Vibeleaderboard
← Course overviewModule 03 · From answering to acting · Beta
Lesson 3.4

Module contents

From answering to acting

Lesson 4 of 5

Which work should I hand over, and how?

Not every task is worth giving to an agent, and a task that is worth it still needs a handover that says what counts as done. This lesson gives five questions for choosing a task, then builds a short written brief for the fencing comparison that later modules carry forward.

Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.

Five questions before delegating

Ask five questions about a task before handing it to an agent. Each one comes from something this course has already taught.

What are the inputs, and can the harness reach them? The work needs particular information: files in a folder the harness can read, a page it can fetch, a record in a connected account. Anything it cannot reach will be missing from the context, and module 1 showed what a model does with missing information. The scope of the inputs also bounds the answer. Module 2 described Simon Willison's January 2026 report, in which a search of his own website answered whether a draft was published there and could not answer whether it was published anywhere. 8 A search is an input with a scope, and the scope is part of the task.

What is the desired result? Name the artifact or the answer, its form, and where it should end up: a file at a path, a message shown to you, a change to a particular record. A result you cannot name is a result you cannot check.

Which operations are available, with what access? This is the previous two lessons: which tools the harness offers, which rules apply to them and what the connected accounts are authorized to do.

Which of those operations can be undone? Module 2's distinction between and irreversible operations decides which actions need an ask rule, which should be unavailable for this task, and which can run freely. Writing a new file into an empty slot is easy to undo. Sending a message is not.

How will you know it is right? Decide now what you will look at afterwards other than the final message: the action record, the output file's contents, a receipt from the service that performed an operation, a calculation you redo yourself.

A is one where all five answers can be written down before the loop starts. The fencing comparison is bounded: named inputs in one folder, one output file, reads plus a single write, a reversible result, and a check that means reading one file and redoing three sums. “Sort out the fence” is not bounded, because its inputs, result and allowed operations are all open. Bounding a task is not the same as making it small. A large task can be bounded if each of its parts can be, and module 8 returns to longer work.

Where the practitioner reports fit

Anthropic's AI Fluency course, a free set of lessons the company publishes about working with its assistant, asks learners to break a project into tasks and, for each one, discuss which parts benefit from human strengths and which could use the assistant's capabilities, recording the major tasks and delegation decisions as a plan they keep for later lessons. 6 That lesson was read, not completed, for this course, and its directions are specific to one product. The idea it contributes is that a delegation decision is made per task and written down. The five questions above are one way to make such a decision concrete when the work will run through a harness.

Google's sample data-analysis agent, mentioned in the previous lesson, shows how the fifth question changes what you can delegate. A question about sales per country is bounded when the steps between question and chart can be inspected: which tables were used, how rows were grouped, what units the totals are in. In the sample's example conversation, the reply describes the database step and does not show the query. 7 When a product shows only descriptions, your means of checking is weaker, and that should change either what you delegate or what you ask to see before trusting a chart.

A third report addresses the first two questions before the loop begins. Thariq Shihipar, a member of Anthropic's technical staff who works on Claude Code, published a guide in July 2026 about finding what a task's description leaves out. Among its practices: before implementation, ask the agent to help find your unknown unknowns; have it interview you one question at a time about ambiguities in the request; point it at examples of the behavior you want; and afterwards ask it for a quiz on the changes as a check on your own understanding. 9 The site's Intel collection holds a summary of the guide, and the author has a profile there. The guide presents these as ways for the person to discover missing requirements. Read it with that scope. A list of blind spots generated by the same model that will do the work is a prompt for investigation: each item is a lead to check, and the list's length says nothing about what it missed. The guide is one practitioner's advice for one product and one model family, and no comparison was run for this course.

The delegation brief

Write the five answers down and the result is a : a short document that states a task's inputs, its output, the actions allowed, the conditions under which the agent should stop, and the evidence that will show completion. Here is an authored brief for the fencing comparison.

