Practice 01: Brief Before You Delegate
The one-sentence brief that goes wrong
Ask for "clean up our Q2 sales data and build a summary" with no further brief, and an eager, literal agent might restructure columns that were needed and drop a breakdown the board expects, then hand back a neatly formatted summary of the wrong thing. Nothing was technically "wrong" with the agent; the brief was one sentence long. The mistake surfaces later, once the clean file and the summary already exist and both need redoing.
The mechanism: Plan Mode
Claude Code has a Plan Mode: given a task, it reads the relevant files and proposes an approach, without changing anything, until the plan is approved. Use it whenever the task matters.
Turning it on is one keystroke, pressed while Claude Code is running: Shift+Tab. The terminal shows plan mode on to confirm. Then type the brief exactly as normal, for example:
Clean up our Q2 sales data and build a summary.
The agent reads the files and replies with its plan, in text, right there in the terminal, instead of touching anything yet. Reading the plan and typing a one-line correction, if one is needed, is done the same way: plain English, back in the same terminal.
The agent reads the relevant files. Nothing is changed yet.
It reports back exactly what it intends to do, including any judgment calls it made along the way.
A correction here costs one sentence. Only once the plan is approved does the agent act.
Catching it before it happens
Run the same one-sentence prompt in Plan Mode instead, and the agent reads the files, then reports back before touching anything: which files it will use, how it will de-duplicate records, and what the summary will contain, including the fact that it plans to drop the region breakdown because that column looked sparse. That's the moment to catch it. One sentence back, "keep the region breakdown," and the plan updates. Only then does the agent act.
Dropping the region breakdown costs about ten seconds to fix in the plan, before anything changes. Caught after the work is already done, the same fix means re-running the clean, rebuilding the summary, and redoing the board cut. There's no measured study behind that gap; it's a plausible, illustrative estimate of what re-running a clean and a summary from scratch tends to cost, offered as an order of magnitude, not a precise, sourced number.
Why it works
The mechanism is a gate: the agent reads your files, proposes a plan, and cannot act until you explicitly approve it.
Misunderstandings are cheap in the plan and expensive in the output, exactly like briefing a new hire. A correction at the plan stage costs a sentence. The same correction after the work is already done costs the rework, plus whatever damage the wrong output caused before anyone noticed it was wrong.
When to use it
Use Plan Mode every time the task matters. For anything touching real business data, that's every time.
Practice 02: the three properties of a well-scoped task, and what to do when a good task turns out too big.
Reply here and it goes straight to Rod. Same as replying to one of his emails.
