How to Build the Workflow for Your AI Employees (Skill + Routine + Task Manager)

This is part three, and it's the one nobody makes a post about because it isn't shiny.

Quick recap. An AI employee needs three things. A skill - the job description, who they are and how they work (part one). A routine - the timer that wakes them up so they turn up without you (part two).

If you've built both, you've hit the problem this post solves. The employee turns up on time, fully briefed, and has absolutely no idea what it's meant to be working on today.

Because a routine starts every run with no memory. It doesn't know what you asked for yesterday. It doesn't know what it finished last time, what's still waiting, what's been sent back with notes. It knows the job. It doesn't know the work.

The workflow is where the work lives and how it moves. And I'm going to be boring and specific about it, because the boring specific version is the one that runs.

The workflow is your task manager

Not a Zap. Not a flowchart. Not an "agent orchestration layer." Your task manager.

Mine is ClickUp. Yours can be anything that has four things: custom fields, statuses, saved filters, and comments. Notion does. Asana does. Most things do.

Here's the idea in one line: every piece of work is a task, and a task is the brief, the inbox, the handoff and the record, all at once.

• It's the brief because the description says what you want.
• It's the inbox because an employee finds its work by filtering for tasks that are assigned to it.
• It's the handoff because moving it from one status to another is how it passes from you to an employee, or from one employee to the next, or back to you.
• It's the record because every comment and every link stays on it, so when it lands back with you, the whole story is in one place.

That's why it has to be the task manager and not a chat. A chat is gone by morning. A task is still there, with everything that happened to it.

This assumes you've got the other two.

If you haven't built a skill yet, start with part one. If you haven't got a routine running, part two. And underneath all of it, your AI brain: The Co-Founder Setup. The workflow connects things - it can't connect things that don't exist yet.

The four things you need in your task manager

This is the entire setup. About thirty minutes, once.

1. A field that says who owns it

One custom field, workspace-wide, called AI Employee. A dropdown. One option per employee - Blog Writer, Inbox Steward, Website Editor, whatever yours are - plus one for you.

That's it. That's how you assign work to an AI without paying for a seat for it. Any task, on any list, can be handed to an employee by setting one dropdown.

2. Statuses that say whose side of the fence it's on

You don't need a status per stage of work. You need to know one thing at a glance: is this with me, or is this with them?

Here's my actual set, on every work list:

Status	Means	Whose
TO SORT	Brain-dumped, not looked at yet	Mine - I sort these daily
TOP PRIORITY	I said urgent	Mine
IDEA	Later, maybe	Mine
TO DO	Decided, not started	Mine
IN PROGRESS	I'm working on it	Mine
W/ AGENT	An AI employee is working on it	Theirs
W/ SAMANTHA	Needs my review or decision	Mine, and it's the one I actually look at
DONE	Closed	-

Two of those do all the work: W/ AGENT and W/ SAMANTHA. Everything else is just my own to-do list. The fence is between those two.

(Client lists get two more - Waiting on client and W/ client - because a client is a third side of the fence. Same idea.)

3. One view that's yours

A saved filter, at the top level of the workspace, called WITH SAMANTHA. Filter: status = W/ SAMANTHA. Group by list. Sort by due date. Favourite it.

This is the only place you look for approvals. Not each list. Not notifications. One view, and if it's empty, nothing is waiting on you. Optionally a second one - AI IN FLIGHT: AI Employee is set and status = W/ AGENT, grouped by employee - so you can see what everyone's holding.

4. The paper trail

A rule, not a setting: every handoff is three things, every time. Change the status, leave a comment saying what you did, link to the output. Status is for the next filter. Comment is for the human. Link is so the work exists somewhere other than a chat window.

An employee that changes a status without a comment has done something you can't check. An employee that comments without a link has done something you can't find. Three things.

How the routing works - this is the whole trick

There's no webhook. No trigger. No "when X happens, fire Y."

Every employee's routine wakes up on its timer and runs one filter:

AI Employee = [me]   AND   status = W/ AGENT

That's it. If there are tasks, it works them, oldest first. If there aren't, it says "queue empty" and goes back to sleep.

So assigning work to an employee is two clicks: set AI Employee to them, set status to W/ AGENT. The next time their heartbeat runs, it's theirs. Handing between employees is the same two clicks, done by the employee: when the Researcher finishes, it sets AI Employee to Writer and leaves the status at W/ AGENT. The Writer's next heartbeat picks it up.

And handing back to you is one click: status to W/ SAMANTHA. It drops into your view. You didn't do anything.

The field says who. The status says whose turn. The filter is the trigger. No tool in the middle.

Putting the three together - one task, start to finish

Let's follow one. A blog post.

You write a task on the content list. The description is the brief - what it's about, the angle, who it's for, what it links to. You set AI Employee = Blog Writer, status = W/ AGENT. Two fields. You close ClickUp.

The Blog Writer's routine wakes up at the top of the hour. It loads the Blog Writer skill - the one from part one, with your voice rules and your never-list. It runs the filter. Finds your task. Reads the brief in the description - not its idea of the brief, the actual one. Does the job the way the skill says. Saves the draft to your vault, in the outputs folder, with a proper filename. Then, on the task: a comment - "Draft done, 1,800 words, angle as briefed. One thing I wasn't sure on: [x]." - a link to the file, and status → W/ SAMANTHA.

You open one view. There it is. You read the draft.

