Three Execution Modes: Default (Agent), Plan, and Ask
You have already fed in the materials and requirements (the previous lesson); this page answers "in what manner does it execute." Given the same request, should it act directly, present a plan before acting, or only answer without acting—WorkBuddy modes are the switch that governs exactly this. It has two independent knobs: execution mode (three levels) decides how it works, and permission mode (two levels) decides whether it needs to ask you first before doing sensitive work. In 10 minutes, once you get these two knobs straight, you will no longer stare helplessly at it "randomly modifying files right off the bat" or "generating a pile of stuff when you just asked a question."
This is lesson 4 of "Daily Use," based on WorkBuddy v5.9.3, with operations following the official "New Task Bar" and "Default Permissions and Security Sandbox" documentation.
- Not yet fluent with input methods? Go back to 2.3 A Complete Guide to Input Methods: Voice, Files, and Prompt Templates
- Want to see the full picture of this stage first? Go back to 2.0 Daily Use: Turn WorkBuddy Into Your Office Buddy
What You Will Be Able to Do After This Lesson
- Distinguish the two dimensions of WorkBuddy modes: execution mode (default Agent / Plan / Ask) and permission mode (default permissions / full access)
- Switch execution modes in the + menu on the left of the input box, and state the applicable scenarios for each of the three modes
- Run a task end to end in Plan mode: fill in information → review the plan → confirm execution (or adjust the plan)
- Use Ask mode for answers only without execution, and explain clearly its relationship with credits
- Understand older terms like Craft in community tutorials and stop getting confused by outdated terminology
Your Current Pain Points
- You ask it to organize a folder and it dives right in; only after it finishes do you realize it misunderstood, and you have to roll back and start over
- You have not thought the requirements through, but it has already finished a version—credits spent, yet the output is unusable
- You only wanted to ask "which number in this spreadsheet is wrong," but it went ahead and generated an analysis report
When to Use This Technique
When you need to:
- Choose an execution approach for the current task among "execute directly," "present a plan then execute," and "answer only without executing"
- Control whether it needs your consent before modifying files, deleting files, or running commands
And you do not want to:
- Redo the entire task after it goes off track, burning credits for nothing
- Get a pile of useless files generated when you only wanted a quick consultation
🎒 Preparation Before Starting
- Completed 2.0 Daily Use: Turn WorkBuddy Into Your Office Buddy, and successfully ran through 1.4 Your First Task
- Have WorkBuddy installed and logged in on your computer, able to create new tasks and find the + menu on the left of the input box
- 10 minutes of reading time
💡 Does it cost credits?
Official rules: all conversation-related requests consume credits, with the amount depending on model pricing and task complexity—simple Q&A consumes less, complex tasks consume more. Third-party experience suggests that Ask avoids triggering file operations and Plan prevents off-track rework, indirectly saving credits (noted in the main text). For the full rules see 2.6 Credits and Pricing.
Core Idea
Mental model:WorkBuddy has two independent knobs—do not mix them up.
Knob one: execution mode—how it gets the work done. Official terminology, three levels:
| Mode | Description | Applicable Scenarios |
|---|---|---|
| Default (Agent) | Executes and completes your task instructions efficiently; can modify files | Document generation, spreadsheet processing, file organization |
| Plan | Plans the task first, then executes after your confirmation | Multi-step tasks, when the scope of changes needs review |
| Ask | Suitable for asking questions and exploring ideas; no execution | Understanding content, confirming requirements |
Knob two: permission mode—whether it needs to ask you first before doing sensitive work. Official terminology, two levels:
| Permission | Description |
|---|---|
| Default permissions | Sensitive operations (such as file modifications, execution outside the workspace) require your confirmation to execute, or you complete the operation yourself |
| Full access | Reduces confirmation steps and allows more operations to execute directly; once enabled, it applies uniformly to all tasks within the current client |
The two knobs do not substitute for each other: execution mode governs "how the task proceeds," while permission mode governs "which actions need your approval." Enabling Plan does not mean it will bypass permission confirmation—the execution phase after plan confirmation is likewise constrained by the current permission mode.
The official advice is straightforward: use the default (Agent) mode for everyday tasks; enable "Plan" when you need to control the plan first; enable "Ask" for consultation only without execution. Back up important files before working on them.
💡 Confused by "Craft" in older tutorials?
In early versions and community articles you can still see the mode name Craft — official release notes include fix records such as "Chat/Craft mode switching" and "ask mode automatically switches to craft after a conversation"; community articles describe Craft as "directly getting hands-on with files," which corresponds to the current default (Agent) mode. Some community articles also mention names like "deep processing" and "full refactoring" (third-party article terminology; the current official documentation has no such naming). When cross-referencing, always use the official current three modes as the standard.
