Chapter 3 — Your First Task: Say What You Mean
What You Will Learn
- Describe a task with four elements: goal, input, constraints, acceptance;
- Ask for a plan before execution;
- Inspect tool calls and file changes;
- Give useful feedback;
- Use
presentto deliver a file.
3.1 Four Elements of a Good Task
Most "the Agent is bad at this" experiences come from vague requests.
A good task contains:[1]
| Element | Question it answers | Example |
|---|---|---|
| Goal | What should exist at the end? | A summary.md file |
| Input | Where does the material come from? | Three files under notes/ |
| Constraints | What must not happen? | Do not modify originals; Chinese; keep links |
| Acceptance | How do we know it is done? | All three titles; 3–5 sentences each; opens correctly |
Bad
Organize my notes.
Better
Read the three Markdown files under notes/ and combine them into notes/summary.md.
Requirements:
1. Keep each original title and source link;
2. Write 3–5 sentences per article;
3. Do not modify the original files;
4. Give me a plan first, then wait for my confirmation;
5. After writing, verify that summary.md exists and tell me which files you read.
The difference is not length. It is completeness.
3.2 Plan Before You Execute
For anything beyond a trivial task, ask for a plan:
Do not change files yet. Give me a 3–5 step plan:
what you will read, what you will write, and how you will verify the result.
Check whether the plan:
- Covers every input;
- States the output location;
- Includes verification;
- Does not quietly expand the scope.
If the plan is wrong, fix the plan first.
3.3 Watch the Tool Calls
While the Agent works, inspect the tool cards:
| Tool | What to check |
|---|---|
| Read file | Is it the right file? |
| Write file | Where did it write? Did it overwrite? |
| Run command | Is it destructive? |
| Web search | Are the sources credible? |
| Subagent | What was delegated? |
A useful habit:
When you see write, delete, overwrite, reset, push, or upload, pause.
3.4 Iterate with Specific Feedback
First drafts are rarely perfect. Good feedback is specific.
Vague
Not good. Rewrite it.
Specific
The second paragraph is too long. Cut it to three sentences.
Keep the three subheadings.
Move each source link to the end of its paragraph.
Do not change the other paragraphs.
Specific feedback = location + problem + standard.
3.5 Verify — Do Not Trust "done"
"Done" is not the same as "correct." Check:
- Does the file exist?
- Does the content meet the requirements?
- Were the originals modified?
- Do the links work?
- Was any input missed?
- Did it invent facts?
Ask for evidence:
List the files you actually read, the files you wrote, and the checks you ran.
Do not just say "done".
3.6 Delivering a File: What present Means
When the output is a file, the Agent may call present.
present means:
Declare an existing file as a final deliverable so the interface can show a file card that opens the current file.
It does not:
- Create the file;
- Upload the file;
- Copy the file contents;
- Put the contents into the model context.
It records paths and optional descriptions.
3.7 A Complete First Task
Prepare
Create:
notes/
a.md
b.md
c.md
Ask
Read the three articles under notes/ and combine them into notes/summary.md.
Requirements:
1. Keep each title and source link;
2. Write 3–5 sentences per article;
3. Do not modify the originals;
4. Give me a plan first;
5. After writing, verify that summary.md exists and tell me which files you read.
Review the plan
Make sure it covers all three inputs and does not modify originals.
Approve and watch
You should see three reads and one write.
Verify
Open summary.md; check titles, links, and length.
Iterate
If the second summary is too short:
Expand the second summary to four sentences, keep its link, and leave everything else unchanged.
Common Pitfalls
- Task too large: do one verifiable thing at a time.
- No acceptance criteria: the Agent cannot know what "good" means.
- Ignoring tool calls: it may edit the wrong file.
- Saying only "rewrite": give location and standard.
- Treating "done" as "correct": ask for evidence.
Summary
- A good task = goal + input + constraints + acceptance.
- Plan before executing.
- Watch write, delete, and command calls.
- Feedback must be specific; verification must be evidence-based.
presentis a delivery declaration, not an upload.
Part I Exercises
- Rewrite a recent request using goal + input + constraints + acceptance.
- In a test folder, ask the Agent to list files read-only and report the tools it used.
- Ask it to produce
summary.md, then check whether the originals changed. - Give one specific revision instruction instead of "redo it".
References
All links are public sources; product behavior may change between versions.