WEBVTT

00:00:00.240 --> 00:00:02.160
You've been using CoWork.

00:00:02.320 --> 00:00:04.720
You've built at least one agent,

00:00:04.880 --> 00:00:09.360
maybe a few, and they kind of work. But if you're honest,

00:00:09.440 --> 00:00:18.335
they all don't work the way you thought they would when you first set them up. And there's at least one agent you built, ran twice,

00:00:18.655 --> 00:00:20.815
and quietly stopped using.

00:00:21.775 --> 00:00:23.935
That's not a co work problem,

00:00:24.175 --> 00:00:26.015
that's a setup problem.

00:00:26.175 --> 00:00:30.740
And it's the same setup problem I had in my own workspace

00:00:30.900 --> 00:00:32.500
when I first started.

00:00:32.740 --> 00:00:34.900
I'm a university instructor.

00:00:35.060 --> 00:00:37.140
One of my courses is writing,

00:00:37.300 --> 00:00:47.585
and a couple of semesters ago, I built an AI grading system to handle part of the routine grading. I do the teaching, it lends a hand grading.

00:00:47.745 --> 00:00:49.905
But it was a mess at first.

00:00:50.225 --> 00:00:52.385
The AI kept hallucinating

00:00:52.385 --> 00:00:53.185
features

00:00:53.345 --> 00:00:55.665
that no longer existed on the platform.

00:00:56.420 --> 00:01:01.220
I had to screenshot my own screen to prove I was correct.

00:01:01.620 --> 00:01:04.660
And I almost quit several times.

00:01:04.820 --> 00:01:06.260
But at one point,

00:01:06.660 --> 00:01:09.300
I realized the problem wasn't the AI.

00:01:09.865 --> 00:01:14.985
It was me. I was giving it a job without a real job description.

00:01:15.145 --> 00:01:19.625
I was asking it to work right without telling it what looked right

00:01:19.865 --> 00:01:23.865
and I hadn't told it what to do when it wasn't sure,

00:01:24.450 --> 00:01:26.210
So it guessed.

00:01:26.290 --> 00:01:28.130
The guesses were confident

00:01:28.290 --> 00:01:31.730
and wrong. Once I fixed three things,

00:01:32.130 --> 00:01:32.930
it worked.

00:01:33.330 --> 00:01:34.450
Not perfectly,

00:01:35.010 --> 00:01:37.250
about 85 to 90%

00:01:37.250 --> 00:01:37.890
of the time.

00:01:38.535 --> 00:01:40.775
But it's been solidly running

00:01:40.775 --> 00:01:41.815
ever since

00:01:41.975 --> 00:01:43.655
with little intervention

00:01:43.655 --> 00:01:44.535
from me.

00:01:44.935 --> 00:01:53.720
The same three things that were missing from the grading system are missing from most co work agents I've seen. And in this video,

00:01:54.040 --> 00:01:58.840
I'm going to show you exactly what they are. And by the end of this video,

00:01:59.160 --> 00:02:07.695
you'll know exactly what to do about it. Let me start with what a well built agent actually looks like versus

00:02:07.855 --> 00:02:10.175
what most people actually have

00:02:10.415 --> 00:02:12.815
because the gap is actually

00:02:12.815 --> 00:02:14.735
more specific than you think.

00:02:15.135 --> 00:02:29.350
Most agents are built the same way. Someone opened co work, created a new project, wrote a few lines of instructions at the top, and then started using it. Co What they ended up with is a vending machine.

00:02:29.510 --> 00:02:35.590
You put something in, you get something back. Here's what each of the three things actually look like.

00:02:38.295 --> 00:02:40.935
The first one is job clarity,

00:02:41.255 --> 00:02:42.695
not prompt clarity,

00:02:43.015 --> 00:02:44.055
job clarity,

00:02:44.455 --> 00:02:46.535
and they're really quite different.

00:02:46.935 --> 00:02:52.160
A prompt tells the AI what to do this time. A job description

00:02:52.320 --> 00:02:55.600
tells the AI what it's actually responsible

00:02:55.600 --> 00:02:57.920
for, what good looks like,

00:02:58.080 --> 00:02:59.840
what mistakes to avoid,

00:03:00.240 --> 00:03:12.305
and how to handle the edge cases you haven't thought to prompt it on yet. Your co work agents need the same thing. Here's an example of my morning brief agent.