Follow Along
Step 1: Find the WorkBuddy mode switch (the + menu on the left side of the input box)
Why: With the wrong mode, the task gets advanced the wrong way — first know where the switch is; every later step starts from here.
Steps:
- Look at the + menu on the left side of the input box and open it
- In the "Mode" section you can see the switches for "Plan" and "Ask only"
- With nothing enabled by default, that is the default (Agent) mode — in the official wording, "the default is Agent mode"
- "Plan" and "Ask only"** are mutually exclusive**: enabling one replaces the other; after turning them off, it automatically returns to the default mode
- When in a non-default mode, the corresponding mode badge appears below the input box; click the × on the badge to exit
You should see: opening the + menu shows the mode switches; after enabling "Plan," a mode badge with an × appears below the input box.
- I found the mode switches in the + menu and have seen the mode badge below the input box
Step 2: Keep the default (Agent) mode for everyday tasks and execute directly
Why: When the task goal is clear and can be stated in one sentence, asking directly for the result is fastest — this is also the everyday default tier officially recommended.
Steps:
- No need to switch anything; keep the default mode
- Enter and send the task as usual: WorkBuddy goes straight into the conversation page to execute; the Agent's takeover and execution process is shown in real time in the conversation interface (there is no transitional page like "Preparing to execute…")
- It can modify files; the output lands in the workspace and can be viewed in the results area
Suitable scenarios (official wording): document generation, spreadsheet processing, file organization.
You should see: after sending, the conversation area immediately starts scrolling with the execution process, and the output appears in the results area once the task completes.
- I have run a task in the default mode and know it gets hands-on directly and can modify files
Step 3: For complex tasks, enable Plan mode — produce a plan first, then execute
Why: For tasks with vague requirements, many steps, or high output cost, running directly tends to go off track. The core of Plan is "produce a plan first, then execute": the model does not immediately produce the final result, but first gathers the missing information and presents a complete, reviewable plan, and only starts executing after you confirm. The official wording is — this lets you correct course within tens of seconds, instead of discovering the approach was wrong only after the result has finished running.
When to enable Plan (official comparison table):
| Recommended to enable Plan | Default mode is faster |
|---|---|
| The task must be broken into multiple steps to complete | Single-point Q&A: looking up a number, asking about a concept, translation |
| Requirements are relatively vague with multiple possible solutions | A clear, small change |
| Key prerequisites are not yet determined | You know exactly what you want; executable in one sentence |
| High output cost; you don't want a wasted run | Continuous lightweight follow-up questions within a session |
A one-sentence rule: If "how to do it" is more worth discussing first than "getting it done," enable Plan mode; if you just need one answer, use the default mode.
Steps (official four-step process):
- Enable Plan: In the "Mode" section of the input box's + menu, turn on the "Plan" switch
- Provide information: The model will ask you questions around key variables (such as time, budget, preferences). It is recommended to answer each item rather than skip — the more information, the better the plan fits reality; if you skip everything, the model will make its own assumptions and mark them in the plan
- Review the plan: Once the plan is generated, the full Plan content can be viewed in the right-side panel
- Confirm execution, two choices:
- Start execution — generate the corresponding plan or files according to the confirmed plan; the execution phase strictly follows the plan
- Adjust the plan — state the changes directly (trim / narrow / add / reorder / start over); the model outputs again and you can iterate repeatedly; execution will not start before confirmation
You should see: after sending the task, it first asks questions or directly presents a plan; the full Plan appears in the right-side panel; it only truly gets to work after you click "Start execution."
- I have run a task in Plan mode: answered its questions (or skipped them), reviewed the plan, and clicked "Start execution"
💡 Third-party experience: Plan is also a money-saving mode
Experience from a Tencent Cloud community article: executing complex tasks directly in the "hands-on file editing" mode means that if the direction is slightly off, you have to go back and forth with edits, doubling the consumption; the safer approach is to first use Plan to confirm the plan and then execute (third-party opinion, for reference).
Step 4: For consultation-type questions, switch to Ask mode — Q&A only, no execution
Why: When you just want to ask questions or explore ideas, the worst thing is it "helpfully" generating a pile of files — Ask mode promises Q&A only, no execution.
Steps:
- Enable "Ask only" in the + menu on the left side of the input box
- Ask as usual: requesting information, confirming concepts, discussing approaches are all fine
- It will not modify files or execute tasks; the mode badge below the input box reminds you that you are currently in the Ask tier
Suitable scenarios (official wording): understanding content, confirming requirements. For example, "What deadline is mentioned in this document" or "Which sections does a weekly report usually include."
