Skip to content

2.3 From Prompting to Delegating

"Prompt engineering" had its moment — magic phrases, secret templates, $49 PDF cheat sheets. Most of it aged like milk. What survived is something simpler and more durable: the practices of a good delegator. This lesson makes the shift from crafting clever prompts to writing briefs a capable worker can run with.

What you will learn

  • Explain why delegation is the right mental model for agentic work.
  • Write a brief using the five elements that reliably improve output.
  • Spot the three most common briefing failures — in other people's prompts and your own.

Builder principle

Talk to an agent the way a great manager talks to a talented new hire: full context, clear outcome, explicit constraints — then room to work.

Why "delegation" and not "prompting"

A prompt mindset asks: what words unlock the good output? A delegation mindset asks: what would a skilled worker need to do this job right?

The second question always wins, because it forces you to supply the things that actually determine quality — context, definition of done, constraints, examples. Magic words don't survive model updates. Good briefs survive everything, including the switch from AI to human helpers, because they were never about the AI. They were about clarity.

The five elements of a brief

Every strong agent brief contains some mix of these:

1. Goal — what does done look like? Not the task, the outcome. "Make the homepage load in under two seconds" beats "optimize the site." A testable finish line changes everything downstream.

2. Context — what does the worker need to know? Who's this for? What exists already? What was tried? Where are the relevant files? (In Chapter 5, most of this moves into CLAUDE.md so you stop repeating it — but the thinking starts here.)

3. Constraints — what are the rules? Budget, tone, tech choices, things not to touch. Unstated constraints are the #1 source of "technically what I asked for, not what I wanted."

4. Examples — what does good look like? One example outranks three paragraphs of description. "Write it like this sample" or "match the style of this file" collapses ambiguity instantly. Models are pattern engines (Lesson 2.1) — feed the pattern.

5. Verification — how will we both check it? Tell the agent how the work will be judged and it will judge itself against that standard while working. "Run the tests when done." "Check it on a phone-sized screen." This one element separates casual users from professionals.

Before and after

Before: Write me an About page for my bakery.

After:

Write the About page for Rise & Grind Bakery (Oakland, family-owned
since 2019, known for vegan cinnamon rolls and hiring locally).

Audience: neighborhood customers and local press.
Tone: warm, proud, no corporate speak — read our Instagram bio
below and match that voice. [bio pasted]
Length: ~200 words, two short sections: Our Story / What We Believe.
Done means: I can read it out loud and it sounds like my aunt
talking about the shop — not a brand deck.

Same request. Different universe of output.

Before: Fix the flaky test.

After:

tests/api/test_orders.py::test_concurrent_checkout fails ~1 in 5
runs on CI (passes locally). Suspect a race on inventory decrement.

Goal: make it deterministically pass without weakening the assertion.
Context: checkout logic in services/orders.py; CI runs pytest -n 4.
Constraints: don't add sleeps; don't mark flaky/skip; don't change
the public API of OrderService.
Verify: run the test 20x locally (pytest --count) and show output.

Note the verification clause — the agent will now prove the fix instead of declaring it.

The three classic failures

  1. The mind-reader assumption. Critical context lives in your head and never reaches the brief. Symptom: "that's not what I meant." Cure: element 2.
  2. The vague finish line. No definition of done, so the agent picks one — and its guess is generic. Symptom: technically-complete-but-mid output. Cure: element 1.
  3. The kitchen sink. Ten goals in one brief. Focus dilutes (remember the whiteboard from 2.2), everything comes back half-done. Cure: one brief, one outcome. Queue the rest.

Try it now

Take your quick win build from Lesson 1.4 and write its first real brief using all five elements. Keep it under 150 words — brevity forces prioritization, and you'll feed this exact brief to Claude Code in Chapter 4.

Then grade yourself: which element was hardest to write? That's the muscle to watch.

Check your understanding

  • Why does the delegation mindset outlast prompt tricks?
  • Name the five elements. Which one makes the agent self-correct while working?
  • Which classic failure do you personally commit most — and what's its cure?
  • Why does one example often beat paragraphs of description?