1.3 Your First Task: From One Sentence to a Deliverable
From a Request to Finished Work
Click 【New Chat】 in the left sidebar and describe the task:
Open the plus menu 【Add context】 beside the input box. Use 【Add files】 to upload material, or 【Choose skills】 / 【Choose works】 to select resources you have added (1.5, 1.6). The agent treats whatever you add as context for this task.
The model beside the input box is used for this task. Choose another available model when needed.
Creating a Task in 4 Steps
For the steps to start a chat, attach a file, and send, see AOS Work: Start Your First Task.
-
Write the task description — the way you would talk to a colleague. State the goal, the input material, the constraints, and the expected output.
-
Add attachments (optional) — use 【Add context】 → 【Add files】 to upload a file. To reference Library material, state its file name and folder path in the task.
-
Send — watch the plan, the tool calls, and the file changes as they happen.
-
View the output — open the generated file in the result panel. Add any changes you need in the same conversation.
Exercise: Review the Deliverable and Refine the Request
Start with the generated weekly report, Weekly Report.md, and continue by listing the owner of each task:
- Open the report from the results and check its
Completed,In progress,Blockers, andNext weeksections. - Request an owner checklist in the same conversation:
The conversation shows a new owner checklist: the product owner confirms the launch date, and the release owner approves the release checklist.
From One Line to a Task Brief: Six Elements
The quality of your description sets the ceiling on the quality of the output. A good one answers six questions:
| Element | What it answers |
|---|---|
| Goal | What problem are you solving |
| Input | Which files, directories, or links to use |
| Action | Analyse, organise, convert, or generate |
| Constraints | What must not change, which standards apply |
| Output | What to deliver, and where |
| Acceptance | What counts as done |
The element people skip most often is acceptance. Without it you can only judge the result by feel — and rework usually costs more than the extra line would have.
Good Versus Bad: The Same Task, Two Ways
Good:
Analyse this sell-in and sell-out data, calculate the sell-through rate per SKU for Italy, sort ascending by sell-through, and output an Excel file with five columns: SKU ID, Product Name, Units Sold, Stock on Hand, Sell-Through Rate. Do not modify the original data.
Bad:
Analyze this spreadsheet.
The difference: the good version states what to do, how to do it, in what format, and what counts as done. A vague description makes the agent guess, and the guess is often not what you had in mind.
Three Example Tasks
Three tasks of increasing scope, each written as a copyable brief.
Example A: Organize Files
1. Input: the "Test Files" folder in Library
2. Action: identify file types, propose a categorisation and rename scheme
3. Constraints: do not delete or overwrite originals; keep duplicates with a sequence number
4. Output: first produce a file inventory table and a categorisation plan
5. Acceptance: file count matches the actual folder and every action is traceable
6. Do not move any files until I confirm the plan.
Example B: Generate Meeting Minutes
1. Must include: conclusions, action items, owners, deadlines, and open questions
2. If an owner or deadline is not clear from the notes, write "TBD" — do not fill in guesses
3. Output: a minutes document and an action-item table
4. Acceptance: every conclusion traces back to the notes, and no action item is missing an owner or a status.
Example C: Word to PPT
1. Audience: department heads; speaking time: 8 minutes
2. Keep every fact and figure from the original — add no unverified numbers
3. Structure: background, current state, problems, options, plan, decisions needed
4. Output: the PPT file plus a slide-by-slide content summary
5. Acceptance: one core idea per slide, numbers matching the original, body text readable in projection mode.
Another Way to Start: Scheduled Tasks
The task above started because you asked for it. There is one other case where you do not say anything at all: you set the time and the content in advance, and it starts on its own.
Say you want every morning at 9 a.m. a summary of yesterday's follow-ups with every customer, delivered to your inbox. Work like that has a shape: it repeats, the rules are fixed, and you want a record kept — which is exactly what a scheduled task is for.
Compile yesterday's customer follow-up records into a summary: for each customer list the last contact date, the current stage, and the next action, and flag any customer with no progress for more than 7 days.
Output one Excel file plus a progress summary of no more than 200 words.
Send the result to my inbox.
The division is simple:
| What you want | Use |
|---|---|
| Something one-off, needed right now | Send a task |
| Something daily or weekly with fixed rules | A scheduled task |
When a scheduled task suits
Daily reports and weekly summaries; periodic checks on stock, quote validity, and contract expiry; time-based notifications to customers or the team.
When it does not
One-off, ad-hoc work; anything that needs your judgement each time it runs.
Full operating detail
Creating a schedule, changing its timing, and reading failures — see Scheduled Tasks. This chapter makes you aware the route exists; Chapter 2 shows how to combine it with Skills and Works.
When the Output Is Wrong: Three Failures and Their Fixes
A first run that misses your expectation is normal. Almost every miss falls into one of three buckets:
| Symptom | Most likely cause | What to fix first |
|---|---|---|
| The direction is not what you wanted | The goal was too broad | Replace "take a look at it" with a specific judgement target, e.g. "find the 10 slowest-selling SKUs and likely reasons" |
| It used the wrong source, or missed a file | The input was not pinned down | Name the files, directories, and tables in the description, or attach them explicitly with the 【Add context】 menu |
| Format, column names, or counts don't line up | No constraints and no acceptance criteria | Add "the output must contain these columns", "do not modify the original data", "counts must match the source" |
One rule governs all three: write the sentence you would say during rework into the task description instead. After two or three rounds, your own template becomes second nature.
Summary
What you write is not a prompt. It is a task brief. The clearer the brief, the more usable what comes back — from a colleague, and from an agent.
With descriptions handled, the next lesson lays out the resource types — and where the scheduled task fits among them.
1.2 The Interface: Three Zones and Two Exits
Three zones, three jobs. Read them once and most interface confusion disappears.
1.4 Resource Map: The Four Kinds of Resources in Explore
The rest of the book uses these names directly. This lesson sets out their types, their sources, and how each one takes effect.