Task: Compare the fencing quotes for the north boundary.

Inputs: the files in Garden/Suppliers. Parkside's quote is a web link given in
fencing-brief.txt.

Output: Garden/fencing-comparison.txt, one row per supplier, with the price
including installation, the delivery time and the guarantee.

Allowed actions: list and read files in Garden/Suppliers; fetch the Parkside link;
write the output file and nothing else.

Not allowed: sending messages of any kind; changing any existing file.

Stop and report if: a quote cannot be retrieved after one retry; a quote covers a
different length than 40 m; finishing would need an action not listed above.

Completion evidence: the output file exists and I have read it; the action record
shows each quote being read; every price in the table matches the quote it came
from, with installation added where a quote lists it separately.

Each line belongs to a layer from the previous lesson, and it matters which. The task, inputs and output lines are context for the model. “Allowed actions” and “Not allowed” are instructions until the harness's rules match them. The brief says no messages; the harness still has a send_email tool, and the previous lesson showed what an instruction alone does about that. So a brief is written and then the harness is configured to agree with it where it can: for this task, leave email out or deny it, ask before the write, allow the reads. Where the harness cannot enforce a line, the line stays an instruction and you know it.

A is a situation, named in advance, in which the agent should stop and hand the decision back to a person rather than choose a way around it. Stop conditions are where the loop returns to you. Without them the model decides for itself what to do about a failed fetch, and line 19 of the trace is what such a decision can look like.

is what you will check to accept the work, named before the work starts so that the check does not depend on the final message. The brief's version names the file, the record and a comparison you can do by hand. It also fixes a meaning the request left open: “price” means the installed price, so Hedgerow's £300 is added. A brief written before the loop is where such meanings get settled.

How the brief would have changed the trace

Run the authored trace again in your head with this brief and a harness set to match it. Lines 2 to 12 are unchanged. After the first timeout, one retry is allowed, so line 13 runs. After the second, the stop condition applies: the model should stop before writing and report which quote could not be retrieved. The organizer can then authorize a partial comparison or supply the missing quote. The harness cannot execute a send_email call because that tool is unavailable. The model could still suggest sending in words; no message would leave through this harness.

What the brief does not do is also worth stating. It does not guarantee the table is right; the completion evidence exists because it cannot. It does not stop the model from misreading a quote. And it is itself context, a document the model may weigh against other instructions in ways module 4 examines. Notice, too, that this brief says nothing about who needs the comparison, by when, or where the Garden folder lives. Those gaps are deliberate; they are where the next module begins.

Carry this forward

  • Choose a task by its inputs, desired result, available operations, reversibility and means of checking. A bounded task is one where all five can be written down before the loop starts.
  • Practitioner methods such as a blind-spot pass or an interview help find missing inputs and requirements. A list generated by the same model is a set of leads to check, not proof of completeness.
  • A delegation brief states inputs, output, allowed actions, stop conditions and completion evidence. Its “not allowed” lines are instructions until the harness's rules or an enforced boundary match them.
  • Completion evidence is named before the work starts, so acceptance never depends on the final message alone.

Lesson check

A fictional harness can list, read and write files in one folder and has no other connections. Which of these requests is a bounded task in this lesson's sense?

Choose an answer for each question.

This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.

A brief written this well can still meet a failure: a link that never answers, a quote that disagrees with the brief, an action nobody allowed. The last lesson is about what to do then.

Sources · Reviewed 2026-09-29
  1. 8. Simon Willison: First impressions of Claude Cowork, Anthropic's general agent
  2. 6. Anthropic AI Fluency: Project planning and delegation
  3. 7. Google ADK: archived data-science sample at b1910c6
  4. 9. Thariq Shihipar (claude.com): A field guide to Fable: finding your unknowns
  5. 10. VibeLeaderboard Intel: A Field Guide to Fable: Finding Your Unknowns

Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.