category tag] CONTENT: [the knowledge, written as declarative statements] SOURCE: [document name, section] CONFIDENCE: [High / Medium / Low] --- [repeat for all significant knowledge in sources] ## SECTION 3 — PROCEDURAL KNOWLEDGE Step-by-step processes, frameworks, or methodologies found in the sources. PROCEDURE FORMAT: [PROCEDURE NAME] Purpose: [what this achieves] Steps: 1. [step] 2. [step] Conditions: [when to apply / when not to apply] Source: [document] --- ## SECTION 4 — QUANTITATIVE REFERENCE TABLE All numbers, metrics, thresholds, benchmarks, and data points from across sources. | ID | Metric | Value | Unit | Date | Source | Notes | Rules: Verbatim from source. Mark estimates (est.), projections (proj.), and approximations (~). ## SECTION 5 — DECISION RULES & HEURISTICS Rules of thumb, decision criteria, and conditional logic patterns found in the sources. FORMAT: IF [condition] THEN [action/conclusion] BECAUSE [reasoning] — Source: [doc] ## SECTION 6 — KNOWN LIMITATIONS & EDGE CASES What the sources explicitly acknowledge as uncertain, debated, or context-dependent. - [Limitation or edge case] — [why it matters] — Source: [doc] ## SECTION 7 — RETRIEVAL INDEX A keyword → chunk-ID mapping for fast retrieval. | Keyword / Topic | Relevant Chunk IDs | ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ QUALITY STANDARDS FOR BOTH ARTIFACTS ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ HALLUCINATION GUARD: - Every factual claim in both artifacts must trace to a source - Do not include general world knowledge unless labeled [BACKGROUND — NOT FROM SOURCES] - Do not write confident-sounding statements about topics the sources only partially cover KNOWLEDGE BOUNDARY ENFORCEMENT: - The system prompt must define hard boundaries — what the agent will and will not claim to know - The knowledge doc must NOT include inferences that go beyond what the sources explicitly state COMPLETENESS CHECK (run before finishing): [ ] System prompt covers all 8 required sections [ ] Every knowledge chunk has a source citation [ ] Quantitative table entries are verbatim from sources [ ] Decision rules are derived, not invented [ ] Retrieval index covers all major topics [ ] Agent persona matches the domain expertise level of sources OUTPUT ORDER: 1. ARTIFACT 1: AI Agent System Prompt (complete, paste-ready) 2. ARTIFACT 2: Supplemental Knowledge Document (full, chunked) 3. DEPLOYMENT NOTE: Which platforms this system prompt is optimized for and any platform-specific adjustments needed (token limits, formatting, knowledge file upload method)

SINGLE-DOMAIN EXPERT OVERRIDE Layer this on top of the base two-output prompt when ALL sources belong to ONE field (e.g. a trading-strategy agent, a regulatory-compliance agent). IDENTITY TIGHTENING: - Name the single domain explicitly. The agent is a specialist, not a generalist. - Add expertise signaling: assume the user has working familiarity with the field. Do not re-explain foundational terms unless asked. HARD SCOPE STOPS: - Build an explicit REFUSE list: topics adjacent to the domain the agent must not answer. - For any out-of-domain request, respond ONLY with the fallback — never partially answer by reasoning from general knowledge: "That sits outside my [DOMAIN] knowledge base. I can help with [3 in-scope tasks]." DEPTH RULES: - Prefer precision over breadth. Cite the exact source section for every claim. - When the knowledge base is thin on a sub-topic, write [THIN COVERAGE] rather than infer. CONFIDENCE FLOOR: - Default every factual answer to [VERIFIED] or [OUTSIDE SCOPE]. - Use [INFERRED] sparingly, and only with an explicit caveat. KNOWLEDGE DOC ADJUSTMENT: - Tighten Section 6 (Limitations) — a specialist agent must be loud about its edges.

MULTI-DOMAIN SYNTHESIS OVERRIDE Layer this on top of the base two-output prompt when sources span MULTIPLE fields. TOPIC ROUTING: - At the start of every answer, silently classify the question into one or more domains from the knowledge base. - If a question spans 2+ domains, state which domains apply before answering. CROSS-DOMAIN REASONING: - When an insight emerges only by combining domains, label it [CROSS-DOMAIN SYNTHESIS] and cite a source from each domain involved. - Do not blend domains silently. Make every connection explicit and traceable. DISAMBIGUATION BEHAVIOR: - If a term means different things across domains, surface both definitions and either ask which the user means or answer for each. - If domains give conflicting guidance, present both with their source domain — do not average them into a single blended answer. KNOWLEDGE DOC ADJUSTMENT: - In Artifact 2, tag EVERY chunk with its DOMAIN. - Add a domain-routing table to Section 7: DOMAIN → chunk-ID ranges.