Picking a target worth the clock

You have a kit, a mock backend that streams convincingly, and a written cut list. The remaining input to a practice run is the blank Target product field at the top of the log entry, and it decides more about what the run measures than anything you do afterwards. Fill it badly and three hours produce a row in a table and no information.

There are two ways to fill it badly, and they fail differently. A target that is too small — a chat box with a system prompt — leaves you two hours of clock and no decisions to spend them on, so the run quietly becomes an afternoon of styling. A target that is too vague — “something for a logistics company” — spends hour one on the specification you were supposed to arrive with, and the phase table records that as domain comprehension when it was really indecision.

Four tests a target has to pass

This is this course’s own screen, not a finding from anywhere. It exists because the runs have to vary in the ways that matter and hold constant in the ways that do not.

• There is a decision a human must accept, correct or reject before anything downstream fires. Not a confirmation dialog. A real judgment the agent got roughly right and a person has to adjust, with the evidence for it visible next to the control.
• You do not already know the domain. If you can name the entities and their relationships from memory, the run measures recall. Freight demurrage, clinical prior authorisation, lease abstraction, insurance subrogation — the interesting ones are the ones where you have to read something first.
• There is structure to render, not just text. An agentic surface earns its complexity from extracted fields, confidences, source references and a run trace. A stream of prose has none of that and skips the whole part of the build worth timing.
• It is unlike the previous run’s target. Rule 5 of the practice log. Six variations on invoice approval measure how well you remember invoice approval.

A fifth, softer one: the failure case should be interesting. A target where a wrong answer is obviously wrong teaches you nothing about designing the correction step, which is where most of the real design work in an agentic surface lives.

What deliberate practice actually requires

Repetition against a clock is not automatically practice, and it is worth being precise about the difference rather than gesturing at the popular version. Ericsson — whose work the ten-thousand-hours story is a distortion of — restated his own criteria in a 2019 paper with Kyle Harwell. Four conditions: individualised training designed by a well-qualified teacher; a communicated goal the trainee can internally represent; an activity giving immediate feedback; and repeated revised attempts approaching that goal.

Map the six runs onto them honestly, including the weak spot:

And the popular version, since somebody will bring it up: Ericsson himself pointed out that Gladwell never used the term deliberate practice and folded in activities such as public performance and ordinary work that violate the criteria. The ten thousand hours was a group average for the top tier of Berlin violin students at age twenty, against roughly 7,800 hours for the next group down. It was never a threshold. If you invoke it, you are invoking something Ericsson disputed.

The target bank

Pick targets in a batch, in advance, when the clock is not running. Choosing under pressure costs clock, and worse, it biases toward whatever you already know how to build — which is precisely the target that measures nothing.

Six candidates, six domains, each with the approval step named. Shape yours like this:

Domain                 Agent does                    Human approves
Freight billing        Flags demurrage overcharges   The disputed line items
Clinical intake        Drafts prior-auth packets     Codes and clinical rationale
Commercial leases      Abstracts renewal terms       Extracted dates and options
Claims subrogation     Ranks recovery candidates     Which files get pursued
Field service          Builds a dispatch plan        Route and parts substitutions
Trade compliance       Classifies goods for tariff   The HS code and its basis

Each row is unfamiliar enough to need reading, structured enough to render, and carries a decision a person has to own. None of them is a chat interface, and none of them is a domain you have shipped.

Four ways a target goes wrong

• The general assistant. “An agent that helps with X.” No approval step, no artifact, and the build collapses into a message list. The run ends early and you fill the time with styling.
• The one you have opinions about. If you already know the schema, hour one disappears and the run looks fast for a reason the kit had nothing to do with. That contaminates the comparison against your baseline as badly as starting from an existing project would.
• The visual problem in disguise. A target whose hard part is a chart, a canvas or a diagram. Real work, wrong course. The clock goes to rendering, and the log records nothing about agentic surfaces.
• The one with nothing to get wrong. If the agent cannot plausibly produce a bad answer, there is nothing for the human to correct, and the correction interaction is the part you are here to get fast at.

What this does not cover

The bank says what to build against. It says nothing about what to build with, and the runs are where that question gets answered — v0, Claude Code, Cursor and Figma Make will come up in the room, and an opinion assembled from vendor pages is worth less than nothing there. Building a position from your own runs, and knowing what does and does not count as evidence for it, is the next lesson, An evidenced position on the tools.

Nor does this lesson tell you what six filled-in rows mean once you have them. Comparing runs whose targets deliberately differ is harder than it looks, and the reading of the log gets its own lesson at the end of this module.