Tips & Best Practices

This article summarizes MBAgent veteran user experience, helping you upgrade from "knowing how to use" to "using it smoothly".

---

1. Prompt Writing

1. Structured Prompts

Write complex requirements structured so AI understands more accurately:

[Goal]
Organize today's orders from 10 stores into Excel.

[Input]
- Mbbrowser environment IDs: 223223, 223224, ..., 223232
- Time range: Today 00:00 - Now

[Output Requirements]
- File path: reports/orders-20260924.xlsx
- Columns: store ID, order number, amount, customer, status
- Sort: By amount descending
- Orders > 1000 highlighted in red

[Exception Handling]
- Environment can't start: log to errors.log, continue next
- Order page has no data: mark as "No orders today"

2. Provide Background

Tell AI the necessary business background:

I run 10 Amazon stores, mainly for the European/American market.
My customer IDs are always 32-char strings.
Order statuses have 5 types: pending payment, pending shipment, shipped, completed, cancelled.
Please organize the order data following these business rules.

3. Give Examples

The most effective prompts often contain examples:

Please rewrite this paragraph more professionally:

Original: We sold a lot in September and made quite a bit of money.

Rewrite example:
Original: Today there are especially many orders and we can't process them.
Rewrite: Today's orders have surged; we recommend increasing processing staff.

Original (to rewrite): We sold a lot in September and made quite a bit of money.

4. Counter-examples

❌ Too vague:

• "Help me with it"
• "Write something good"
• "What do you think"

❌ Too complex:

• "Please write me a Python script using requests library to call Mbbrowser ApiServer, open 10 Mbbrowser environments, process concurrently, asynchronously scrape..."

✅ Characteristics of good prompts:

• Clear goal
• Contains key parameters
• Specific output requirements
• Explains exception handling

---

2. Token Saving Tips

1. Choose the Right Model

Scene	Recommended Model	Saving
Daily Q&A	deepseek-v4-flash / gpt-5.4	85%
Doc summary	gpt-5.4 / deepseek-v4-pro	90%
Code generation	claude-sonnet-5	55%
Complex reasoning	claude-opus-5.5	0% (can't save)

2. Close Unnecessary Features

In "Settings → Preferences", close:

• ❌ Show thinking process (save 30-50% Token)
• ❌ Auto memorize every message (save 10%)
• ❌ Show Token and cost statistics (doesn't save money, but reduces interference)

3. Use the /compact Command

For long sessions, use /compact to compress history:

/compact

MBAgent compresses the history conversation into a summary, saving 50-70% Tokens while preserving key info.

4. Open New Sessions in Time

If the topic changes, opening a new session is more Token-efficient than continuing:

[Session A] Organize today's orders
[Session B] Write an English email ← New session, AI won't "remember" session A's order data

5. Delete Unrelated Attachments

If you uploaded an attachment but no longer need it, delete that attachment message in the session to avoid AI "seeing" it again in the next session.

---

3. Multi-Model Comparison

When to Use Multi-Model Comparison?

• ✅ Critical decisions (contract, legal, medical)
• ✅ Creative writing (different models have different styles)
• ✅ Learning purposes (see how different models handle the same problem)

Operation Method

1. Type message in the input box
2. Hold Ctrl and click multiple model buttons
3. Multiple models give answers side-by-side
4. Compare effects

Empirical Values

Task Type	Best-performing Model
Chinese writing	claude-sonnet-4 / qwen3-max
English email	claude-sonnet-4 / gpt-5
Code generation	claude-sonnet-4 / deepseek-v3
Complex reasoning	claude-opus-4 / o1
Multimodal (image)	gpt-5 / gemini-2.5-pro
Long doc analysis	claude-sonnet-4 (200K context)
Math problems	o1 / claude-opus-4 extended thinking

---

4. Mbbrowser Operations Tips

1. Multi-Account Concurrency

Let MBAgent operate multiple Mbbrowser accounts simultaneously:

Please simultaneously open the environments for these 10 stores:
- Store A
- Store B
- Store C
- ...
Collect order data in parallel, concurrency 5.

Best practices:

• Concurrency = 5-10 (not exceeding 10)
• Reserve independent time per environment (avoid triggering anti-scraping)
• Set reasonable timeout (5 min per environment)

2. Failure Retry

Mbbrowser environments may fail due to network, IP, cookies:

If a certain environment fails to start:
- Retry 3 times (30s between each)
- If retries still fail, log error to errors.log
- Skip that environment, continue to next

3. Screenshot Save

Let MBAgent save key page screenshots:

When collecting order data, screenshot each store's order page, save to screenshots/{store ID}-{timestamp}.png.

4. Exception Alerts

Notify immediately when encountering important anomalies:

If you discover the following situations, please immediately email alerts@company.com:
- Order amount anomalies (> 50000 or < 1)
- Environment cannot start
- Account frozen

---

5. File Processing Tips

1. Large File Block Processing

If the file is too large (> 1 MB), let AI process in blocks:

This CSV file has 100,000 rows. Please process in batches:
- 5000 rows per batch
- Save intermediate result after each batch
- Finally merge all results

2. Incremental Update

Use incremental update when processing large files:

Only process the new orders for 2026-09-24 this time, do not re-process historical ones.

