Subagent Prompting

Sub-Agent Prompting Template

Overview

This skill provides a standardized template for invoking sub-agents with deterministic inputs and predictable outputs. Use this when:

• Orchestrating specialist sub-agents (code auditor, test auditor, etc.)
• Designing multi-agent workflows with conditional branching
• Ensuring consistent prompt structure across sub-agent invocations

4-Part Template (Required)

Every sub-agent invocation MUST include all four parts. Incomplete prompts lead to unpredictable behavior.

GOAL (What Success Looks Like)

State the high-level objective, not just the action. Good goals are outcome-focused.

## GOAL

[Describe the desired end state, not the process]

Examples:
- GOOD: "Identify all security vulnerabilities that could allow unauthorized data access"
- BAD: "Review the auth file"

- GOOD: "Refactor authentication module for improved maintainability without breaking existing tests"
- BAD: "Refactor the code"

CONSTRAINTS (What You Cannot Do)

Explicit boundaries prevent scope creep and unexpected changes.

## CONSTRAINTS

- [Hard limit 1: e.g., "Do NOT modify any files"]
- [Hard limit 2: e.g., "Do NOT add new dependencies"]
- [Hard limit 3: e.g., "Maintain backward API compatibility"]
- [Resource limit: e.g., "Complete within 50 tool calls"]

Examples:
- "Identify issues only - do NOT implement fixes"
- "Read-only analysis - no file modifications"
- "Focus only on files in src/auth/ directory"

CONTEXT (What You Need to Know)

Provide all information required to complete the task. Sub-agents run in isolated context and cannot access parent conversation.

## CONTEXT

### Files to Analyze
- `path/to/file1.ts` - [brief description of relevance]
- `path/to/file2.ts` - [brief description of relevance]

### Related Context
- Previous findings: [summary of relevant prior work]
- Architecture notes: [relevant design decisions]
- Known issues: [existing problems to be aware of]

### Standards to Apply
- [Coding standard or guideline reference]
- [Security policy reference if applicable]

OUTPUT (What to Deliver)

Specify concrete deliverables with exact format requirements.

## OUTPUT

### Primary Deliverable
Write findings to: `logs/{agent-name}-{timestamp}.md`

### Output Format
[Specify structure: YAML, Markdown sections, etc.]

### Summary Requirements
Return to main thread: [max 200 tokens summary of key findings]

### Diagnostic Output
Write to: `logs/diagnostics/{agent-name}-{timestamp}.yaml`

Pipeline Syntax (F# Conceptual Notation)

Understanding the Notation

F# pipe syntax (|>) is a conceptual notation for planning and documentation. It is NOT directly executable Claude Code syntax.

// This is documentation, not executable code
Agent1 (task) |> Agent2 (task) |> Agent3 (task)

Purpose: Visualize workflow dependencies and conditional logic before implementation.

Mapping to Task() Invocations

Each pipeline stage maps to a sequential Task() call from the main thread:

// Conceptual pipeline
CodeAuditor (security) |> CodeAuditor (architecture) |> (if findings > 0 then IssueDebugger else Done)

Actual execution:

1. Main thread invokes: Task(description="Security audit", subagent_type="sonnet", prompt="[4-part prompt]")
2. Main thread reads log output, extracts findings
3. Main thread invokes: Task(description="Architecture audit", subagent_type="sonnet", prompt="[4-part prompt]")
4. Main thread reads log output, extracts findings
5. IF findings.count > 0: Main thread invokes IssueDebugger
6. ELSE: Pipeline complete

Pipeline Patterns

Code Review Pipeline:

CodeAuditor (security)
|> CodeAuditor (architecture)
|> TestAuditor (coverage)
|> (if findings > 0 then IssueDebugger else Done)

Fix Validation Pipeline:

IssueDebugger (root cause)
|> Implementer (apply fix)
|> CodeAuditor (verify quality)
|> TestAuditor (verify tests)
|> (if issues > 0 then IssueDebugger else Done)  // Loop until clean

Test Audit Pipeline:

TestAuditor (classify all)
|> (if mock_heavy > 0 then VerificationScriptCreator else Done)
|> Implementer (rewrite flagged)
|> TestAuditor (re-verify)

Key Constraint

Sub-agents CANNOT spawn other sub-agents. All pipeline orchestration happens from the main thread.

