Chapter 05 of 8
Asking for work.
There's no secret phrasing. Brief Claude like a sharp new hire, ask for the plan on big jobs and proof on everything, and let your recorded calls write the rest of your briefs for you.
Most people's first requests to Claude sound like a text to a friend: "can you help me with our website?" Claude will happily help, and what comes back will be vague, because the ask was vague. There's no secret phrasing that fixes this. What fixes it is asking the way you'd brief a sharp new hire on their first day, someone clever and fast who knows nothing about your business yet.
This chapter is about that brief: what goes in it, when to ask for a plan instead of the work, and why you should never accept "done" without seeing it. Then it turns to the best brief of all, the one you never have to write, because you already said it out loud on a call.
The four parts of a good ask.#
A good ask has four parts, and none of them are clever. The first is the outcome: what exists when this is done? Not "work on the pricing page" but "a new pricing page, live on the site." The second is context, whatever Claude can't know on its own: the prices are in the attached sheet, our customers are contractors who read on their phones, the last agency made it look cheap.
The third is the bar, which is how you'll judge the result. "It should look like it belongs on the best site in our industry" sets a very different target than silence does. The fourth is boundaries, the lines not to cross: don't touch any other page, show me a preview link first, don't make anything live without me. Stacked together, they look like the figure below, and each one fits in a sentence.
The four parts of a good ask
Each part, with a one-line example from the same pricing page job.
Notice what's missing from that stack: the steps. Don't tell Claude how to do the work unless you genuinely care about the how. When you dictate the steps, you get your plan done well. When you describe the outcome, you get Claude's best plan, and it's often better than yours.
Here is the difference in practice. The weak version:
Can you help me with our website?
And the strong version of the same request:
Our homepage gets traffic but almost no form fills. Find out why, fix the top three problems on a preview link, and show me before and after at phone size.
The second one names the outcome, gives the context, sets the bar and draws the boundary, all in two sentences. It takes ten seconds longer to say and saves an hour of back and forth. One more tip: talk instead of typing. Use your Mac's dictation or the voice input in the Claude app. Talking gives Claude more context in less time, and more context beats clever wording every time.
For big jobs, ask for the plan first.#
Those four parts are enough for anything that takes minutes. For anything that will take hours, like a new website or moving a business off an old system, don't ask for the work yet. Ask for the plan. The desktop app has a permission mode built for this called Plan mode, where Claude reads, researches and proposes, and changes nothing until you approve. Have it write the plan into a file called PLAN.md in the project folder: what it will build, in what order, the shortest possible list of things it needs from you, and how it will prove each part works.
How a big job runs
Four beats, and you only have to be there for one of them.
As the figure shows, your job sits in the middle. A written plan catches misunderstandings while they're still cheap, before they've cost you a day. It also outlives the session: because PLAN.md lives in the project, the next session reads it and carries on where the last one stopped, even days later or on the other machine. Once the plan looks right, say go, and tell Claude to run the whole thing without checking in unless it hits something on your never-without-me list, which the next chapter covers in full.
When you dictate the steps, you get your plan done well. When you describe the outcome, you get its best plan.
Always ask for proof.#
Whether the job took ten minutes or three days, it ends the same way. "Done! Everything should work now" is a sentence, not evidence. Build proof into the ask from the start. That means a screenshot at phone size, which Claude takes by opening the page in its built-in browser at a phone's width; the preview link; the test output, which is the pass or fail report from the automated checks it ran; and the before and after numbers. It costs Claude a minute and saves you from discovering on Monday that "done" meant "written but never tried."
Proof also changes how Claude works. When it knows it has to show a working page on a phone, it checks the page on a phone before telling you it's finished, and catches its own mistakes along the way. The single most useful habit in this whole book is also the shortest: before you accept anything, ask to see it.
Your calls are already the brief.#
Everything so far assumes you sit down and write the ask. Often you don't have to. People say more in ten minutes of talking than in an hour of typing. On a call, a client explains their business, what annoys them, the numbers, the deadline and what the last agency got wrong, all without being asked. Normally most of that evaporates the moment you hang up.
Claude is excellent at exactly the part people are worst at: reading a long, rambling conversation and pulling out every decision, number, name, promise and open question, with nothing invented. Hand it the transcript file, whether that's the document Google Meet saves to Drive or the transcript Zoom produces with a cloud recording, and there's no prompt to write. The call is the prompt.
There is one rule here you can't bend, though: transcript, not summary. A summary is somebody's opinion of the call. A transcript is the call. The AI meeting notes built into most video tools are summaries, and they read beautifully, which is exactly what makes them dangerous.
A new client, two calls. The first was fully recorded and transcribed, and from it Claude caught every commitment word for word. On the second call, only the AI notes were switched on. They looked tidy, but they paraphrased. Numbers got rounded, names dropped out, and a client's "we'll probably want a booking page" quietly became "the client wants a booking page." Same client, same Claude, and a much shakier record of what was actually agreed.
