Recipes

Copy-paste Agent team configurations. Each recipe includes: group setup, Agent roster, SOUL.md highlights, and collaboration chain.

Recipe 1: AI writing studio

Scenario: Automated content — one message triggers a three-step Chinese + English article pipeline.

Setup

Setting	Value
Group name	📝 Writing studio
Members	Novelist + Editor + Translator
Collaboration chain	Novelist → Editor → Translator
Leader	Novelist

SOUL.md highlights

Novelist: Emphasize creativity and literary quality; output 500–1000 characters.

Editor: Preserve meaning; improve rhythm and wording; mark changes.

Translator: Literary translation; fidelity, fluency, elegance.

Example

You: Write a short prose piece about "a day in the life of an indie developer"

→ Novelist: ~800-character prose
→ Editor: Polished (12 edits, 3 transition paragraphs added)
→ Translator: English version (preserving tone)

Outcome

One message, within ~2 minutes:

• Original prose
• Edited version
• English translation

---

Recipe 2: Full-stack dev team

Scenario: Code generation + review + documentation in one flow.

Setup

Setting	Value
Group name	💻 Dev squad
Members	Code assistant + Editor
Collaboration chain	Code assistant → Editor
Leader	Code assistant

SOUL.md highlights

Code assistant:

## Code standards
- Ship complete, runnable code — no fragments
- Include JSDoc comments
- Include error handling
- Include usage examples

Editor (as code reviewer):

## Code review
- Check security (SQL injection, XSS, etc.)
- Check performance
- Check style consistency
- Give fixes in diff-style code blocks

Example

You: Build an Express REST API with JWT auth

→ Code assistant: Full code (routes, middleware, DB)
→ Editor: Review (3 security notes + 2 performance tweaks + style alignment)

Code assistant — reasoning first, then runnable code

---

Recipe 3: Private knowledge assistant

Scenario: Upload documents; the Agent becomes your domain expert.

Setup

Setting	Value
Group name	📚 Knowledge base
Members	Code assistant
Agent tools	rag_query + memory + web_search
Knowledge base	Project docs, API specs

SOUL.md highlights

## Retrieval rules
- Use rag_query on the knowledge base first
- If empty, use web_search
- Answer only from retrieved content — no fabrication
- Cite which document each fact came from

Example

You: How do we call the payments API?

→ Code assistant: [rag_query] → finds payment section in your API docs
→ "Per your API docs, the payment endpoint is POST /api/v1/payments..."

Good uploads

• API docs / Swagger exports
• Product manuals / user guides
• Technical specs / RFCs
• Meeting notes / project wikis
• Study notes / reading summaries

---

Recipe 4: Multi-model arena

Scenario: Same question, different models — compare answers.

Setup

Setting	Value
Group name	⚔️ Model arena
Members	GPT contender + Claude contender + Qwen contender

Create three Agents with identical SOUL.md but different models:

Agent	Model
GPT contender	gpt-4o
Claude contender	claude-sonnet
Qwen contender	qwen-max

Example

You: @GPTContender @ClaudeContender @QwenContender Explain Transformer attention in one paragraph

→ GPT contender: [GPT-4o answer]
→ Claude contender: [Claude answer]
→ Qwen contender: [Qwen answer]

Compare quality, style, and depth to pick what fits you.

---

Recipe 5: Daily / weekly report generator

Scenario: Auto-generate work summaries on a schedule.

Setup

Setting	Value
Agent	Xiajiao (虾饺) Butler
Tools	memory_search + manage_schedule
Schedule	Cron: Friday 17:00

SOUL.md highlights

## Weekly report
When the weekly job runs:
1. memory_search for this week’s conversation memories
2. Group by project / topic
3. Output:
   - Done this week
   - In progress
   - Next week
   - Risks / blockers

Schedule it

You: @XiajiaoButler Every Friday at 5pm generate my weekly report

Xiajiao Butler: [manage_schedule]
Scheduled: every Friday 17:00 — auto weekly report

---

Recipe 6: Customer support team

Scenario: RAG-backed answers plus translation for global customers.

Setup

Setting	Value
Group name	🎧 Customer support
Members	Code assistant (RAG) + Translator
Collaboration chain	Code assistant → Translator
Knowledge base	Product docs, FAQ, API docs

Flow

Customer (Chinese): What are your API rate limits?

→ Code assistant: [rag_query] → finds rate limits → friendly reply
→ Translator: English version for overseas customers

---

Recipe 7: Morning news digest

Scenario: Daily search + structured summary.

News Agent uses web_search and presents results in a structured table.

Setup

Setting	Value
Agent	Xiajiao (虾饺) Butler
Tools	web_search + manage_schedule

SOUL.md highlights

## News digest
When the news job runs:
1. web_search for today’s tech headlines
2. Pick the 5 most important
3. Each item: title + one-line summary + link
4. Clean layout for quick scanning

Schedule

You: @XiajiaoButler Every weekday 8:30am send me a tech news digest

---

Recipe 8: Interview coach

Scenario: Mock technical interviews with an Agent as interviewer.

Setup

Setting	Value
Agent	Custom “Interviewer”
Tools	memory (track performance)

SOUL.md

# Interviewer

You are a senior technical interviewer with 10 years of experience.

## Flow
1. Start with self-introduction
2. Tailor questions to the candidate’s background
3. Increase depth gradually
4. 2–3 follow-ups per question

## Style
- Professional and friendly
- Positive feedback on good answers
- On wrong answers, guide thinking — do not give the answer immediately
5. End with an evaluation report

## Rubric
- Technical depth (1–10)
- Clarity of thought (1–10)
- Communication (1–10)
- Overall recommendations

## Memory
- memory_write after each session
- Next session: adjust difficulty using history

---

Recipe 9: Competitive analysis

Scenario: Structured competitor research.

