Gdr prompt

GDR Prompt Creation

Create focused Gemini Deep Research prompts using NSB v0.4's layered constraint architecture.

Core Concept

GDR's iterative research can cascade into unproductive territory. NSB prevents this through three constraint layers applied in sequence:

Layer	Function	Operates On
Anchors	Immutable facts (budget, hardware, timeline)	Context
Walls	Hard exclusions—entire domains OFF-LIMITS	Search vectors
Gates	Conditional inclusion—PERMITTED ONLY IF	Source evaluation

Key insight: Walls collapse the search space before GDR begins iterating. Gates filter during synthesis. Both are needed.

Workflow

1. Extract Anchors

Ask: "What are the non-negotiable facts about your situation?"

Target categories:

• Physical: Space constraints, hardware specs, location
• Economic: Budget ceiling, ongoing costs acceptable
• Temporal: Deadlines, timeline requirements
• Role/Scope: What problem this solves, what it doesn't need to do

Anchors are facts, not preferences. If it could change based on research findings, it's not an anchor.

2. Derive Walls

For each anchor, ask: "What entire categories does this eliminate?"

Pattern: NO [domain] ([brief justification tied to anchor])

Good walls:

• Eliminate domains that always violate anchors
• Are keyword/topic level (not source-specific)
• Include parenthetical justification to prevent drift

Bad walls:

• Exclude things that sometimes have value
• Are so narrow they're really gates
• Lack anchor justification

3. Construct Gates

For categories with mixed utility, ask: "Under what specific, verifiable conditions would this be valuable?"

Pattern: [Category]: PERMITTED ONLY IF [verifiable condition]

Good gates:

• Have objectively verifiable conditions
• Apply to categories that survived the walls
• Specify quality thresholds or relevance checks

Bad gates:

• Use subjective criteria ("high quality", "relevant")
• Duplicate what walls already exclude
• Can't be evaluated by the model

4. Frame Research Questions

Questions come after all constraints. They assume constraints are enforced.

Effective question types:

• Champions: "Which X best fits these constraints?"
• Anti-portfolio: "What commonly recommended X should be avoided given these constraints?"
• Hidden costs: "What are the non-obvious failure modes?"
• Timing/strategy: "When/how should I execute?"

5. Define Output Structure

Specify concrete deliverables:

• Short lists with rankings
• Avoid lists with justifications
• Checklists for verification
• Price/threshold anchors

Quick Reference

# Deep Research: [Topic]

## I. ANCHORS (Immutable Context)
[Category] Reality:
- [Fact]: [Specific number/constraint]

## II. WALLS (Domain Exclusions)
- NO [domain] ([anchor justification])

## III. GATES (Conditional Inclusion)
- [Category]: PERMITTED ONLY IF [verifiable condition]

## IV. RESEARCH QUESTIONS
1. **[Type]**: [Question assuming constraints enforced]

## V. OUTPUT STRUCTURE
1. **[Deliverable]**: [Contents]

References

For full methodology including worked examples and design principles, see references/nsb-methodology-v0.4.md.

For a blank template, see assets/template.md.

Common Failure Modes

Problem	Symptom	Fix
Walls too broad	Valuable sources excluded	Convert to gate with conditions
Walls too narrow	Research still wanders	Elevate to domain-level exclusion
Gates unverifiable	Model ignores them	Add specific, objective criteria
Missing anchors	Walls feel arbitrary	Ask "why is this excluded?" → find anchor
Questions before constraints	Constraints ignored	Restructure prompt order