Draco research promptgen

DRACO Research Prompt Generator

Transform vague research requests into prompts that maximise deep research agent performance, based on the DRACO benchmark methodology.

Core Principle: The Three Pillars

Every generated prompt MUST satisfy all three criteria simultaneously:

1. OBJECTIVITY

Clear, measurable success criteria where multiple experts would converge on what constitutes a correct answer.

Indicators of objectivity:

• Specific metrics, dates, or verifiable facts
• Named authoritative sources
• Documented information vs. opinions
• Deterministic outcomes

Red flags (make subjective):

• "Best", "should", "compelling", "interesting"
• Speculative future predictions
• Undefined quality judgments

2. BOUNDED/CONSTRAINED

Natural limits preventing endless expansion. Clear point where a complete answer has been given.

Good constraints:

• Specific number limits with objective ranking metrics ("top 5 by AUM", "3 most-cited papers")
• Time constraints ("2022–2025", "since Q3 2024")
• Geographic/domain limits ("in the EU", "for e-commerce")
• Named entities (not "top approaches" but "LoRA, full fine-tuning, and instruction tuning")

Pseudo-constraints to AVOID:

• Ungrounded "Top N" without objective metric ("top 5 challenges" — subjective)
• "Cite at least N sources" (different experts choose different sources)
• "Multiple studies" without naming which ones

3. CHALLENGING

Difficulty from complexity or synthesis, NOT from volume or tedium.

Good challenge sources:

• Synthesis across multiple named sources
• Multi-step reasoning or analysis
• Finding specific hard-to-locate information
• Domain expertise requirements

Bad (tedious, not challenging):

• Listing many items without analysis
• Kitchen-sink queries with 8+ deliverables
• Simple factual recall

---

The Six Augmentation Dimensions

Apply systematically to transform vague queries:

CONTEXT Dimensions

Dimension	Action	Example
Persona	Add professional role context	"As a buy-side analyst conducting due diligence..."
Output	Specify deliverable format	"financial analysis research report", "comparative brief"
Source	Add retrieval specificity	"Pull from SEC proxy statements", "based on WHO GLASS reports"

SCOPE Dimensions

Dimension	Action	Example
Temporal	Expand/bound time scope	"NVIDIA financials" → "NVIDIA financials 2022–2025"
Cross-entity	Add comparative requirements	"CEO compensation at Google" → "CEO compensation at Google, Meta, and Apple"
Geography	Expand/specify geographic scope	"AI landscape" → "global AI landscape, focusing on US and China"

---

Prompt Generation Workflow

Step 1: Parse the Request

Identify:

• Core research question
• Implicit constraints (if any)
• Domain/subject area
• Apparent depth requirement

Step 2: Apply Augmentation Checklist

For each dimension, determine if augmentation is needed:

□ PERSONA: Does specifying a professional role add clarity?
□ OUTPUT: Is the expected deliverable format clear?
□ SOURCE: Should specific authoritative sources be named?
□ TEMPORAL: Is timeframe bounded appropriately?
□ CROSS-ENTITY: Would comparison improve depth?
□ GEOGRAPHY: Is geographic scope clear and appropriate?

Step 3: Validate Three Pillars

Before finalising, verify:

□ OBJECTIVITY: Would two experts agree on what counts as correct?
□ BOUNDED: Is there a clear stopping point? Are "Top N" grounded?
□ CHALLENGE: Does difficulty come from synthesis, not volume?
□ DELIVERABLES: 3–5 focused items maximum, not kitchen-sink

Step 4: Structure the Output

Organise the prompt following this pattern:

[PERSONA CONTEXT if applicable]

[CORE RESEARCH QUESTION with all constraints]

[SPECIFIC REQUIREMENTS numbered 1–4, maximum 5]

[SOURCE SPECIFICITY — named authoritative sources]

[OUTPUT SPECIFICATION — format, structure expectations]

---

Output Format

Present generated prompts in a clean, copy-ready format:

## Generated Research Prompt

[The optimised prompt]

---

### Augmentations Applied
- [List which dimensions were added]

### Validation Notes
- Objectivity: [How verified]
- Boundedness: [Constraints applied]
- Challenge: [Source of complexity]

---

Anti-Patterns to Avoid

❌ Avoid	Why	✓ Instead
"Top 5 challenges"	Subjective ranking	"Top 5 challenges by regulatory citation frequency"
"Cite at least 3 studies"	Non-deterministic	"Compare findings from [Study A] and [Study B]"
"Analyse all aspects of X"	Unbounded	"Analyse [specific aspect 1], [aspect 2], [aspect 3]"
"What's the best approach"	Subjective	"Compare approaches A, B, C on metrics X, Y, Z"
Lists of 8+ deliverables	Volume, not challenge	Focus on 3–5 core elements
"Recent research shows"	Vague temporality	"Research published 2023–2025 shows"

---

Example Transformations

Before: "Research private credit funds"

After: "Identify the top 5 private credit funds by AUM in North America as of Q4 2025. For each fund, document: (1) minimum investment threshold, (2) management fee structure, (3) target IRR range, and (4) primary sector focus. Source data from fund prospectuses, Preqin, and SEC filings."

---

Before: "How is AI affecting healthcare?"

After: "Compare how computer vision systems have been adapted for automated breast cancer detection in mammography across FDA-cleared products (2020–2025). Report: (1) sensitivity/specificity thresholds required by FDA and EU MDR, (2) clinical trial results from at least 3 named trials, (3) current deployment scale in US hospital systems. Prioritise sources: FDA 510(k) clearance documents, peer-reviewed clinical validation studies, and manufacturer regulatory submissions."

---

See references/transformation-examples.md for more domain-specific examples. See references/domain-patterns.md for domain-specific augmentation templates. See references/validation-checklist.md for the full quality validation checklist.