by Wondering how to set up autonomous AI agents that complete work in the background without constant babysitting? Want a step-by-step process for training AI to run your workflows on its own and even manage other AI agents?
In this article, you’ll discover how to build AI playbooks you can trust to run autonomously, remove yourself from the review loop, and set up an AI middle manager that oversees your other AI agents.
Why Trust, Not Agent Building, Is the Key to Autonomous AI
Building an agent doesn’t make AI autonomous. According to Rachel Woods, AI can only run in the background without supervision once it can be trusted to do the work the way it’s supposed to be done. Without that trust, an agent runs off doing random things, and someone has to go back and undo that work.
The fix is to document the work step by step in what Rachel calls a playbook. When AI follows a playbook exactly, the result is predictable, and that predictability is what makes it possible to stop checking the work. Across industries and roles, Rachel hasn’t found a type of work that can’t be playbooked, and on an average day, she gets about 8 hours of autonomous work from AI.
#1: Build an AI Playbook Using Four Components: Trigger, Inputs, Steps, and Output
An AI playbook is a step-by-step process that AI follows to complete one specific type of work the way its owner wants it done. Most people haven’t yet worked out the structure of their own work, so figuring out that structure is the hard part. For example, connecting AI to an email account and asking it to get to inbox zero produces work that has to be checked tag by tag, with constant feedback that often contradicts itself.
Rachel’s email inbox triage playbook, which runs every hour and keeps her at inbox zero, started with a playbook outline.
How to Scope and Stack AI Playbooks
The most difficult part of the process is setting the bookends of a playbook. The common trap, and one that also arises when delegating to people, is defining something too large, such as having AI run an entire morning or building a single playbook for a whole marketing campaign.
A marketing campaign, for example, includes many deliverables: briefs, emails, social media posts, landing pages, follow-up lead magnets, and often several channels. One playbook for all of it won’t produce value or earn enough trust to run autonomously until every piece works correctly.
The better approach is one playbook per deliverable, each working independently, with an overall playbook that ties them together. Rachel calls this stacking, and it’s also how AI ends up managing itself. Her morning routine playbook triggers other playbooks and collects their outputs, a setup built brick by brick rather than overnight.
Pro Tip: The outline comes first, and the detailed instructions get filled in later, rather than writing the whole playbook from start to finish like a book written from chapter one.
How to Build a Playbook Outline in 5 Steps
- Pick one type of work that’s clear enough to hand off to AI.
- Define the trigger, which is what kicks off the playbook. It can be a schedule, such as the hourly run of the email playbook, or an event, such as a message sent to the AI or a new calendar invite.
- Define the inputs, which are what the AI needs to do its job. For the email playbook, the input consists entirely of new emails.
- Write the steps as instructions a person would need to do the work correctly. For the email playbook, the steps were based on how Rachel would ideally teach someone to work through her inbox, using her decision-making process and message tagging.
- Define the output, or what the finished result looks like. For the email playbook, the entire inbox is triaged and handled, responses are drafted, and anything the AI can’t figure out is escalated to Rachel to review later.
The trigger, inputs, instruction steps, and output description components apply to outlining any type of work.
#2: Use the CRAFT Cycle to Build and Improve Each AI Playbook
The CRAFT cycle is Rachel’s five-step process for building an AI playbook and improving it until AI can run it without review. CRAFT stands for clear picture, realistic design, AI-ify, feedback loop, and team rollout. The same five steps apply to every playbook and are repeated to improve it over time. The cycle avoids the all-or-nothing outcome that a new playbook tends to trigger: either reviewing its output forever or deciding it doesn’t work and abandoning it.
The first CRAFT cycle is useful even for workflows that someone still runs and reviews every day. A second cycle on the same playbook is what makes the work autonomous.
The first version usually keeps a human in the loop. Running it shows what that person checks for and which mistakes the AI makes, and those observations reveal where the playbook needs to be more specific.
Go Deeper Than an Article Can Take You
Join thousands of small-business marketers at Social Media Marketing World — three days of AI, social media, and marketing strategy taught by world-class experts. April 1–3, 2027 in Anaheim.
Every session is pitch-free and built around tactics you can use yourself, no big team required. Your All-Access ticket includes an extra day of workshops, full access to AI Business World, and recordings of everything.
GET MY ALL-ACCESS TICKET — SAVE $900The second cycle starts with a new, clear picture built on what was learned, then changes the playbook and removes the human from every step, where the AI now has what it needs to succeed.
These cycles go quickly once a person gets into the rhythm of running them.
Clear Picture: The playbook gets written in its most ideal form, documenting the full process being handed off. The outline is the starting point, with more specific, nuanced instructions added to steps that need them, especially when the work itself is nuanced.
Realistic Design: The first version gets scoped down. Steps that will be easy for AI go to AI, and a human stays in the loop on the harder ones. The goal is a playbook that’s working at some level, so there’s something to learn from and improve. A roadblock isn’t a reason to abandon the playbook. AI keeps improving, so a step that isn’t realistic now may be in six months, and AI can handle the rest of the process in the meantime.
AI-ify: The playbook gets set up in an AI tool and put into use. This is where its owner starts confirming the AI can do what’s expected.
Feedback Loop: Rachel considers this the most important step. Running the playbook shows both where the AI does the work correctly and where it doesn’t, and both need a response. AI is predictive and guesses at what’s wanted unless it’s told otherwise, so good behavior isn’t guaranteed to recur.
Tired of Guessing Which AI Moves Matter?