Setup

Setting	Value
Group name	🔍 Competitive analysis
Members	Code assistant + Editor
Collaboration chain	Code assistant → Editor
Tools	web_search + memory

Example

You: Compare Dify, FastGPT, and Coze

→ Code assistant: [web_search] → structured comparison
→ Editor: Polish + conclusions and recommendations

---

Recipe 10: Study notes

Scenario: Capture notes; Agent organizes and retrieves them.

Setup

Setting	Value
Agent	Custom “Study buddy”
Tools	memory_write + memory_search + rag_query

SOUL.md

# Study buddy

Help organize and retrieve study notes.

## Mode
- On new notes: extract concepts, memory_write
- On questions: memory_search first, then rag_query

## Principles
- Feynman style: explain hard ideas simply
- Link concepts together
- Tag source and date

---

Recipe 11: Multilingual technical docs

Scenario: One source document → English + Japanese.

Setup

Setting	Value
Group name	🌐 Multilingual docs
Members	Editor + English translator + Japanese translator
Collaboration chain	Editor → English translator → Japanese translator
Leader	Editor

SOUL.md highlights

Editor (preprocessor):

## Translation prep
When you receive a technical document:
1. Proofread grammar and wording
2. Mark terms as [term: preferred translation]
3. Mark do-not-translate spans (code blocks, identifiers, brands)
4. Output a clean source for downstream translators

English translator:

## Technical EN rules
- Keep code-block comments as-is; add English on the next line if needed
- Do not translate API paths or parameter names
- Follow Google Developer Documentation Style Guide
- Prefer active voice

Japanese translator:

## Technical JA rules (JP)
- Use です/ます style
- Prefer katakana for technical terms (サーバー, デプロイ)
- Follow Microsoft Style Guide for Japanese

Example

You: Translate this deployment doc... (paste source)

→ Editor: fixes 3 phrases, 8 term glosses
→ English translator: full English
→ Japanese translator: full Japanese

One submission, three language versions.

---

Recipe 12: Code migration assistant

Scenario: Move a legacy stack to a new one.

Setup

Setting	Value
Group name	🔄 Code migration
Members	Code assistant (analyze) + Code assistant (rewrite)
Collaboration chain	Analyze → Rewrite

SOUL.md highlights

Analyze Agent:

## Code analysis
When code arrives:
1. Identify language, framework, patterns
2. List dependencies and external APIs
3. Mark core business logic
4. List migration risks
5. Output a structured report

## Format
- Dependency map: old → suggested new
- Risk: 🔴 high / 🟡 medium / 🟢 low
- Mark logic that must not change

Rewrite Agent:

## Rewrite
From the analysis report:
1. Rewrite module by module
2. Preserve behavior; change implementation only
3. Add TypeScript types
4. Add unit tests

## Rules
- Keep public API (inputs/outputs) stable
- New code should be strictly better than a mechanical port
- Summarize key diffs

Example

You: Migrate this Express app to Fastify: (paste code)

→ Analyze Agent:
  - express → fastify, body-parser → built-in, cors → @fastify/cors
  - Risks: 3 middlewares need custom plugins
  - Core: 5 controller functions

→ Rewrite Agent:
  - Full Fastify project
  - Types
  - Migration notes
  - 3 unit tests

---

Advanced patterns

Dynamic Leader

Use a “dispatcher” Agent as Leader to route work:

# Dispatcher

You route user tasks to the best Agent.

## Team
- @CodeAssistant: programming and technical questions
- @Editor: copy, polish, translation
- @Novelist: creative writing
- @Translator: Chinese ↔ English

## Rules
1. Infer intent
2. @mention the best fit
3. For multi-step work, @mention in order
4. Do not answer yourself — only route

Expert voting

Multiple Agents propose; a “judge” synthesizes:

# Judge

You evaluate technical proposals.

## Flow
1. Wait for all experts
2. Score each on:
   - Feasibility (1–10)
   - Cost (1–10)
   - Risk (1–10)
   - Maintainability (1–10)
3. Output score matrix + recommendation

## Format
| Dimension | Plan A | Plan B | Plan C |
|-----------|--------|--------|--------|
| Feasibility | X/10 | ... | ... |
...
Recommendation: Plan X because ...

Memory-driven personalization

## Adaptive behavior
- If memory says preferred stack, default to it
- If memory says “keep it short”, cap replies at ~200 characters
- If memory has past projects, connect context
- On first contact, ask preferences once, then memory_write

Pattern summary

Pattern	Best for	Traits
1:1 chat	Q&A, code	Simplest
Group + @mention	Flexible collaboration	Manual routing
Group + chain	Repeatable pipelines	Auto handoff
Group + Leader	Primary + helpers	Default routing
Scheduled jobs	Digests, reports	No manual trigger
RAG	Doc Q&A	Evidence-based
Multi-model	Pick a model	Same prompt, different models
Dynamic Leader	Smart routing	Auto assignment
Expert voting	Design review	Many angles
Memory	Personalization	Learns over time

Related docs

• SOUL.md guide — Strong Agent personas
• SOUL.md templates — Copy-paste templates
• Multi-Agent chat — Groups and @mention
• Collaboration flow — Chains
• Model configuration — Model per Agent
• Platform comparison — How Xiajiao (虾饺) differs