← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

Anthropic’s AI
Agent Playbook

The team that builds Claude shared how they actually work alongside AI agents. Here are the four rules, made practical for you.

In June 2026, Anthropic published Building effective human-agent teams, a look at how its own teams work with Claude agents. The line I keep coming back to is "Work now looks a lot more like a multiplayer game, with teams of humans setting the strategy, and Claude executing the work."

Here are the four rules to steal, whether you run one AI assistant or a whole team of them. Each comes with a prompt you can paste into Claude or ChatGPT today.

Rule 1: Write Everything Down

01

Give it the full picture

An agent can’t read your mind or overhear a hallway conversation. Anthropic puts it bluntly. "For an agent, if it’s not written down and accessible, it doesn’t exist."

The more you put in writing (your goals, your notes, how you like things done), the better it gets. Most people fail right here. They assume the AI knows more than it does, then blame the AI.

This prompt interviews you and builds a brief you can reuse in every chat.

The How I Work Operating Brief
You are going to help me build a reusable operating brief about me and how I work, so that you and any AI I work with can give me genuinely tailored help instead of generic answers. This brief becomes the context you use for everything we do together. First, interview me. Ask me one focused question at a time to fill in anything missing or vague below, and don't move on until each answer is specific. Dig where my answers are fuzzy. Then write the finished brief in clean, labeled sections: 1. WHO I AM: my role, my business or work, who I serve, and what a typical week looks like. 2. WHAT I'M TRYING TO ACHIEVE: my top 1 to 3 goals right now, the timeframe, and how I'll know I succeeded (the actual numbers or outcomes). 3. HOW I LIKE THINGS DONE: my tone and voice, the formats I prefer, my standard for what "good" looks like, and a clear list of things to always avoid. 4. WHAT ELSE YOU SHOULD KNOW: my strengths, the things I find tedious or get stuck on, and where I most want your help. 5. TOOLS & RESOURCES: the tools, accounts, files, and assets you can draw on. When it's done, give me the brief in a tight format I can save, and tell me exactly how to paste it back in at the start of any future chat so I never have to re-explain myself. Here's what I can tell you to start (fill in or correct me): - What I do: [role / business] - What I'm trying to achieve: [the bigger goal] - How I like things done: [tone, format, standards, what to avoid] - Tools and accounts you can use: [list]
Rule 2: A Role And A North Star

02

Stop throwing random tasks

Anthropic gives every human and agent a defined role with the right tools for the job, then sets a north star, an ambitious goal that helps decide which work matters.

For you, that means telling the AI who it is, what it can use and the bigger goal. Once it knows the destination, it starts suggesting things you didn’t think to ask for.

The Role And North Star Charter
I want you to operate as a proactive teammate, not a task-taker. Set up our working relationship as a clear operating charter, then run by it from now on. YOUR ROLE: You are my [role, e.g. marketing assistant / chief of staff / research analyst]. YOUR TOOLS AND ACCESS: [what you can use, e.g. my Google Drive, my email drafts, web search]. If a job needs a tool you don't have, tell me instead of faking it. MY NORTH STAR: The one bigger goal we are working toward is [the big goal, e.g. grow my email list to 50,000 subscribers this year]. Every suggestion you make should ladder up to this. HOW I WANT YOU TO OPERATE: 1. Think in priorities, not tasks. At any moment you should know the top 3 highest-impact things we could be doing to move the north star. 2. Be proactive. Don't wait for me to hand you work. Propose the next best move, and tell me WHY it matters and what it should produce. 3. Push back. If I ask for something low-impact, or you see a better path, say so directly before doing it. I want a teammate with a spine, not a yes-machine. 4. Flag what you need. If you're missing context, a tool, or a decision from me, ask for it specifically instead of guessing. TO START: Based on the north star above, ask me any questions you need. Then give me a prioritized game plan, the 3 highest-impact moves to make next, why each one matters, and what success looks like for each. Finally, recommend the single one we should start on today, and what you need from me to begin.
Rule 3: Never Trust It Blindly

03

One does, one checks

Anthropic points to what’s often called a Doer-Verifier setup. The idea, in their words, is to "give one agent the job of doing the task and another agent the job of checking the first agent’s work."

The checker works best when it’s fresh. So paste the finished work into a new chat (or have Claude Code spin up a subagent) with the prompt below. A reviewer that never saw the reasoning behind a mistake is far more likely to catch it.

The Fresh-Eyes Verifier
You are a skeptical reviewer. You did NOT do this work and you did not see the reasoning behind it. Assume it contains mistakes until you prove it doesn't. WHAT I ASKED FOR: [paste the original request or task] THE WORK TO CHECK: [paste the draft, or attach the file] WHAT IT'S FOR: [who will see it and what happens if it's wrong] Check it against three things: 1. Completeness: did it do everything I asked, or only part of it? 2. Accuracy: is anything wrong, missing, outdated or made up? Check facts, numbers, names, dates, links, logic and any claim that sounds too confident. If you can't verify something, say so instead of guessing. 3. Quality: is this good, or just acceptable? Would it impress the person it's for, or embarrass me? Give me: - A ranked list of every issue, most serious first, each with the exact line or section and why it matters. - The corrected version, with your fixes applied. - A short "what almost slipped through" note at the end. Don't tell me it's fine without running the full check. If it really is fine, say which checks it passed and why.
Rule 4: Build Trust Slowly

04

Start small, then hand over more

Don’t hand an agent everything on day one. Anthropic says its teams "grant agents autonomy in proportion to demonstrated reliability, then expand it deliberately."

Review closely at first and give feedback. As it nails one kind of task, let it run more of that on its own.

The Trust Ladder For Delegating
Help me safely hand off more of my work to you over time, the way strong teams grant autonomy in proportion to proven reliability. I don't want to micromanage forever, but I'm not going to trust you blindly on day one either. Here is the work I eventually want to delegate: [describe the task or area, e.g. drafting and scheduling my emails]. Design a 4-level trust ladder for this work, and explain exactly how we move up it: - LEVEL 1: You draft, I approve everything before anything happens. - LEVEL 2: You do it, I review before it goes live, but you flag the risky parts for me. - LEVEL 3: You handle routine cases on your own and only escalate the unusual ones. - LEVEL 4: You own it end to end and simply report back. For each level, tell me: what you'll do, what you'll still run past me, and the specific track record you need to earn before we move up (how many times, and at what quality). Then tell me which level we should honestly start at today, and what you'll show me each time so I can verify your work and build trust fast. Keep me honest: if I try to skip ahead before you've earned it, remind me why we climb in order.

The Idea Under All Of It

You set the strategy and the AI does the execution. The people who win won’t be the ones with the fanciest prompts. They’ll be the ones who learned to run the team. Read Anthropic’s full post ↗