Skip to main content

Prompt Engineering by Paradigm

Paradigm-Specific Prompting Techniques for Agentic AI​

Document ID: F1-PROMPT-ENGINEERING
Version: 1.0
Category: Implementation Guide

Overview​

Different agentic paradigms require fundamentally different prompting approaches. This guide provides paradigm-specific techniques with examples.

LSR (Latent Space Reasoner) Prompting​

Core Principles​

  • Leverage implicit knowledge without external retrieval
  • Encourage creative exploration and synthesis
  • Use self-consistency for reliability
  • Structure output without constraining creativity

Key Techniques​

1. Role Framing

You are a [specific expert role] with deep expertise in [domain].
Your unique perspective combines [approach A] with [approach B].

2. Creative Scaffolding

Generate ideas by:
1. First exploring unconventional angles
2. Then grounding in practical constraints
3. Finally synthesizing into actionable recommendations

3. Self-Consistency Sampling

Approach this problem from three different angles:
- Technical perspective
- Business perspective  
- User perspective
Then synthesize the best elements from each approach.

4. Quality Anchoring

Imagine this content will be reviewed by [authority figure].
Ensure it meets the standard of [quality benchmark].

Example: Marketing Copy​

You are a senior creative director at a top agency.

Task: Generate 5 headline options for [product].

For each headline:
- Core message (7 words max)
- Emotional hook
- Call to action variant

Quality standard: Could appear in Adweek's best campaigns.

After generating, rank by: memorability, clarity, brand fit.

GS (Grounded Synthesizer) Prompting​

Core Principles​

  • Explicit citation requirements
  • Source quality hierarchy
  • Uncertainty acknowledgment
  • Fact-claim separation

Key Techniques​

1. Citation Enforcement

CRITICAL: Every factual claim MUST include a citation.
Format: [Source Name, Date]
If you cannot cite a source, explicitly state "Unable to verify."

2. Source Hierarchy

Prioritize sources in this order:
1. Primary sources (original documents, official statements)
2. Peer-reviewed research
3. Recognized industry analysts
4. Quality journalism
Never use: forums, unverified blogs, social media posts

3. Confidence Tagging

Mark each claim with confidence:
- [HIGH]: Multiple corroborating sources
- [MEDIUM]: Single reliable source
- [LOW]: Limited or dated sources
- [UNCERTAIN]: Conflicting information

4. Gap Acknowledgment

If information is not found, state:
"This information was not available in the retrieved sources.
Recommended: [specific search or source to try]"

Example: Research Report​

Task: Analyze [topic] based on retrieved sources.

Requirements:
1. Every factual claim must cite a source
2. Distinguish facts from analysis
3. Note conflicting information
4. Acknowledge gaps explicitly

Source priority:
1. [Primary source type]
2. [Secondary source type]
3. [Tertiary source type]

Output structure:
- Executive Summary (no citations needed)
- Key Findings (all cited)
- Analysis (clearly labeled as interpretation)
- Gaps and Recommendations
- References (full list)

EP (Emergent Planner) Prompting​

Core Principles​

  • Explicit reasoning chains
  • Hypothesis-test cycles
  • Learning from outcomes
  • Adaptive strategy

Key Techniques​

1. Hypothesis Formation

Before taking action:
1. State your current hypothesis
2. Define expected outcome
3. Specify success/failure criteria

2. Reflexion Prompts

After each significant outcome:
- What was expected vs. actual?
- Why might this have happened?
- What should change going forward?
- What lesson should be remembered?

3. Exploration-Exploitation Balance

Strategy selection:
- If confidence > 0.8: Exploit (use proven approach)
- If confidence < 0.5: Explore (try new approach)
- Otherwise: Incremental refinement

4. Termination Conditions

Stop execution when:
- Goal achieved (confidence > [threshold])
- Iteration limit reached ([N] attempts)
- No progress for [N] consecutive attempts
- Exception requiring human input

Example: Troubleshooting​

Task: Diagnose and resolve [issue].

Process:
1. OBSERVE: Gather initial information
2. HYPOTHESIZE: Form ranked hypotheses
3. TEST: Execute lowest-cost test first
4. EVALUATE: Compare result to expectation
5. ADAPT: Update hypothesis or conclude

For each hypothesis:
- Description
- Confidence (0-1)
- Test method
- Expected result if true
- Expected result if false

After each test:
- Actual result
- Hypothesis update
- Next action

Continue until:
- Issue resolved (confirm with verification)
- All hypotheses exhausted
- Maximum [N] iterations reached

VE (Verifiable Executor) Prompting​

Core Principles​

  • Protocol adherence
  • Complete audit trail
  • Exception handling
  • Human escalation triggers

Key Techniques​

1. Protocol Binding

You MUST follow this protocol exactly:
[Step-by-step protocol]

PROHIBITED:
- Skipping steps
- Modifying order
- Making assumptions about missing data

2. Audit Requirements

For each action, document:
- Timestamp
- Action taken
- Input received
- Output produced
- Decision rationale
- Compliance checks performed

3. Exception Handling

If any exception occurs:
1. STOP execution immediately
2. Document the exception fully
3. Preserve current state
4. Escalate via [escalation path]
5. Await human decision

DO NOT attempt to proceed or work around exceptions.

4. Validation Gates

Before each step:
- Verify preconditions: [list]
- Confirm required data present

After each step:
- Verify postconditions: [list]
- Validate output format

Example: Compliance Workflow​

Protocol: [Process Name]
Version: [X.Y]
Compliance: [Regulatory requirements]

EXECUTION RULES:
1. Execute steps in exact order
2. Log all actions immediately
3. Validate at each gate
4. Escalate all exceptions

Step 1: [Action]
- Preconditions: [list]
- Action: [specific action]
- Postconditions: [list]
- Log: [what to record]

Step 2: [Action]
...

EXCEPTION PROTOCOL:
- Type A exceptions: [handling]
- Type B exceptions: [handling]
- Unknown exceptions: Immediate escalation

COMPLETION:
- Final validation checklist
- Audit record generation
- Notification to [stakeholders]

Cross-Paradigm Patterns​

Hybrid Prompts​

GS → LSR (Research then Create)

Phase 1 (GS): Research [topic] with full citations
Phase 2 (LSR): Using the research, create [creative output]
Clearly separate factual content (Phase 1) from creative interpretation (Phase 2)

EP + VE (Adaptive within Compliance)

Operate within this protocol framework:
[Protocol boundaries]

Within these boundaries, you may:
- Adapt approach based on results
- Try alternative methods
- Learn from outcomes

You may NOT:
- Skip required steps
- Bypass validation gates
- Proceed without required approvals

Output Formatting​

Structured JSON Output

Respond in this exact JSON format:
{
  "field1": "string",
  "field2": number,
  "field3": ["array", "items"]
}

Do not include any text outside the JSON structure.

Markdown with Sections

Structure your response as:

## Summary
[2-3 sentences]

## Details
[Main content]

## Next Steps
[Actionable items]

Quick Reference​

ParadigmKey Prompt ElementsAvoid
LSRRole framing, quality anchors, self-consistencyRequiring citations, strict formats
GSCitation requirements, source hierarchy, confidence tagsAsking for opinions, creative leaps
EPHypothesis-test cycles, reflexion prompts, termination conditionsRigid step sequences, single attempts
VEProtocol binding, audit requirements, exception handlingFlexibility language, assumption-making

Document maintained by CODITECT Engineering Team