New AI tools and strategies launch every week — and figuring out which ones matter takes hours you don't have.The AI Business Society filters the noise for you, with expert-led training you can put to work the same day — plus a community of marketers sharing what's actually working.
I'M READY FOR REAL AI RESULTSFor example, if the AI adds a useful summary to the top of a document without being asked, the fix is to add “include a summary at the beginning” to the playbook, or to tell the AI it got that right and have it update the playbook. If the summary was a waste of space, an instruction to never include one goes in instead.
Each round of feedback adds the specificity that lets the AI follow documented instructions rather than guess.
Claude can edit its own skills, so it can apply changes from the feedback loop directly. That self-editing needs constraints. Rachel’s team uses a major-and-minor feedback framework: the AI can make minor changes on its own, but major changes get discussed first. The alternatives are to review every change or to build a playbook for updating playbooks. Without one of these safeguards, the AI can go off the rails.
Team Rollout: This is the most overlooked step, and it’s where the return on investment comes from because a playbook without a formal rollout sits on the shelf.
Rachel’s favorite change management tactic here is habit stacking. Applied to playbooks, habit stacking means defining exactly when in the broader workflow each playbook gets used. Everyone using it also needs training on how to start it, how to review its work, and what the expectations are.
#3: Store AI Playbooks Independently From the AI Tools That Run Them
A playbook is a method, not a type of AI tool, so it works in any tool that can follow instructions. Rachel runs playbooks on ChatGPT, Claude, and Grok. Many of her students run them in Microsoft Copilot, and Gemini and Airtable work, too.
The easiest starting point is to save playbooks as Claude skills, along with any other skills or saved instructions.
The most durable approach is what Rachel calls “own the playbook, rent the tech.” Playbooks get written and stored in Google Docs, Google Drive, Notion, or another document storage system rather than inside an AI tool, which turns each one into an independent asset. Why? Keeping playbooks only inside an AI tool is like a team member keeping important work only in a personal drive. Stored separately, the same playbook can run in any AI tool.
Rachel’s team stores its playbooks in a mix of Google Drive and GitHub. Google Drive is the easier option, since Claude can be told where to find specific playbooks, which keeps them up to date. GitHub suits more advanced users and lets team members on separate accounts access the same playbooks and collaborate on improving them.
#4: Set Up an AI Middle Manager to Oversee Autonomous AI Agents
An AI middle manager is an AI agent that oversees other AI agents and the playbooks they run, taking over the management work a person would otherwise do. Once several playbooks are running, the next step is to hand that management layer to AI as well.
The skill set used here is the same one used to manage people: breaking work down so it can be delegated, managing outcomes, and sitting on top of the system rather than doing the work in the weeds. Rachel says anyone who hasn’t managed before needs to develop it.
Managers also build a process of their own, such as morning huddles, keeping tasks on track, and checking in with specific team members, and that process is itself a playbook. After managing a set of AI playbooks for a while, the management layer can be playbooked too, and that move up a level can repeat.
Rachel calls this a clear and direct path to having the equivalent of a whole role, or even a whole team, doing work underneath one person.
How to Get AI Agents to Work Together
Rachel uses two methods to get AI agents to coordinate, both of which rely on defined playbooks rather than open-ended instructions.
Use Subagents: In Claude, one agent can spawn subagents that run playbooks and report back to the main agent, which then manages them. This is the more advanced option.
Use a Task Management Tool: For a company already using Asana, ClickUp, or a similar tool, a manager agent can assign work to other agents there, just as a manager assigns work to people. The other agents monitor the tool and are triggered by activity inside it, and the manager agent follows up to ensure tasks are completed.
The choice depends on the use case. Team systems benefit from the task management method because the audit trail shows what the AI did, didn’t do, and got wrong, which is why Rachel’s company uses Asana for them. For a personal routine, that’s overkill. Rachel simply talks to her main agent, which she has named Pepper.
Pro Tip: Rachel says making up custom vocabulary for AI is a sign of reaching a new level with it. Specific terms make the AI less confused and reduce how much it has to guess. Because she uses the word “feedback” constantly, she calls the playbook for overall feedback and observations about her business and systems her “nitpicking” playbook.
Rachel Woods, Founder of AI Momentum Protocols, is an AI Operations Strategist who helps teams stop chasing AI tools, and build real playbooks that create operational momentum with AI. Check out her AI Operator Accelerator Bootcamp Follow her on LinkedIn
Other Notes From This Episode
- Connect with Michael Stelzner @Stelzner on Facebook and @Mike_Stelzner on X.
- Watch this interview and other exclusive content from Social Media Examiner on YouTube.
Listen to the Podcast Now
This article is sourced from the AI Explored podcast. Listen or subscribe below.
Where to subscribe: Apple Podcasts | Spotify | YouTube Music | YouTube | Amazon Music | RSS
✋🏽 If you enjoyed this episode of the AI Explored podcast, please head over to Apple Podcasts, leave a rating, write a review, and subscribe.
Stay Up-to-Date: Get New Marketing Articles Delivered to You!
Don't miss out on upcoming social media marketing insights and strategies! Sign up to receive notifications when we publish new articles on Social Media Examiner. Our expertly crafted content will help you stay ahead of the curve and drive results for your business. Click the link below to sign up now and receive our annual report!
Where Do You Actually Stand With AI?
Most marketers use AI, but they don't know what's holding them back from real results.
Our 3-minute quiz pinpoints the one gap keeping you stuck — and gives you the exact next step to fix it.
GET MY RESULTS