Custom Agent vs Built-in Agent

Agent Selection Priority

Before spawning a pipeline stage:

1. Check .claude/agents/ for a custom agent matching the task
2. If custom agent exists: Use its name field as subagent_type - this invokes the custom agent
3. If no custom agent: Use general-purpose and provide full specialized instructions in the prompt

Important: 4-Part Prompt Always Required

Regardless of agent type, every Task invocation MUST include the 4-part prompt (GOAL/CONSTRAINTS/CONTEXT/OUTPUT) and follow subagent-output-templating for output format.

Built-in Agent Types

Example: Custom Agent EXISTS

# .claude/agents/security-auditor.md exists with name: security-auditor
Task(
    description="Security audit",
    subagent_type="security-auditor",  # Invokes the custom agent
    prompt="""
## GOAL
Identify security vulnerabilities in calculator.ts

## CONSTRAINTS
- Do NOT modify files

## CONTEXT
File: src/calculator.ts

## OUTPUT
Write to: logs/security-audit-{timestamp}.yaml
"""
)

Example: Custom Agent DOES NOT Exist

# No custom agent for this task - use general-purpose
Task(
    description="Security audit",
    subagent_type="general-purpose",
    model="sonnet",
    prompt="""
## GOAL
Identify security vulnerabilities in calculator.ts using OWASP Top 10 patterns.

## CONSTRAINTS
- Do NOT modify files
- Focus on injection, auth, and data exposure risks

## CONTEXT
File: src/calculator.ts
This is a calculator module with arithmetic operations.

## OUTPUT
Write findings to: logs/security-audit-{timestamp}.yaml
Format: YAML with findings array, severity, and recommendations
"""
)

Note: When no custom agent exists, the prompt must include all specialized instructions that would otherwise be in the custom agent's markdown file.

Model Selection Guidance

Task-Type Based Selection (Objective)

Model selection is based on task type, not subjective complexity. This keeps selection deterministic and objective.

Selection Rules

1. Determine task type from the action verb:

   • Lookup/Execute/Run/Fetch → Haiku
   • Review/Analyze/Audit/Classify → Sonnet
   • Write/Fix/Implement/Apply → Opus

2. Custom agent override: If a custom sub-agent has agent: in its frontmatter, use that model instead of these rules.

3. Always specify model: Every Task() invocation must include subagent_type.

Pipeline Example

Orchestrator (Opus) writes initial code
     ↓
Sub-agent (Sonnet) reviews → finds issues
     ↓
Sub-agent (Opus) fixes
     ↓
Sub-agent (Sonnet) re-reviews
     ↓
[Loop until clean]

Anti-Patterns

Diagnostic Output (Required)

When this skill is used to invoke a sub-agent, the sub-agent MUST write diagnostic output.

Diagnostic File Location

logs/diagnostics/{skill-name}-{YYYYMMDD-HHMMSS}.yaml

Diagnostic Format

skill: subagent-prompting
timestamp: 2026-01-10T12:30:45Z
diagnostics:
  model_requested: sonnet
  model_actual: sonnet
  context_type: main
  parent_vars_accessible: true
  hooks_fired: []
  execution_time_ms: 1250
  completion_status: success
notes: "Skill invoked successfully"

When to Write Diagnostics

• At the END of skill execution (success or failure)
• Include actual model used (may differ from requested)
• Record execution time for performance tracking

Quick Reference

Prompt Checklist

[ ] GOAL: Outcome-focused objective stated
[ ] CONSTRAINTS: Hard limits explicitly listed
[ ] CONTEXT: All required files and background provided
[ ] OUTPUT: Log path, format, and summary requirements specified
[ ] DIAGNOSTIC: logs/diagnostics/ path included

Task() Invocation Template

Task(
    description="[3-5 word summary]",
    subagent_type="sonnet",  # or "haiku", "opus"
    prompt="""
## GOAL
[Outcome-focused objective]

## CONSTRAINTS
- [Limit 1]
- [Limit 2]

## CONTEXT
[Files, background, standards]

## OUTPUT
Write to: logs/{agent}-{timestamp}.md
Diagnostic: logs/diagnostics/{agent}-{timestamp}.yaml
Summary: [max 200 tokens]
"""
)

References

For extended examples and edge cases, see references/examples.md.