⚠️ Ask ≠ free chat
Official credit rules: all conversation-related requests consume credits, and the amount consumed depends on model pricing and task complexity. Ask only means "no execution," not "no billing." Third-party experience suggests that conversations which don't need to generate or modify files should use Ask, which avoids triggering extra file operations and thereby indirectly reduces consumption (third-party opinion). See 2.6 Credits and Pricing for the credit rules.
You should see: it gives a text answer, and no new files appear in the workspace.
- I have asked a question in Ask mode and confirmed it only answers and does not touch files
Step 5: Switch permission modes as needed: default permissions vs. full access
Why: The execution mode governs "how to do it," while the permission mode governs "whether sensitive operations need confirmation" — this is the second independent knob, and it must be set correctly before handling important files.
Steps:
- The permission entry is at the bottom of the input box: a persistent permission status button that shows "Default permissions" by default; click it to view the current permission status and its explanation
- Switch between the two tiers via the "Allow full access" switch; the "Configure task" section of the official "New task bar" documentation also lists a "Permission management" item
- When switching to full access, a "Confirm allowing full access?" dialog pops up, clearly listing the capabilities to be unlocked (file operations, terminal commands, internet access); you must check the risk acknowledgment and then click "Allow full access"
Comparison of the two tiers (condensed from the official table):
| Comparison item | Default permission | Full access permission |
|---|---|---|
| File access scope | Within sandbox constraints; going out of bounds requires asking for permission | Anywhere on your computer |
| Deleting / modifying files | Confirmation required each time | Mostly executed directly |
| Interruption frequency | Relatively high | Very low |
| Scope of effect | Current task | Applies uniformly to all tasks within the current client |
| Risk level | Low | High |
Three rules you must remember:
- Takes effect globally: once full access permission is enabled, it applies uniformly to all existing tasks and subsequently created tasks within the current client; after it is turned off, all tasks revert to the default confirmation flow. It does not carry over across devices—each client is managed independently
- Turn it off when done: the official recommendation is to keep the default permission in daily use and only switch to full access for a short time in tasks that are recoverable and isolatable
- Default permission does not block everything: routine operations inside the sandbox proceed as usual; only out-of-bounds and high-risk operations (deleting, modifying, system commands, sending data out, etc.) require confirmation—everyday writing, analysis, and document generation are basically unaffected
What you should see: clicking the permission button at the bottom of the input box shows an explanation of the current status; when switching to full access, you first see a risk confirmation dialog.
- I can find the permission button at the bottom of the input box and state that full access permission takes effect globally and should be turned off when done
For detailed steps, refer to the official documentation: New task bar (configuring tasks), Default permission and security sandbox.
Checkpoint ✅
- Can switch among the three execution modes in the + menu on the left of the input box, and knows that "Plan" and "Ask" are mutually exclusive and automatically revert to default after being turned off
- Can recite the official applicable scenarios: Default = document generation / spreadsheet processing / file organization; Plan = multi-step tasks / cases requiring review of the scope of changes; Ask = understanding content / confirming requirements
- Has run a task end to end in Plan mode: supplying additional information (or skipping to let it make its own assumptions and flag them) → reviewing the plan in the right-side panel → starting execution or adjusting the plan
- Knows that Ask mode also consumes credits: all conversation-related requests consume credits
- Can find the permission button at the bottom of the input box and state that full access permission takes effect globally and is recommended to be turned off when done
Pitfall reminders
❌ Treating "execution mode" and "permission mode" as the same thing
Symptom: you turned on Plan and assumed it wouldn't touch files carelessly; or you left full access permission on and assumed it would still present a plan for your review before executing.
Cause: these are two independent dials—the execution mode (Agent / Plan / Ask) determines how a task proceeds, while the permission mode (Default / Full access) determines whether sensitive operations need confirmation. The execution phase after a Plan-mode plan is confirmed is still subject to the current permission mode.
Correct approach: set them separately. To review plans, turn on Plan; to reduce confirmation dialogs, adjust the permission switch. For how to combine the two, see "Advanced tips" below.
❌ Forcing a run in default mode even with vague requirements, only to find the direction was wrong after it finishes
Symptom: you toss it a one-liner, it runs for a few minutes and delivers a version, you open it and it's completely not what you wanted, so you have to redo it.
Cause: default (Agent) mode executes directly; when requirements are vague, there are multiple possible solutions, or premises are undetermined, the direction easily goes astray. Third-party experience (Tencent Cloud Community article): diving straight into complex tasks means that even a slight deviation in direction requires back-and-forth changes, doubling the cost.
Correct approach: for multi-step tasks, vague requirements, undetermined premises, or high-cost deliverables, turn on Plan first to align on the plan before executing—in the official wording, "correct the direction within tens of seconds instead of discovering the approach was wrong only after the result finishes running."
