Anatomy of a Directed Edit — Worksheet
What you will do. For each of six task descriptions, decide whether the right fit is a suggester, a coding agent, or neither. For the agent-suited ones, write a directed prompt that exercises all four moves — locate, plan, write, verify — then predict whether you would reach for Cowork or Claude Code first, and why.
When done: keep this sheet. Lesson 3.2 reuses Task 2 as the first directed-edit exercise; Lesson 3.5 revisits Task 5 when scoping a larger change.
Reference — the directed-prompt shape
Every directed prompt names four things, in this order:
- Goal. What should be true at the end, in one sentence.
- Suspected location — or explicit permission to search. “The bug is probably in utils.py.” Or: “Find it first and tell me where.”
- Required plan step. “Propose a plan in one or two sentences before writing.”
- Required verify step. “Run the existing tests and show me the output.” Or: “Print the function’s output for these three inputs and confirm they match the goal.”
A prompt that names all four is a directed prompt. A prompt that skips any of them is a suggester request dressed as a directed one.
Worked example — Task 1
Task 1. Add a docstring to the parse_date function in utils.py.
CategoryAgent (borderline). A suggester could do it, but an agent can read the function first and produce an accurate docstring.
Directed promptIn utils.py, add a docstring to the parse_date function. Before writing, read the function body and tell me in one sentence what it appears to do. Then write the docstring, show me the diff, and confirm no existing test in tests/test_utils.py broke.First-choice tool
Cowork. Single-file, folder-shaped, low-stakes work; Cowork’s chat shape fits.
One-sentence reasonA suggester might hallucinate what parse_date does if it’s unusual; an agent that reads the function first produces an accurate docstring, and the cost of using the agent is negligible.
Tasks 2–6 — fill in your own
Write your answers before checking the key. Your reasoning is what this worksheet is teaching.
Answer key (cover this until you have tried)
Task 2 — sort-files .tar.gz bug
Agent. This is exactly the Lesson 3.2 starter. Cowork is fine for a single-file bug fix; Claude Code also works. Both are practiced in 3.2.
Task 3 — what does “ontology” mean
Neither. Factual question, not a code task. Ask a plain chat model (suggester-only) or a dictionary.
Task 4 — reorganize downloads
Agent. Folder-shaped, cross-file, requires actually moving files and reporting what happened. Tool: Cowork. Folder-shaped work is Cowork’s sweet spot. (Note: this smells like the tidy.py starter from Lesson 3.2 — deliberately so.)
Task 5 — React loading spinner
Agent. Multi-file, repo-shaped change with a test-pass verify signal. Tool: Claude Code. Cross-file changes with test runs are where Claude Code’s terminal-native flow earns its weight. You will revisit this in Lesson 3.5 as a scoping exercise.
Task 6 — email regex
Suggester — or, more honestly, a reference (every programming book has this). The “right” email regex depends on which RFC you follow and usually calls for a real library (e.g., Python’s email.utils) for anything serious. An agent is overkill and a suggester’s answer risks being subtly wrong for edge cases — prefer the library.
Reflection — three sentences
Sign and date
Where this worksheet ends up
Keep this page. Lesson 3.2 uses Task 2 as the first directed-edit exercise. Lesson 3.5 revisits Task 5 when scoping a larger change. If your answers on those two tasks change significantly after those lessons, that is a sign the lessons landed.