00:03:12.545 --> 00:03:15.985
I wrote, pull news from these five sources,

00:03:16.385 --> 00:03:19.665
summarize each item in exactly two sentences.

00:03:19.825 --> 00:03:25.220
The first sentence states what happened. The second states why it matters to a nontechnical

00:03:25.220 --> 00:03:26.100
professional.

00:03:26.340 --> 00:03:29.460
Flag anything relevant to AI workflow,

00:03:29.620 --> 00:03:30.980
productivity tools,

00:03:31.220 --> 00:03:33.060
or education technology.

00:03:33.300 --> 00:03:35.620
Do not flag product launches

00:03:35.825 --> 00:03:39.825
unless they're from a company with more than 10,000,000 users.

00:03:39.905 --> 00:03:43.825
If there's nothing worth reporting on a given day, say so.

00:03:44.065 --> 00:03:46.145
That level of specificity

00:03:46.385 --> 00:03:47.985
separates an agent

00:03:48.065 --> 00:03:49.905
that gives you information

00:03:49.905 --> 00:03:51.800
that it thinks you want

00:03:51.960 --> 00:04:00.840
versus what you really need. And here's the thing, you already know how to do this. You've been delegating work to people for years.

00:04:01.000 --> 00:04:06.955
You tell them what good looks like, what to watch out for, and how to handle the exceptions.

00:04:07.115 --> 00:04:09.115
Write your agent instructions

00:04:09.195 --> 00:04:16.315
the way you'd brief a new hire on a job they're going to do every week without you watching.

00:04:16.795 --> 00:04:18.315
That's job clarity.

00:04:20.930 --> 00:04:26.130
The second thing most people never build is a context anchor.

00:04:26.450 --> 00:04:27.970
The context here

00:04:28.210 --> 00:04:35.415
is that every session you use in co work starts with co work knowing your instructions,

00:04:35.895 --> 00:04:38.215
but not really knowing who you are.

00:04:38.695 --> 00:04:42.295
It doesn't know your voice unless you've told it your voice.

00:04:42.535 --> 00:04:44.375
It doesn't know your audience

00:04:44.540 --> 00:04:47.100
unless you've described your audience.

00:04:47.420 --> 00:04:50.220
It doesn't know what you've already covered,

00:04:50.380 --> 00:04:52.060
what you're looking toward,

00:04:52.220 --> 00:04:53.500
what you care about,

00:04:53.820 --> 00:04:56.300
or what your non negotiables are.

00:04:56.620 --> 00:04:58.700
Here's what this looks like in practice.

00:04:59.345 --> 00:05:01.105
Before building this,

00:05:01.425 --> 00:05:03.025
my LinkedIn agent

00:05:03.105 --> 00:05:06.145
created content that was technically fine,

00:05:06.465 --> 00:05:08.625
but it really didn't sound like me

00:05:08.865 --> 00:05:10.705
as of course you would expect.

00:05:11.105 --> 00:05:19.440
A context anchor fixes that. It's a file. I call mine claud dot m d that loads automatically

00:05:19.440 --> 00:05:20.560
in the background

00:05:20.800 --> 00:05:24.720
every time I open any session within this project.

00:05:25.120 --> 00:05:30.235
It contains everything any AI needs to do my specific work.

00:05:30.475 --> 00:05:33.915
If you don't have a file like this in your workspace,

00:05:34.235 --> 00:05:38.475
that's the single highest leverage thing you can build today.

00:05:40.795 --> 00:05:42.155
The third thing,

00:05:42.475 --> 00:05:46.250
and this is really the one most people are surprised about,

00:05:46.650 --> 00:05:49.370
is a failure protocol.

00:05:49.530 --> 00:05:52.330
Most agents are built to succeed.

00:05:52.890 --> 00:05:56.730
Almost zero are built to fail gracefully.

00:05:57.130 --> 00:06:07.955
When your morning brief agent, for example, can't find anything worth reporting, what does it do? When your research agent hits a paywall and can't pull a source.