❌ Assuming Ask mode is "free chat"
Symptom: treating Ask as an unlimited Q&A machine and chatting about everything in it.
Cause: the official credit rules are explicit—all conversation-related requests consume credits; Ask only means "no execution," not "no billing."
Correct approach: for single-point Q&A (looking up numbers, asking about concepts, translation), either Ask or default mode works; if you care about the credit bill, first check 2.6 Credits and pricing.
❌ Leaving full access permission on long-term just to click fewer confirmations
Symptom: finding confirmation dialogs annoying, you keep full access permission on for all tasks.
Cause: full access permission is a global switch (it applies uniformly to all tasks in the current client), and high-risk operations are no longer confirmed step by step. The official documentation explicitly lists scenarios where enabling it is not recommended: you are handling production data, customer data, financial data, or the only copy of something; the working directory is near the Desktop, Downloads, the root of your personal documents, or the root of a project repository; the task will batch-delete, batch-rename, or overwrite files; you need to run scripts or third-party tools you don't understand; or your computer contains important files without a backup.
Correct approach: keep the default permission in daily use; only enable full access for a short time in trusted, isolated, recoverable tasks, and turn it off when done. One official exception: unattended scheduled tasks are recommended to use full access permission—under the default permission, sensitive operations require your confirmation, and if you're not at the computer, the task will stall waiting for confirmation (see 3.3 Scheduled tasks for details).
Advanced tips (optional)
🔧 Combo for important files: Plan + default permission + backup beforehand. The official advice is "back up before handling important files"; Plan lets you review the plan before taking action, and the default permission adds a checkpoint before sensitive actions. For an in-depth discussion of backups, the sandbox, and other security mechanisms, see 5.4 Data security and privacy.
🔧 Use Ask as a "requirements clarifier before taking action." The official positioning of Ask includes "confirming requirements"—before starting a big task, use Ask to clarify concepts, premises, and definitions first, then switch back to default mode to execute; that's more cost-effective than redoing work after going off track.
🔧 Pair unattended scheduled tasks with full access permission. The official scheduled-task documentation recommends full access permission (the task runs unattended, and tool calls need no per-step confirmation), because under the default permission no one is there to click confirm and the task will stall. For the configuration process, see 3.3 Scheduled tasks: letting WorkBuddy work on its own on schedule.
🔧 A practical Plan reference. For the complete process of using Plan to confirm the plan first and then generate a PPT, see A4 Generating a PPT in one sentence: from source material to presentation deck.
Lesson summary
✅ Two dimensions: WorkBuddy modes = execution mode (three levels) + permission mode (two levels), two independent dials, set separately.
✅ Execution modes: Default (Agent) executes directly and can modify files; Plan presents a plan first and executes after confirmation; Ask only answers questions without executing. The entry point is the + menu on the left of the input box; "Plan" and "Ask" are mutually exclusive and automatically revert to default when turned off; when not in default, an indicator appears below the input box—click × to exit.
✅ Plan flow: turn on the switch → the model asks questions to fill in information (if all are skipped, it makes its own assumptions and flags them) → review the plan in the right-side panel → start execution / adjust the plan (remove, narrow, add, reorder, start over; nothing executes before confirmation).
✅ Two permission levels: default permission (sensitive operations confirmed first, keep it in daily use); full access permission (takes effect globally, reduces confirmations; enable briefly for trusted, isolated tasks and turn off when done). The entry point is the persistent button at the bottom of the input box.
✅ Credits: all conversation-related requests consume credits—Ask isn't free either; third-party experience holds that Ask triggers fewer file operations and Plan prevents going off track, which can save credits indirectly.
✅ Old naming: Craft in early versions and community articles ≈ the current Default (Agent) mode.
Key settings recap:
| Setting | Recommended value | Description |
|---|---|---|
| Execution mode | Default (Agent) | Use directly for everyday tasks, can modify files |
| Plan | Enable for multi-step / vague requirements / high-cost outputs | Plan first, then execute; no action before confirmation |
| Ask only | Enable when only consulting without executing | Understand content and confirm requirements; still consumes credits |
| Permission mode | Default permissions | Sensitive operations require confirmation; keep this for daily use |
| Full access | Enable briefly for unattended / trusted isolated tasks | Takes effect globally; turn it off once done |
Next lesson preview
With the execution approach settled, the next lesson tackles "which model to send for the job." In the next lesson you'll learn 2.5 Model selection and customization: 14 built-in models, API keys, and Ollama.
You will learn:
- How to choose among the Fast / Balanced / Ultimate model tiers
- The list of 14 built-in models and their applicable scenarios
- Connecting custom models via API key, Tencent Cloud, or Ollama
- How model multipliers affect credit consumption