WorkBuddy Tutorial: Automate File Archiving with AI Agents, Skills & Safe Permissions
This step-by-step tutorial teaches beginners how to use Tencent's WorkBuddy desktop AI workbench to safely automate file archiving, covering installation, workspace setup, read-only inventory, plan-then-execute workflows, conflict handling, permission management, scheduled tasks, and skills—emphasizing verification over trust.
01 Install WorkBuddy and Prepare a Practice Directory
WorkBuddy supports Windows 10+ (x64; ARM64 per current release notes) and macOS 12+ (Monterey). Download from the official site https://www.workbuddy.cn/ and choose the installer matching your OS and architecture (Windows x64, Mac ARM64, Mac x64). On Windows, run the installer and click Finish. On macOS, open the .dmg and drag WorkBuddy to Applications.
After launch, sign in via QR code or browser authorization. Check the About page for version number and the models/quota available to your account (personal vs. enterprise may differ).
Create a Dedicated Practice Folder
On your desktop or an easy-to-find location, create WorkBuddy-练习区 with three subdirectories:
WorkBuddy-练习区/
├── 01_输入/
├── 02_输出/
└── 03_备份/01_输入holds practice materials. 02_输出 receives only newly generated artifacts. 03_备份 stores original copies (note: this is a convenience backup inside the same workspace, not a substitute for an independent off-site backup).
Add Mock Files for the Exercise
Inside 01_输入, create two folders 来源A and 来源B. Populate them with .txt files as follows:
来源A:
销售-2026-07-周报.txt # Normal classification
客服-2026-08-记录.txt # Normal classification
销售-2026-13-汇总.txt # Invalid month (13) → should be held
来源B:
销售-2026-07-周报.txt # Same name as A but different content → conflict
销售-客服-2026-09-联合事项.txt # Ambiguous department → should be held
财务-2026-09-报表.txt # Normal classificationThese mock files let you observe how WorkBuddy handles duplicates, invalid dates, and ambiguous labels without risking real data.
02 Understand the Workspace and Choose the Right Mode
Three key concepts:
Task – the assignment (e.g., “archive Q3 files by department/month and list exceptions”). Rules, steps, and context live here.
Workspace – the actual filesystem directory (here WorkBuddy-练习区). Operations without explicit paths default to this directory.
Artifacts – the final deliverables (sorted folders, exception lists, execution reports). Always verify them in the filesystem, not just in chat.
Select the workspace via the file picker when creating a task. Different tasks have independent chat contexts, but if they operate on the same filesystem directory they can still overwrite each other’s files.
Workspace vs. Knowledge Base
The workspace is the local directory for the current task. The Knowledge Base (资料库) stores long-term references and task outputs, governed by account/team permissions. Files added from the Knowledge Base become read-only context for the task; do not confuse the two.
Ask, Plan, Agent – Which to Use?
Mode | What It Does First | Use for Archiving Task
--------------------|----------------------------------------|-----------------------------
Ask (Q&A) | Discusses rules, does not execute | Clarify “how to judge duplicate files”
Plan (Planning) | Produces an execution plan for review | Review archive paths, exception handling, permissions
Default Agent (Execution) | Reads/writes files within granted permissions | Actually inventory, copy files, generate listsUse Ask to discuss rules, Plan to review a plan before execution, and Agent to perform file operations. Even in Agent mode you can explicitly constrain it to read-only. Ask’s file counts are not a substitute for a real inventory.
03 From a Vague Request to an Executable Archiving Plan
Start with a Fuzzy Requirement
Help me organize Q3 materials, archived by department.
Missing decisions: how to classify “销售-客服-联合事项”, which version to keep when names clash but content differs, whether to auto-correct invalid months (e.g., 2026-13), and where to place results.
Explicit Rules Table
Must Agree On | This Exercise’s Rule | Why It Matters
--------------------|---------------------------------------------------|-----------------------------
Input scope | Recursively scan 01_输入 under current workspace | Avoid reading wrong sources
Output root | 02_输出/部门/月份/ | Keep artifacts contained
Classification basis| Only explicit department tag and valid month in filename | Prevent misclassification from body text
Conflict handling | Same name, different content → hold both; repeated forwards → list only | No silent version selection
Forbidden actions | No delete, overwrite, move, or modify source files | Ensure recoverability
Deliverables | Sorted files, routing list, pending-confirmation list, quantity summary | Verifiable outputThese concrete rules reduce ambiguity far better than “please be careful”. If rules are still missing, a skill like grill-me can surface the gaps, but you must confirm the final rules.
