By this point in the series you know how to prompt well. The next level of mastery is not another technique — it is a working relationship. Two people can use the same tool and get very different results: one treats Claude as a search box and gets generic answers; another hands over everything, including decisions that needed their judgment, and ships work they never really evaluated. Both are using Claude. Neither is collaborating with it.
AI-human collaboration means deciding, task by task, what to delegate, what to work on together, and what to keep — then moving work back and forth through short draft-and-refine loops with specific feedback. Delegate effort, never accountability: use Claude for volume and first drafts, as a reviewer and challenger of finished work, and as a thinking partner, while the final judgment stays with you.
Real collaboration is closer to managing a very fast, very well-read colleague than to operating software. That colleague has no memory of yesterday unless context is provided, cannot see priorities unless they are stated, and will produce confident-sounding work whether or not it is right. This chapter covers how to divide the work, run the loop, use Claude to examine work, and build a rhythm that improves over time.
1 How Do You Decide What to Delegate to Claude and What to Keep?
Before writing a prompt, decide the division of labor. Most poor results trace back to a mismatch at this stage, not a weak prompt.
A Simple Sorting Test
For any task, three questions settle who does what.
Three Modes of Working
The answers map onto three modes, each with a different balance of effort between Claude and the person.
Formatting, summaries, first drafts, data cleanup.
Strategy, writing with a point of view, problem-solving.
Final decisions, sensitive conversations, anything where your judgment is the point.
Stating the role up front keeps Claude from jumping to a conclusion that wasn't asked for. The same question gets a different, more useful answer when the mode is named.
2 How Does a Draft-and-Refine Loop Work?
The best results rarely come from a single prompt. They come from a short loop where each round makes the work more specific to the person doing it.
The Basic Loop
Giving Feedback That Moves Work Forward
Vague feedback gets vague changes. Naming what to keep is as important as naming what to change — without it, a revision can quietly undo parts that were working.
Produces change without direction — Claude has to guess what "right" means.
Names the problem, the part to protect, and the change wanted.
Knowing When to Stop
The third-revision rule: If the third revision is not clearly better than the second, the problem is usually the brief, not the draft. Step back and rewrite the brief instead of polishing further.
3 How Can Claude Act as Reviewer, Challenger, and Thinking Partner?
Many people use Claude only to produce. Some of its most valuable uses are to examine work that already exists.
Assigning a specific viewpoint makes the review sharper than a general request for feedback.
A challenger is useful precisely because it has no stake in being agreeable. Ask for disagreement explicitly, because the default tone leans cooperative.
Sometimes the most useful thing Claude can do is not solve the problem yet.
4 How Do You Build a Collaboration Rhythm That Gets Better?
Collaboration becomes powerful when it stops being improvised each time.
Keep Your Context Reusable
Write down the background that keeps getting re-explained: role, audience, style, standards. Keep it in a short note to paste at the start of a session, or in a project where it persists. Repeating it costs minutes; not having it costs consistency.
Capture What Works
When a prompt gives excellent results, save it with a note on when to use it. Over a few weeks that becomes a personal library more valuable than any generic list.
Review Your Own Habits
Once a month, look back at recent work and ask where delegation is off in either direction.
Both errors are common, and they point in opposite directions: doing too much by hand, and trusting too much without checking. A monthly look back catches whichever one is drifting.
5 What Are the Most Common Collaboration Mistakes?
The same five mistakes account for most weak results from working with Claude, whatever the task.
Handing over judgment along with the work. Delegate effort, not accountability. You own the result.
Doing everything yourself out of habit. Volume, formatting, and first drafts rarely need your direct effort.
Vague feedback loops. "Make it better" produces change without improvement. Be specific about what to keep and what to change.
Never asking Claude to disagree. If every response feels agreeable, the challenge needed probably isn't arriving.
Re-explaining context every time. Capture it once and reuse it.
- Decide the division of labor before writing the prompt — most poor results start with a mismatch there
- Use the three-question test: private context, accountability, or volume and first-draft effort
- Name the mode up front — delegate, collaborate, or retain — and state Claude's role
- Run short loops: brief, draft, react, refine, decide — and rewrite the brief if round three isn't better than round two
- Give feedback that names what to keep as well as what to change
- Use Claude to review, challenge, and ask questions — and ask for disagreement explicitly
- Save reusable context and winning prompts, and audit your own delegation habits monthly
Foundation Option
Take three tasks from the coming week and sort each into Delegate, Collaborate, or Retain using the three-question test. Run one of them with the role stated up front.
Advanced Option
Take a piece of work already finished and run it through the Reviewer and Challenger prompts. Note which objections had not been considered.
Reflection Questions
Which kind of task tends to be over-delegated, and which over-kept? Where has vague feedback cost extra rounds of revision? What background would be worth saving to stop re-explaining it?
Ask three questions about each task. Does it require context only you have? If so, provide that context or keep the task. Does it carry consequences you are accountable for, such as a decision or a message to a customer? If so, you make the final call. Is it mostly volume, structure, or first-draft effort? If so, delegate it freely.
Delegating means Claude does the work and you review the result, which suits formatting, summaries, and first drafts. Collaborating means working back and forth, which suits strategy and writing with a point of view. Retaining means you do the task yourself, perhaps with Claude as a sounding board, which suits final decisions and sensitive conversations.
A draft-and-refine loop is a short cycle of five steps: you brief the goal, audience, and constraints; Claude drafts; you react with specific feedback; Claude revises only what you pointed to; and you decide whether the work is ready. If the third revision is not clearly better than the second, the problem is usually the brief, so rewrite it instead of polishing further.
Be specific about what to change and what to keep. A comment like "make it better" produces change without improvement, while naming the exact problem, the parts to leave untouched, and the desired adjustment gives a targeted revision. Naming what to keep matters because a revision can otherwise quietly undo parts you liked.
Assign a specific viewpoint and ask for disagreement explicitly. For example, ask Claude to review a proposal as a skeptical finance director and list the weakest points and likely questions, or to argue against your plan as strongly as possible and name the most serious objection. The default tone leans cooperative, so the request for challenge has to be stated.
Coming Up Next: In Chapter 27, MCP Connectors & Integrations, learn how to connect Claude to the tools and data already in use, so it can work with real information instead of whatever gets pasted into the chat.