3. Back Up Important Files

Before AI modifies files, let it back up first:

Before modifying ~/reports.xlsx, please back it up to backups/reports-{timestamp}.xlsx.

---

6. Knowledge Base Usage Tips

1. Difference Between KMS and AI Memory Bank

Dimension	AI Memory Bank	KMS Knowledge Base
Suitable content	Preferences, rules	Documents, references
Size	KB level	MB-GB level
Retrieval method	Auto-injected to prompt	Tool-call retrieval

2. KMS Auto Retrieval

After creating KMS, AI will automatically retrieve relevant content in conversations. You don't need to explicitly say "go check KMS".

3. KMS Incremental Update

KMS supports incremental update:

/kms ingest ./new-docs/

Only adds new files, doesn't rebuild the index.

---

7. Efficiency Improvement Workflows

1. Morning Routine

08:50 - MBAgent auto-runs "Today's sales report"
09:00 - You open MBAgent, daily report is already generated
09:05 - Browse daily report, find 2 stores with sales anomalies
09:10 - Let AI investigate the anomaly
09:30 - Receive anomaly investigation report
09:35 - Send handling email

2. Multi-Account Batch Operations

09:00 - AI starts 10 Amazon stores
09:05 - AI checks each store's inventory
09:10 - AI adds low-inventory products to restock list
09:15 - Organize restock list
09:20 - You review the list
09:30 - One-click order restock

3. Customer Inquiry Auto Response

10:00 - Customer leaves a message on the website
10:01 - MBAgent receives notification (via webhook)
10:02 - MBAgent retrieves relevant product info from KMS
10:03 - MBAgent generates reply draft
10:05 - You review and one-click send

---

8. Avoid Common Mistakes

❌ Don't Do This

1. No Backup Before AI Operates Important Accounts

Counter-example:

Help me clean up old emails in my inbox, delete the important ones, keep the unimportant ones.

Positive example:

Help me clean up old emails in my inbox, but first back up all emails to backup/emails-{date}.eml,
then only delete emails before 2024.

2. Let AI Do Too Many Steps at Once

Counter-example:

Help me complete all the following:
1. Organize data for 100 stores
2. Generate 10 reports
3. Email 50 customers
4. Update database
5. Back up all files
6. ...

Positive example: Split into multiple sessions or multiple tasks; each task focuses on one goal.

3. No Output Format Specified

Counter-example:

Help me organize the order data.

Positive example:

Help me organize the order data:
- Output format: CSV
- Fields: order ID, customer, amount, status, time
- File path: reports/orders-{date}.csv
- Encoding: UTF-8 with BOM (easy to open in Excel)

4. Not Testing Before AI Operates Unfamiliar Sites

Counter-example:

Help me scrape data from XXX site every day.

Positive example:

First help me manually scrape data from XXX site once, confirm we can get it.
Then change to auto-scrape every day.

5. Deploying AI-Written Code Without Testing

Counter-example:

Write a Python script that runs every day.

Positive example:

Write a Python script, run it once in a test environment first,
confirm the output is correct before deploying to production.

---

9. Debugging Tips

1. When Task Fails

1. Check error log: In task detail, see which step specifically failed
2. Run the failed step alone: Split the task into multiple and run alone
3. Simplify input: Test with the simplest input first
4. Switch model: Sometimes it's a model issue
5. Contact Mbbrowser customer support: Could be a Mbbrowser issue

2. When AI's Answer Is Wrong

1. Point out the error: "This isn't right; it should be X"
2. Provide correct example: Give AI a correct example directly
3. Multi-round fix: Let AI fix step by step
4. Re-ask: Rephrase and start over

3. When Token Consumption Is Abnormal

1. Check Token statistics: Below each reply round
2. Check for large attachments: Uploaded attachments consume Tokens
3. Compress history: Use /compact command
4. Open new session: Avoid history being too long

---

10. Team Collaboration

1. Shared KMS

Create a shared KMS for the team:

/kms create --scope=team --name=company-handbook
/kms mount company-handbook

All team members can access.

2. Shared Task Templates

Export task → send to colleague → colleague imports.

See Start Menu Tasks.

3. Shared API Key

⚠️ Sharing personal API Keys is not recommended. Suggest:

• Team purchases Mbbrowser official package together
• Each person uses independent sub-account
• Share knowledge via Mbbrowser RAG

---

11. Long-Term Learning

Recommended Learning Path

1. Week 1: Basic usage (chat, upload attachments)
2. Week 2: Mbbrowser integration
3. Week 3: Task panel and scheduled tasks
4. Week 4: Knowledge base and memory bank
5. Week 5: Custom workflows
6. Ongoing: Follow Mbbrowser official changelog

Recommended Reading

• MBAgent official changelog
• Mbbrowser usage tutorial
• Cursor / Antigravity AI programming tools usage

Join the Community

• Mbbrowser official WeChat group
• Mbbrowser user forum
• GitHub Issue (Mbbrowser official repo)

---

Next: FAQ → Check here first when encountering problems.