Either it's good - status → DONE, and the Publisher's filter (if you've built one) picks it up from there. Or it needs work - you leave a comment saying exactly what, set status back to W/ AGENT, and the next heartbeat picks it up with your notes on it.

If it bounces three times, the routine stops trying, sets W/ SAMANTHA and summarises the loop. That's the cap from part two. A task that's been round three times isn't a writing problem, it's a brief problem, and the brief is yours.

That's the loop. Skill does the work. Routine turns up. Workflow says what, where, and when to stop.

Adding a reviewer in the middle

Once it's working with one employee, the upgrade that matters most is putting a second one between the worker and you.

A Reviewer is an employee whose skill is: read the original brief, read the output, check it against the brand voice and the last thing you approved, and either pass it or list the fixes. It never edits. It comments.

The chain becomes: you → Writer → Reviewer → you. The Writer hands to the Reviewer (AI Employee = Reviewer, still W/ AGENT). The Reviewer either sets W/ SAMANTHA with "approved, ready for you," or sets AI Employee back to Writer with a comment listing what to fix. Max three loops, then it escalates.

What you see in your view is now only things that have already passed a check. Most of the "no, not like that" happens before it gets to you.

Templates

Four. Copy them.

The task brief

What goes in a task description so an employee can work it without asking. If you find yourself explaining things in comments later, they belong here.

WHAT: [one sentence - the thing to produce]
FOR: [who it's for / where it goes]
ANGLE: [the take, the point, the thing that makes it this and not generic]
SOURCE: [links, files, the brain dump - where the raw material is]
LINKS TO: [what it should reference or connect to]
DONE LOOKS LIKE: [format, length, where it's saved, what it's called]
NOT THIS: [anything you already know you don't want]
HUMAN DECISION NEEDED ON: [anything the employee should stop and ask about
rather than guess - leave blank if nothing]

The handoff comment

What an employee (or you) writes on the task at every status change.

[DONE / SENT BACK / ESCALATED]: [one line - what happened]
Output: [link]
Notes: [anything the next person needs - decisions made, things unsure on]
Next: [who it's going to and why]

The setup checklist

Once, about thirty minutes, all in the task manager's UI.

[ ] Custom field, workspace-level: "AI Employee" - dropdown - one option per
    employee + one for you
[ ] Statuses on every work list: TO SORT · TOP PRIORITY · IDEA · TO DO ·
    IN PROGRESS · W/ AGENT · W/ [YOU] · DONE
[ ] Saved view, workspace-level: "WITH [YOU]" - status = W/ [YOU] -
    group by list - sort by due date - favourite it
[ ] Optional view: "AI IN FLIGHT" - AI Employee is set AND status = W/ AGENT -
    group by AI Employee
[ ] Automation: status → W/ [YOU] ⇒ notify you (so approvals reach your phone)
[ ] Automation: status → DONE on anything public ⇒ nothing. You press publish.

New employee in fifteen minutes

Once the first one works, every one after it is this.

[ ] Copy the skill template, fill the fields (part one)
[ ] Copy the heartbeat routine, change the name + the filter (part two)
[ ] Add them to the AI Employee dropdown
[ ] Give them ONE task, status W/ AGENT, and watch the first run
[ ] Read the output, fix the skill from the corrections, run it again

The rules that stop it going wrong

Same shape as the last two posts. Short, because each one is a thing that's already gone wrong for me.

1. An employee only touches tasks where AI Employee = itself AND status = W/ AGENT. Not "anything that looks like its job." That filter is the only reason nothing gets done twice.
2. Every handoff is status + comment + link. All three. No exceptions.
3. Nothing public without your DONE. No employee posts, sends, publishes or deploys. Everything public stops at W/ SAMANTHA and you press the button. This is the rule that lets you sleep.
4. The brief is the job. Employees read the original description, not the last comment. If the description isn't enough, the fix is the description.
5. If it isn't in the task manager, it doesn't exist. A voice note to Claude in a chat is not a task. A task is a task. This is the rule I break most and regret most.
6. Stuck is a signal. Anything sitting at W/ AGENT with no change in 48 hours gets surfaced - in the morning brief, in a manager's report, somewhere you'll see it. Employees don't get to quietly stall.
7. Dates, not times. Tasks get a due date. Only calendar blocks get a time. An employee doesn't need to know it's 3pm; it needs to know it's Tuesday.

What it looks like on a Monday

You open one view. In it: three drafts that have already passed the reviewer, one thing an employee got stuck on with a note saying why, and one decision only you can make.

You read three things, say yes to two and "not like that" to one, make the decision, close the view. Twenty minutes.

Everything else that happened - the research, the drafts, the inbox filing, the status pages rewriting themselves overnight - happened without you, because it was in the task manager, assigned to someone, at a status that meant "go."

That's the system. Skill. Routine. Workflow. The employee who knows the job, the timer that makes them turn up, and the queue that tells them what to do and when to stop and ask.

Start with the checklist

Do the setup checklist this week - the field, the statuses, the view. Thirty minutes. Then give one employee one task at W/ AGENT and watch what happens.

You'll find the brief wasn't enough. You'll add a line to the skill. You'll do it again. By the third task it'll come back right, and you'll have built something that doesn't need you for that job anymore.

Then you build the next employee, and it takes fifteen minutes.

If you haven't got your AI brain set up yet, that's the foundation for all of this: start here.

And if what you want is the whole system - the pricing, the positioning, and the tech and AI that lets you run a business on three days a week - that's what I teach.

Watch the free training →