Step 1: Read-Only Inventory
Before any processing, run a read-only scan of 01_输入 using an execution-capable mode with an explicit “read-only” constraint. This produces a current-state report listing every file, its path, and any anomalies (invalid months, name clashes, ambiguous departments). Treat “final version” in a filename as a clue, not proof.
Step 2: Write Classification & Halt Rules into Plan
Switch to Plan mode. Encode “unsure → stop” as explicit conditions. Avoid contradictory instructions like “don’t ask, just finish”. Require the plan to show expected destinations and exception lists before any file mutation. Key review points:
What files it will process (source paths, estimated counts; distinguish estimates from actual reads).
What it will modify (target directory vs. overwriting sources – risk profiles differ).
How it handles exceptions (name conflicts, unclassifiable, duplicates, permission errors) with explicit halt/report rules.
How it will verify (file counts, paths, content hashes, exception lists; later tasks add formula checks, slide overflow tests, etc.).
Step 3: Approve Plan, Then Execute Copy
Confirm the exact files to copy, target paths, and held items. Then issue an execution directive like:
请按我已经确认的归档计划执行。
只复制计划中明确允许的文件到 02_输出/部门/YYYY-MM/。
禁止删除、覆盖、移动、重命名或修改 01_输入 的任何文件。
目标文件已存在时,不得覆盖,先跳过并登记为冲突。
待人工确认的文件一律不复制。
完成后请在 02_输出生成:
- 文件去向清单(源路径、目标路径、处理状态)
- 待人工确认清单(文件、原因、建议核对项)
- 执行摘要(输入数、复制数、跳过数、异常数)
如执行中出现未预料的目录访问或权限请求,请先暂停让我确认。During execution, scrutinize any permission prompts – see exactly what path and action are requested before approving.
Step 4: Open Artifacts and Verify Independently
In Finder/Explorer, inspect 02_输出. Expected structure (approximate):
02_输出/
├── 销售/
│ ├── 2026-07/
│ │ ├── 销售-2026-07-周报.txt (from 来源A)
│ │ └── 销售-2026-07-周报.txt (from 来源B – conflict, held)
│ └── 2026-09/
│ └── 销售-客服-2026-09-联合事项.txt (held – ambiguous)
├── 客服/
│ └── 2026-08/
│ └── 客服-2026-08-记录.txt
├── 财务/
│ └── 2026-09/
│ └── 财务-2026-09-报表.txt
├── 待人工确认清单.md
└── 执行摘要.mdSuccessful copy count ≠ input file count. Held items (conflicts, invalid months, ambiguous labels) are correct behavior. Do not force them into folders to inflate completion rate.
Run a read-only verification pass:
请核对本轮归档结果,不要修改任何文件。
检查:每份输入文件是否都有明确处理状态;
已复制文件的目标路径是否正确;
异常与跳过项是否全部登记;
源文件是否保持原状;
能读取内容的副本是否与源文件内容一致(优先比较哈希)。
给出实测数字及无法验证的项目,不要自行宣布全部通过。Spot-check a few files, especially conflict groups and the pending list. Trusting the same agent’s “verified” claim is not independent verification.
Failure? Diff First, Don’t Full Re-run
If expected vs. actual counts mismatch, do not immediately re-run the whole batch – that can duplicate or overwrite correct results. Instead, ask for a minimal fix:
本轮任务未通过验收。
计划允许复制【按批准计划填写】份;实际复制【填写实测数】份。
请不要立即重跑。
先列出差异:
- 哪些输入对应的目标文件缺失?
- 哪些文件出现在错误的目录?
- 哪些文件的内容与源文件不一致?
- 需要人工确认的文件是否全部进入清单?
请给出仅针对异常项的最小修复方案,
未经我确认,不修改已有产物和源文件。This is idempotency in practice: check existing results, distinguish done/new/conflict, and apply only the necessary corrections.
