---

name: oberagent description: CRITICAL - Invoke FIRST before ANY Task tool call. This skill gates all agent dispatch. If you are about to use Task tool, STOP and invoke this skill. Enforces oberprompt principles. Triggers on "Task tool", "launch agent", "spawn agent", "dispatch agent", "parallel agents", "subagent", "background agent", "Explore agent", "general-purpose agent", "Bash agent", "Plan agent".

Skill: oberagent

Meta-skill that enforces prompt engineering best practices before dispatching agents.

The Iron Law

CopyNO AGENT PROMPT SHIPS WITHOUT COMPLETING THE AGENT PROMPT CHECKLIST

This applies to:

• "Quick" agent dispatches
• "Simple" Task tool calls
• Agents that "just need to search"
• Your 10th agent after 9 successes

Skipping this workflow = accepting subagent failures.

---

Required Workflow

Copy1. Invoke oberprompt skill (MANDATORY - loads prompt engineering guidance)
      ↓
2. Define Agent Purpose (what outcome, not what actions)
      ↓
3. Select Agent Type (match to purpose)
      ↓
4. Select Model Tier (match capability to task complexity)
      ↓
5. Identify Applicable Skills (subagents don't inherit skill awareness)
      ↓
6. Write Prompt (following oberprompt template + skill instructions)
      ↓
7. Validate (checklist)

Step 1 is non-negotiable. The oberprompt skill provides the constraint budget, progressive disclosure patterns, and validation checklist that make agent prompts effective. Without it, you're guessing.

---

Step 1: Invoke oberprompt

Before anything else, invoke the oberprompt skill.

CopyInvoke skill: oberprompt

This loads:

• Constraint budget guidelines for your model tier
• Progressive disclosure patterns
• Validation checklist
• Anti-patterns to avoid

Do NOT skip this step. Even if you "know" oberprompt principles or invoked it earlier in the conversation, invoke it again. Each agent dispatch is a fresh decision point requiring the full checklist.

---

Step 2: Define Agent Purpose

Before writing ANY prompt, answer:

If you can't answer these clearly, you're not ready to dispatch an agent.

---

Step 3: Select Agent Type

Match agent to purpose, not to your first instinct.

---

Step 4: Select Model Tier

Not every task needs Opus. Match model capability to task complexity. Using the right model improves speed, reduces cost, and often produces better results for simpler tasks.

Decision Table

Anti-Patterns

Progressive Model Selection

1. Start with the suggested default from the decision table
2. If agent fails or gives shallow results, consider whether:
   • The prompt needs improvement (most common)
   • The task is more complex than expected (upgrade model)
3. Document model choice in your dispatch reasoning

---

Step 5: Identify Applicable Skills

Subagents don't inherit skill awareness. They start fresh without knowing which skills you have access to. You must explicitly pass relevant skills.

Process

1. Review your available skills - Check what skills exist that could apply to the agent's task
2. Match skills to the agent's work - What type of work is the agent doing?
3. Include skill invocation in prompt - Tell the agent to invoke relevant skills first

Skill-to-Task Mapping

Example

Without skill inheritance (gap):

CopyReview the implementation plan and identify any issues.

With skill inheritance (correct):

CopyFirst invoke the code-foundations skill. Then review the
implementation plan and identify any issues with the design.

When to Skip

• Explore agents doing pure search/navigation - no skills needed
• Bash agents running simple commands - no skills needed
• Agents doing research only - typically no skills needed

When in doubt, pass the skill. Extra skill invocation is cheap; missing it causes failures.

---

Step 6: Write the Prompt

Use oberprompt guidance. Since you invoked oberprompt in Step 1, apply:

• Constraint budget for your model tier
• Progressive disclosure (start simple, add constraints only on failure)
• Outcome-focused framing (what you need, not how to get it)

Template

Copy[MODEL]: haiku | sonnet | opus (from Step 4 decision)

[SKILLS]: "First invoke [skill-name]" (if applicable - see Step 5)

[OUTCOME]: What you need to know/have when agent completes

[CONTEXT]: Only if agent lacks necessary background (usually unnecessary)

[SCOPE]: Optional narrowing (directory, file types, etc.)

Remember: The Task tool accepts a model parameter. Use it to specify the chosen tier.

Examples

Bad prompt (over-constrained):

CopySearch the codebase for authentication. Look in src/ directory.
Use Grep to find "auth" and "login". Then use Read to examine
each file. Return a list of all files with line numbers. Make
sure to check both .ts and .js files. Don't include node_modules.
Summarize what each file does.

Good prompt (outcome-focused):

CopyFind where user authentication is implemented and explain the
current approach. I need to understand the pattern to add OAuth.

Bad prompt (vague):

CopyLook at the code

Good prompt (clear outcome):

CopyFind how API errors are handled and surfaced to users. I'm seeing
inconsistent error messages and need to understand the current pattern.

---

Step 7: Validation Checklist

Complete EVERY item before dispatching.

Check 0 is the gatekeeper. If you didn't invoke oberprompt, checks 1-7 are based on guesswork.

---

Parallel Agent Dispatch

When dispatching multiple agents:

1. Each agent gets its own oberagent checklist
2. Agents must be independent - no agent depends on another's output
3. Combine results after all complete

Parallel Dispatch Template

CopyAgent 1: [Outcome A - independent]
Agent 2: [Outcome B - independent]
Agent 3: [Outcome C - independent]

If agents depend on each other, run them sequentially.

---

Red Flags - STOP and Reconsider

---

Debugging Failed Agents

When an agent returns poor results:

First attempt: Remove constraints, simplify prompt Not: Add more instructions

---

Integration with oberprompt

oberprompt is invoked at the start of every oberagent workflow (Step 1).

This ensures you have access to:

• Full constraint budget guidelines for your model tier
• Complete progressive disclosure patterns
• Validation checklist
• Anti-patterns and red flags

oberagent orchestrates the workflow; oberprompt provides the prompt engineering substance.