Since then the rule has been simple: every call gets a full recording and a word-for-word transcript, switched on before it starts. If all you have for a call is the AI notes, tell Claude so, and have it mark anything it writes from them as based on notes, not a transcript.
Record everything, and say so.#
If your company uses Google Meet, recording and transcripts need Google Workspace Business Standard or higher. The cheaper Business Starter plan doesn't have them. Whoever runs your Google account can turn both on for the whole company in the Google admin settings, under the Meet video settings, and set them to start automatically for every meeting your people organize. Afterward, the recording and a transcript document land in a "Google Meet" folder in the organizer's Google Drive, where Claude can find them on its own once the Drive connector is switched on. If a client hosts the call on their own Zoom or Meet, your settings don't apply, so ask them to record it.
Then the part that matters more than any setting: always tell people you're recording. Federal law in the US needs only one person on the call to agree, but about a dozen states, including California, Florida, Illinois, Pennsylvania and Washington, require everyone to agree, and so do many countries. Saying "I record calls so I don't miss anything, is that all right?" at the start covers you everywhere. Almost nobody minds.
What you get back.#
When the call ends, you ask for everything at once, and the transcript flows through Claude into four places, as the figure shows. A recap for the client's file, with who was there, what they want in their own words, and every number, decision and promise, quoting exact phrases for anything that could be argued about later. A short follow-up email, drafted in the existing thread and left unsent. A task list with an owner and a due date on every item. And the transcript itself, filed in the client's record.
From a call to finished paperwork
What happens to one recorded call, left to right.
Read the full transcript of today's call with the Parkers. Give me the recap, draft the follow-up in their email thread (don't send it), and list every task with an owner and a date. Then start anything that's clearly ours.
That last line is where it gets good. Anything that's plainly your job and doesn't need the client's input, Claude can start right away, with the same habits as any other ask: a plan first if it's big, proof when it's done. And because every transcript is kept, "what did they say about pricing across all our calls?" takes ten seconds three months later instead of never.
You don't have to write a brief. You already said it out loud.
Voice notes count too.#
The same trick works when there's no one on the other end. Some of the best briefs are five minutes of you talking to your phone in the car. Recent iPhones can transcribe Voice Memos and Notes, and the Claude app has voice input of its own. Ramble on purpose: what you want, why, what bugs you about how it works now, what great would look like. Claude can't open your Apple Notes by itself, so share the transcript to Claude or paste it in and say what you want done with it.
One morning a partner left a short voice memo: move a client's old website off its outdated builder, give it a facelift, automate the weekly market report the client built by hand every Friday, and replace their clunky PDF forms. Claude read the transcript and didn't ask a single question it could answer itself. By that afternoon there was a rebuilt site on a preview link, a page where dropping in a PDF publishes the weekly report automatically, 33 old reports archived, and proper online forms in place of the PDFs. The voice note was the entire brief.
A rhythm for the day.#
Put all of this together and a working day with Claude settles into a simple rhythm, sketched in the figure below. At seven, the morning brief from the overnight work is waiting. In the car, a voice note becomes the first brief of the day, and Claude starts on it in its own session. Calls happen as they always did, except each one ends with the transcript going straight to Claude. In the afternoon you review previews and proof, and at night you hand off the long jobs to run while you sleep.
A day with Claude
Who is working when, 7am to 10pm. Approximate.
Give each job its own session so they don't trip over each other, and check on each when it pings you. Every time something comes back, look at the proof before you say yes. And when it comes back not quite right, don't just accept it. Push back, and save the lesson so you never have to say it twice. That's the next chapter, and it's the most important one in the book.
If you do only one thing from this chapter this week, make it the recording. Turn on recording and transcripts, and hand the next transcript to Claude. You won't go back.
Common questions.
How do I prompt Claude?
Describe the result you want, not the steps. Add the background Claude can't guess, say how you'll judge it, and name anything it must not do. Talking works as well as typing.
How do I use Claude Code effectively?
Put your standing rules in CLAUDE.md so you don't repeat them. Give each session one job, ask for a plan before big work, and ask for proof that it works.
What are Claude prompt best practices?
Be specific about the result, share examples or files to follow, and give Claude a way to check its own work, like numbers that must tie. More background beats clever wording. Anthropic also publishes a prompting guide for deeper technique.
Should I use plan mode?
Yes, for big jobs, or when you're not sure how it should be done. In plan mode Claude writes a plan and changes nothing until you say go. For a small, clear fix, skip it.
What do /compact and /clear do in Claude Code?
/compact shrinks a long conversation into a summary so there's more room. /clear starts fresh. Use /clear between unrelated jobs and /compact when one long job needs more room.
Can Claude take meeting notes?
Claude doesn't sit in on the live call. Your meeting tool records it and writes the transcript. Claude then turns that transcript into notes, decisions, tasks and a follow-up draft in a few minutes.