04 Containing Risk When Exceptions Occur
Workspace Is Not a Perfect Sandbox
Separating input/output/backup aids auditing, but the workspace is not an impenetrable boundary. With sufficient permissions, a task can reach outside. The 03_备份 folder inside the workspace is not a reliable independent backup for critical data.
WorkBuddy distinguishes Default Permission (allows routine ops in workspace, prompts for out-of-scope access, critical deletes, script execution) from Full Access (disables many confirmations). Full Access applies to all current and future tasks in the client, not just the current one. Keep Default Permission daily; don’t use Full Access to dismiss dialogs.
When a permission dialog appears, read the requested path and action. Be especially cautious with bulk deletes, script runs, or external service calls.
One Conflict Illustrates the Danger
Read-only inventory reveals two files named 销售-2026-07-周报.txt with different content. Both would target 销售/2026-07/. Blind copy would silently overwrite one, leaving only a single file – loss may go unnoticed. After inventory, feed the conflict list into Plan so the agent adjusts: both excluded from copy, both logged in pending-confirmation with source paths and reasons. Verify the plan against the original conflict list before approving execution of the remaining files.
Out-of-Scope Directory Access
If WorkBuddy requests access outside the practice folder, pause. Ask it to explain the reason and impact. Even if the reason sounds plausible, seek a narrower, recoverable alternative. Real production data (contracts, PII) must comply with organizational confidentiality policies.
Connector Permissions Need Separate Confirmation
Connectors (email, cloud docs, automation) extend reach to external services. Before authorizing, verify the account scope and whether the connector can send, delete, or publish – actions often harder to undo than local file ops.
Recovery When Things Go Wrong
Sources intact, outputs wrong: Check exception list, fix only erroneous artifacts; don’t wipe and redo without confirmation.
Sources modified: Restore from an independent backup outside the task’s writable scope; verify version/content before proceeding.
No usable backup: Don’t let AI “auto-restore”. Preserve the scene, check system version history, recycle bin, or other existing backups.
WorkBuddy’s Safety Center offers file backup and delete protection, but their coverage depends on local config. They add recovery chances, not a substitute for verified backups.
05 Next Steps After Mastering the Basics
Models, Experts, Skills, Connectors
Model – understands/generates; pick by multiplier (lower = cheaper). Use low-multiplier models for Q&A, read-only inventory, rule-based archiving, list generation. Reserve high-multiplier for multi-step analysis, long research. Enable vision only when images are involved.
Expert – provides role-specific experience.
Skill – packages a workflow, tools, and operational requirements for a task type. Does not automatically grant file/external access; depends on skill definition and granted permissions.
Connector – links WorkBuddy to email, docs, etc.
Model Selection Tips
The model picker shows multipliers. Start with Auto. Use Ask/Plan to nail down goals, input scope, output format before execution (these modes also consume credits but reduce mid-run rollbacks). Don’t switch models mid-stream because one reply is slow – credits already spent are not refunded. If files won’t read, check directory, permissions, format first – not model or prompt length.
Using Skills
Open Skill Market via the “Expert·Skill·Connector” icon in the left sidebar.
Install a skill (built-in, local package, search, or custom).
Invoke the skill in chat; WorkBuddy calls its capabilities.
Review skill description, source, and requested permissions. Test with practice data first. Skills may read/write local files or call external services – vet third-party skills carefully. Disable or uninstall unused skills.
Scheduled Tasks for Recurring Work
Create in the left “Scheduled Tasks” panel. Define frequency, prompt, workspace, permissions. Click “Run Now” to test, then enable periodic execution. Monitor run history.
Example: “Every 1st of month 9 AM, summarize last month’s materials in workspace, save to 02_输出; read-only sources, no delete/overwrite, no external sends.”
Two types: Local (needs machine + client running, can access local workspace) and Cloud (runs offline, cannot read local files). Under Default Permission, sensitive ops may pause for confirmation; avoid scheduling irreversible actions (delete, overwrite, external send) for unattended runs.
Reference: https://www.workbuddy.cn/docs/workbuddy
First run: get one small task from clear rules → confirmed permissions → open files to verify results. That beats a single complex command. Later, apply the same method to spreadsheets, slides, larger datasets. Automate only after the manual flow is stable.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