00:06:08.195 --> 00:06:15.890
When your LinkedIn agent gets content that's too vague to work with, if you haven't told it what to do in those situations,

00:06:16.210 --> 00:06:17.490
it will do something.

00:06:17.890 --> 00:06:19.330
It will guess.

00:06:19.810 --> 00:06:21.650
That's where the inconsistency

00:06:21.650 --> 00:06:22.530
comes from.

00:06:23.090 --> 00:06:24.770
Not from the agent failing,

00:06:25.405 --> 00:06:27.325
from the agent succeeding

00:06:27.325 --> 00:06:30.605
in the wrong thing. A failure protocol

00:06:30.605 --> 00:06:32.765
is just a set of instructions

00:06:33.005 --> 00:06:37.085
for what to do when the job can't be done as specified.

00:06:37.690 --> 00:06:39.930
I'll give you a concrete example.

00:06:40.170 --> 00:06:44.170
My weekly research brief has a fallback instruction

00:06:44.170 --> 00:06:45.210
that says,

00:06:45.370 --> 00:06:51.965
if fewer than three credible sources are found on a given topic, do not synthesize

00:06:51.965 --> 00:06:53.405
from weak sources.

00:06:53.565 --> 00:06:54.285
Instead,

00:06:54.445 --> 00:06:58.925
flag that topic as low coverage for this week and move on.

00:06:59.245 --> 00:07:00.765
The failure protocol

00:07:00.765 --> 00:07:02.845
didn't make the agent smarter,

00:07:03.085 --> 00:07:04.125
it made it honest.

00:07:04.630 --> 00:07:07.190
And that is infinitely more useful

00:07:07.270 --> 00:07:08.950
than one that produces

00:07:09.190 --> 00:07:11.110
confident sounding garbage.

00:07:11.590 --> 00:07:13.430
Now, before we continue,

00:07:13.670 --> 00:07:19.005
if you found this video useful so far, please give it a like or comment.

00:07:19.245 --> 00:07:22.845
It really helps the channel. Also, subscribe

00:07:22.845 --> 00:07:25.805
if you would like to see more content like this.

00:07:26.125 --> 00:07:28.605
Here's what my workspace looks like now.

00:07:29.270 --> 00:07:30.550
Several agents.

00:07:30.710 --> 00:07:38.470
Each one has a real job description, draws from my Claude dot m d file, and has fallback instructions.

00:07:38.710 --> 00:07:40.070
That's the setup.

00:07:40.470 --> 00:07:42.230
So here's what to do next.

00:07:42.775 --> 00:07:45.735
Open whatever agent you want to improve,

00:07:45.895 --> 00:07:52.135
tell Claude what it's supposed to do, and ask it to help you write a proper job description.

00:07:52.455 --> 00:07:59.970
Something like this. This agent is supposed to do x for me every week, help me write a job description

00:07:59.970 --> 00:08:08.450
that includes what good output looks like, what to avoid, and what to do if it can't complete the task.

00:08:08.690 --> 00:08:13.035
That last part, what to do if it can't complete the task,

00:08:13.195 --> 00:08:15.275
is your failure protocol.

00:08:15.595 --> 00:08:17.595
You're not building it separately,

00:08:17.755 --> 00:08:20.475
you're building it into the job itself.

00:08:20.635 --> 00:08:24.635
You'll have a real instruction set including the fallbacks

00:08:25.090 --> 00:08:26.610
in about ten minutes

00:08:26.850 --> 00:08:29.970
and you can paste it straight back into the agent.

00:08:30.210 --> 00:08:38.530
If you want to build the context anchor I mentioned, the file that loads every time and makes every agent in your workspace

00:08:38.690 --> 00:08:39.890
smarter by default,

00:08:40.405 --> 00:08:49.685
I've put together a free guide that walks you through exactly how to do that. It's called the Portable AI Working Identity.

00:08:49.685 --> 00:08:51.365
Link is in the description.

00:08:51.685 --> 00:08:58.740
And if you're still in the early stages of co work and not sure exactly what you should build first,

00:08:58.980 --> 00:09:01.940
I've got a video that covers exactly that.

00:09:02.100 --> 00:09:04.820
Click this video here to see what it is.