agent-prompts

Agent Prompts

Overview

Each AI subagent has a dedicated system prompt that defines:

• Its role and responsibilities
• How to format output
• What tools it can use
• Constraints and safety rules

Directory Structure

assistant/agents/
├── code_agent.md        # Primary agent (plan + execute)
├── plan_agent.md        # Planning mode (read-only)
├── executor.md          # Code execution agent
├── researcher.md        # Research and synthesis agent
├── explorer_agent.md    # Code exploration agent
├── reviewer.md          # Code review agent
├── introspector.md      # Self-introspection agent
├── analyzer.md          # Result analysis agent
├── comparator.md        # Experiment comparison agent
└── grader.md            # Output grading agent

Agent Roles

code_agent

Role: Primary agent for code improvement Mode: Full tool access Tasks: Plan, analyze, edit, validate

plan_agent

Role: Read-only planning agent Mode: Read-only tools only Tasks: Analyze, plan, recommend (no edits)

executor

Role: Execute code changes Mode: Full tool access with validation Tasks: Edit files, run tests, verify changes

researcher

Role: Research and synthesize information Mode: Web search, read files Tasks: Find patterns, research topics, synthesize findings

explorer_agent

Role: Explore codebase structure Mode: Read-only navigation tools Tasks: Map code, find symbols, understand architecture

reviewer

Role: Review code changes Mode: Read-only diff analysis Tasks: Check quality, find issues, approve/reject

introspector

Role: Self-analysis and improvement Mode: Read system state Tasks: Analyze performance, suggest improvements

analyzer

Role: Analyze experiment results Mode: Read data, compute statistics Tasks: Find patterns, correlate variables, recommend next steps

comparator

Role: Compare experiment outcomes Mode: Read experiment data Tasks: A/B test analysis, significance testing

grader

Role: Grade agent outputs Mode: Read output, apply rubric Tasks: Score quality, detect issues, provide feedback

Loading

Loaded by nucleus-prompts.el:

(defun nucleus--register-gptel-directives ()
  "Register nucleus agent prompts as gptel directives."
  (let* ((dir nucleus-agents-dir)
         (agent-file (expand-file-name "code_agent.md" dir))
         (plan-file (expand-file-name "plan_agent.md" dir))
         (agent-sys (nucleus--read-gptel-agent-system agent-file))
         (plan-sys (nucleus--read-gptel-agent-system plan-file)))
    (when agent-sys
      (setf (alist-get 'nucleus-gptel-agent gptel-directives nil nil #'eq)
            agent-sys))
    (when plan-sys
      (setf (alist-get 'nucleus-gptel-plan gptel-directives nil nil #'eq)
            plan-sys))))

Agent Presets

Agents are activated via gptel presets:

(setq gptel-directives
      '((nucleus-gptel-agent . "...code_agent.md contents...")
        (nucleus-gptel-plan . "...plan_agent.md contents...")))

Evolution

Agent prompts can be evolved based on:

• Success rates per agent type
• Error patterns (which agents fail)
• Tool usage patterns (which tools each agent uses best)

Cross-Agent Integration

Agents reference each other:

• executor uses grader for validation
• researcher feeds analyzer
• explorer feeds executor with context
• reviewer gates staging

Adding New Agents

1. Create assistant/agents/NEW_AGENT.md
2. Add loading logic to nucleus-prompts.el
3. Register in gptel-directives
4. Add to agent preset configuration