| name | 3a-guide |
| description | Use when a developer asks about agent design best practices, patterns, or how to make a better agent. Provides curated guidance on architecture patterns, tool design, prompt engineering, error handling, and other quality aspects of agent development. Triggers on: "best practice", "how should I design", "good agent", "agent pattern", "multi-agent", "tool design", "prompt design", "3a guide", "advice", "is this a good approach", "what pattern", "single vs multi agent", "how to structure my agent". Not for project planning — use 3a-plan. Not for progress tracking — use 3a-track. Not for stage-gate reviews — use 3a-review. Not for implementation — delegate to aws-agents.
|
| allowed-tools | Read Grep Glob |
| metadata | {"type":"skill","version":"0.1.0"} |
guide
Curated design guidance for building better agents.
When to use
- Developer asks "what's the best way to..." about agent design
- Developer is choosing between patterns or approaches
- Developer wants to validate a design idea before committing
3a-plan needs a reference to guide an architecture decision
Do NOT use for:
- Creating a project plan or roadmap -> use
3a-plan
- Checking progress -> use
3a-track
- Readiness reviews -> use
3a-review
- Running
agentcore commands or writing code -> delegate to aws-agents
Input
$ARGUMENTS can be:
- A topic: "multi-agent", "tool design", "prompt"
- A question: "should I use one agent or two?"
- Empty — list available guidance topics
Process
Step 1: Identify the topic
Match the developer's question to an available reference:
If no reference matches, answer from general knowledge but flag: "이 주제에 대한 정리된 가이드는 아직 없습니다. 지금 논의한 내용을 /3a-decide로 기록해두시겠어요?"
Step 2: Load and present
Load the matching reference. Don't dump the entire document — extract the section relevant to the developer's specific question.
Present as a conversation, not a lecture:
- State the key tradeoffs concisely
- Ask a clarifying question to narrow down
- Give a recommendation based on their context
Step 3: Connect back to the plan
If the developer is in an active project (.3a/ exists):
- Suggest recording the decision via
/3a-decide
- Check if this changes anything in
.3a/roadmap.md
If no active project:
- The guidance stands alone — no need to force a planning workflow
Available references
References are added incrementally as patterns emerge. Current catalog:
- Multi-agent patterns — Single agent vs coordinator vs pipeline vs fan-out vs hierarchical
- AWS Prescriptive Guidance — AWS 공식 가이드 9종 카탈로그 (패턴, 프레임워크, 보안, 서버리스, 경제성, 거버넌스, 멀티테넌트 등)
Adding new references
New references should follow this structure:
# [Topic Title]
## When to use this reference
[Trigger conditions]
## [Pattern/Approach sections]
[Each with: when to use, tradeoffs, examples]
## Decision Guide
[Questions to ask in order to narrow down the choice]
## Boundary with aws-agents
[What 3A covers vs what aws-agents covers]
Output
- Targeted advice for the developer's specific question
- Clear tradeoffs, not just "it depends"
- A recommendation when enough context is available
- Suggestion to record the decision as an ADR
Quality criteria
- Advice is specific to the developer's context, not generic
- Tradeoffs are concrete (latency, complexity, cost), not abstract
- Always distinguishes "what pattern to choose" (3A) from "how to implement it" (aws-agents)