How to Give Work to AI: A Beginner Task Brief
- Previous: Choose a safe workspace
- Goal: turn one job into a task brief that AI can understand and you can review.
- Estimated time: 8 minutes.
Step 2 · Choose a Workspace set up a dedicated workspace and finished a read-only inventory → this guide explains how to write a clear task brief covering goals, scope, steps, red lines, and deliverables.
💡 Core takeaway: Do not simply tell AI to “help me handle this.” Tell it what to do, what materials to use, what actions are prohibited, and what final deliverables to produce.
1. Why One Sentence Fails
Requests like these may sound natural, but none of them are specific enough:
Help me tidy this up.
Analyze this material.
Get the files sorted out.
The AI does not know:
- what your real objective is;
- which files it is allowed to use;
- whether to copy, move, or modify;
- whether to output a document or a spreadsheet;
- which details must not be guessed;
- what standard determines that the job is done.
The more ambiguous the brief, the more the outcome depends on the AI's guesswork.
2. The Six-Part Brief
2.1 Goal: Why Do This?
Do not state only an action; explain the intended purpose as well.
Goal: organize product materials by model so that sales reps can find them quickly.
2.2 Scope: What Can AI Read?
Only process files in the "02-to-process" folder.
Do not read or modify other directories.
2.3 Steps: In What Order?
1. Inventory the files first;
2. Identify likely product models;
3. Propose an organization plan;
4. Process at most 3 confirmable files as a sample;
5. If fewer than 3 files can be confirmed, process the actual confirmable count; do not invent files or expand scope;
6. Wait for my confirmation before continuing.
2.4 Forbidden Actions: What Must Not Happen?
Do not delete source files.
Do not overwrite source files.
Do not send, upload, or publish any content externally.
2.5 Deliverables: What Must Be Delivered?
Please produce:
1. the organized folders;
2. a product material index;
3. a processing log.
2.6 Uncertainty: What If It Is Unsure?
Place unverified items into "04-needs-review".
Do not invent product specifications based on filenames.
3. Copyable Universal Template
Please help me complete [task name].
一、Goal
I want this work to achieve: [explain purpose and expected result].
二、Information scope
Use only: [folder, files, or allowed material].
Do not read or process anything outside this scope.
三、Steps
1. Inventory existing material first;
2. Share your understanding and proposed plan;
3. Complete a small sample first;
4. Wait for my confirmation before processing the rest.
四、Forbidden actions
- Do not delete source files;
- Do not overwrite source files;
- Do not send, upload, or publish;
- Do not present guesses as facts.
五、Deliverables
Please deliver: [document, spreadsheet, folder, checklist, or report].
Also list every file created, copied, moved, or modified.
六、Uncertainty
Mark uncertain items as “needs review” and pause; do not guess on your own.
Before acting, restate in your own words:
1. The goal as you understand it;
2. Which materials you will use;
3. What actions you will not take;
4. What you plan to deliver.
I will confirm before you execute.
4. Why Ask AI to Restate First?
Having AI restate the task catches misunderstandings before any work runs.
For example, you might say:
Organize product images by model.
The AI might understand it as:
Organize products by image color.
If it restates first, you can correct the misunderstanding before any file is moved. This takes only tens of seconds, yet prevents massive rework.
How Can a Beginner Tell That the AI Really Paused?
Do not rely solely on whether it says “understood.” Confirm at least these 4 points:
- it explicitly states “waiting for your confirmation” or equivalent wording;
- it has not started running commands, and no created, modified, or moved files appear;
- it only restates the goal, scope, red lines, deliverables, and uncertainty handling rules;
- if the restatement is wrong, correct the rules first, and do not send “start,” “execute,” or “continue.”
The screenshot uses fictional teaching material in an already-installed Windows 11 environment with CC Switch and the “Laifa AI” configuration for reference testing. It illustrates the “restate first, then wait for confirmation” interaction pattern; it does not indicate that OpenAI's official ChatGPT sign-in workflow has been verified. The image still shows older folder wording, so it also cannot prove that the revised “read only 02-to-process” boundary was tested.
5. Why Start with a Small Sample?
Do not tell AI to process 500 files on the very first try.
Recommended workflow:
Process at most 3 files first
→ if fewer than 3 can be confirmed, process the actual confirmable count
→ inspect classification and format
→ refine the rules
→ then process remaining items
“At most 3” is an upper limit, not a quota that must be forced. When only 1–2 files can be confirmed, process only 1–2; do not read outside the scope, invent files, or overwrite existing outputs just to meet a target number.
5.1 How to Verify the “At Most 3” Limit Worked
When testing, you can deliberately prepare 4 fictional, fully recoverable input files and inspect the output directory. In the Windows 11 controlled reference test below, the Case D input directory contained 4 files:
In the recorded state after testing, the output directory only showed 3 newly added summary files corresponding to the first three inputs:
The August 3, 2026 machine manifest confirmed that “4 inputs added only 3 outputs,” but it did not preserve the complete causal chain from real Codex task submission through execution and disk writing, and the interaction text was in English. It is therefore partial reference evidence for the count rule, not proof that the current Chinese tutorial has been fully tested end-to-end. When practicing on your own, still require Codex to restate, wait for confirmation, execute, and have you independently verify inputs and outputs.
A small sample is not extra overhead; it verifies at very low cost:
- whether AI understands the task;
- whether the output format is appropriate;
- whether classification rules are correct;
- whether important fields are omitted;
- whether forbidden actions are respected.
6. Three States During Task Execution
| State | How AI should handle it |
|---|---|
| Confirmed | Process according to rules |
| Plausible but unevidenced | Mark as “inference” and await verification |
| Cannot determine | Place in needs-review; do not keep guessing |
In particular, product specifications, certifications, pricing, lead times, and customer details must never be written as facts without an explicit source.
7. Completion Check
Before moving to the next step, make sure:
- the task brief clearly states the goal;
- the allowed scope of information is specified;
- execution order is defined;
- deleting, overwriting, sending, and publishing are forbidden;
- final deliverables are stated;
- handling rules for uncertain items are defined;
- AI is asked to restate first and run a small sample.
📋 Common Pitfalls
| Pitfall | Consequence | How to avoid |
|---|---|---|
| 🔴 One-line vague prompt without requesting AI restatement | AI proceeds on guesses, producing off-target output and requiring massive rework | Force the AI to restate the goal, scope, and deliverables before execution |
| 🔴 Omitting explicit red lines and prohibitions | AI may overwrite, delete original files, or trigger unauthorized external uploads | Explicitly list prohibitions against deletion, overwriting source files, and external sharing |
| 🟡 Leaving deliverable formats and log requests unspecified | AI outputs disorganized code snippets or raw text that cannot be easily reviewed | State the deliverable format (Markdown list, table) and require a file-change log |
| 🟢 Letting AI guess whenever details are ambiguous | Specifications and certifications get hallucinated and recorded as facts | Establish a clear uncertainty rule: unsupported inferences must be flagged as "needs review" |
8. Next Step
Now that you know how to write a task brief, run the complete workflow on a low-risk practice case.
📄 Step 4 · Complete Your First Task — Follow step-by-step instructions to organize product materials and test your task brief.
👉 Continue to Step 4: organize product materials with Codex →