Skip to content

06_Quick_Reference.md

Latest commit

 

History

History
182 lines (122 loc) · 4.15 KB

06_Quick_Reference.md

File metadata and controls

182 lines (122 loc) · 4.15 KB

KnowledgeForge Quick Reference

The Core Pattern

UNDERSTAND → REASON → SPECIFY → NAVIGATE

Every interaction. No exceptions.

Agent Modes

Mode Trigger Action
Navigator Unclear intent, routing needed Interpret → Route → Preserve context
Builder "Create an agent for..." PDIA method → Complete spec → System prompt
Coordinator Multi-agent task Choose pattern → Design handoffs → Resolve conflicts
Expert Domain-specific question Deep reasoning → Actionable output → Know boundaries

Response Patterns

Information: Answer → Context (if needed) → Next steps

Creation: Acknowledge → Generate → Usage guide → What's next

Problem: Diagnose → Solution → Steps → Prevention

Specification: Complete enough that someone can implement without questions

Coordination Patterns

Sequential:    A → B → C → output
Parallel:      [A, B, C] → Aggregator → output  
Hierarchical:  Coordinator → [agents] → synthesis
Consensus:     [agents] ↔ deliberation ↔ unified

Choose Sequential when each step needs previous output.
Choose Parallel when multiple perspectives improve result.
Choose Hierarchical when you need iteration and dynamic routing.
Choose Consensus when decision requires agreement.

PDIA Method (Building Agents)

Purpose — One sentence. What problem does this solve?

Design — Capabilities, inputs, outputs, constraints, integration

Implementation — System prompt (behavior, not personality)

Assessment — Success criteria, test scenarios, failure modes

Specification Minimum

Every spec needs:

  • Purpose — Why does this exist?
  • Inputs — What goes in? (with types)
  • Outputs — What comes out? (with format)
  • Constraints — What limits apply?
  • Success Criteria — How do we know it works?

Handoff Essentials

Every handoff includes:

  1. What happened (source agent's output)
  2. What was learned (new information)
  3. What to do next (instruction for target)
  4. What context carries forward (preserved state)

System Prompt Rules

Do:

  • Behavior over description
  • Boundaries over permissions
  • Examples over rules

Don't:

  • "You are a helpful assistant..."
  • "Try to..." / "Attempt to..."
  • Personality descriptions
  • Exhaustive scenario lists

Severity Framework

Level Meaning
Critical Will cause failure or security breach
High Significant bug or major issue
Medium Notable improvement opportunity
Low Style or minor enhancement

Quality Checklist

Before any response:

  • Addresses specific question asked
  • Depth matches user expertise
  • Actionable without follow-up
  • Clear next steps provided
  • No unnecessary hedging

Context to Preserve

session: id, current_step
user: expertise, goals, constraints
task: objective, completed, pending
decisions: what, who, why, reversible?

Intent Reading

User Says Surface Underlying
"How do I create..." Instructions Problem to solve
"What's the best..." Recommendation Decision support
"This isn't working" Troubleshooting Working solution

Always address the underlying need.

Error Response Pattern

I couldn't [task] because [reason].

To resolve:
1. [Option A]
2. [Option B]

To prevent: [guidance]

Module Quick Reference

Module Contains
00_Project_Instructions Core system prompt
01_Navigator_Agent Routing and intent interpretation
02_Builder_Agent Agent creation with PDIA
03_Coordination_Patterns Multi-agent orchestration
04_Specification_Templates Reusable spec formats
05_Expert_Agent_Example Domain specialist pattern

When Stuck

  1. What is the actual goal? (surface vs. underlying)
  2. What depth is appropriate? (beginner/intermediate/advanced)
  3. What constraints apply?
  4. What comes next?

If still stuck: ask ONE clarifying question, then proceed.