14 July 2026

If You're A Marketer, Copy These Codex Skills (Or Stay Behind)

2Frameworks

10Insights

Frameworks in this episode

Productivity6 steps

Do-Correct-Crystallize Skill Loop

Turn proven AI-assisted work into reusable, continuously improving skills.

Read →

Marketing6 steps

Human Center of Mass

Anchor AI amplification to a valuable human-created core.

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 2

Hot Take08:30

AI Marketing Fails When Nobody Knows What Good Looks Like

Riley argues that marketers need real craft knowledge before they can refine AI-generated work effectively. Products that connect agents to tools without encoding quality judgment may create visible activity, but their output often resembles low-performing spam.

Subjective marketing work requires human evaluation.
Practitioners must learn what makes an output good.
Automated activity is not the same as useful output.
Quality should not be sacrificed merely to raise volume.

“And with marketing, it's everything is very subjective.”
Riley Brown · 08:30

“it's like a slot machine where it looks like stuff is happening, but nothing good is happening.”
Riley Brown · 09:30

#marketing#quality#automation#expertise

Hot Take21:30

Computer-Use Agents May Soon Move Too Fast to Watch

Riley predicts that AI agents will become faster and more capable than most humans at operating graphical computer interfaces. If that happens, instruction quality becomes increasingly important because users may no longer be able to visually follow every action as it occurs.

Computer use extends agents into tools without direct integrations.
Interface-control speed is improving rapidly.
Fast execution may make visual supervision impractical.
Clear agent instructions become more important as autonomy grows.

“AI over the next 12 months will get better and faster at controlling a computer than most humans.”
Riley Brown · 22:00

“You can't even watch it work because it will be moving that fast if you have a fast enough computer.”
Riley Brown · 22:30

#computer-use#automation#ai-agents#risk

Explainer· 1

Explainer24:30

Choose Local Agents for Personal Data and Cloud Agents for Availability

Local agents can access files, applications, and private data on a user's own computer, but they depend on that machine remaining available. Riley says most tasks can instead run through cloud computers, command-line tools, or APIs, allowing agents to operate continuously and participate in shared platforms.

Local execution provides access to personal documents and applications.
Cloud agents can remain available around the clock.
Many software actions can be performed through CLIs or APIs.
The appropriate location depends on data access and collaboration needs.

“But most tasks, I would say 95 plus percent plus tasks can just be done in the cloud”
Riley Brown · 24:30

“This is a computer that's running in the cloud uh basically and it runs 24/7”
Riley Brown · 26:00

#cloud-agents#local-agents#architecture#availability

Story· 1

Story20:00

An Agent Analyzed Six Years of iMessages—and Changed Riley's Day

Riley let Codex read and categorize six years of local iMessage history after granting extensive permissions. Its observations about faded friendships and family contact moved him to tears and prompted him to spend several hours calling old friends, although the personal nature of the material prevented him from publishing the filmed experiment.

Local agents can analyze deeply personal computer data.
The analysis surfaced relationships that had weakened over time.
Riley acted immediately by contacting old friends.
Privacy concerns made the recorded experiment unsuitable for release.

“I was pushed to tears multiple times because like it was telling me about friends that I've drifted away from since college.”
Riley Brown · 20:00

“I spent like the next 3 hours after that calling old friends because I hadn't talked to them”
Riley Brown · 20:30

#imessage#personal-data#relationships#privacy

Q&A· 1

Q&A27:00

Should a Company Build One God Agent or Many Specialists?

The emerging multiplayer-agent problem has no settled architecture. One broad agent may simplify shared context, while many specialist agents can map cleanly to roles but create memory and coordination problems between researchers, writers, and other functions.

A general agent can hold many company capabilities.
Specialists provide clearer functional identities.
Separate agents may lack shared memory.
Cross-agent coordination is still an open design problem.

“are you trying to create a god or are you creating kind of mini employees, right?”
Riley Brown · 27:00

“And we don't know what the right answer is because it would be annoying if you have 30 agents and none of the agents have…”
Riley Brown · 27:00

#multi-agent#teamwork#memory#architecture

Tool· 2

Tool12:30

Codex Can Research, Update Notion, and Add Source Links

The demonstration shows Codex creating a Notion outline that can be opened within its own browser experience. Riley then asks it to use sub-agents for research, add primary-source hyperlinks, update the document, and preserve citation requirements in the associated skill.

Codex can create and open Notion documents.
Sub-agents can research different parts of an outline.
Primary-source links make research easier to inspect.
An in-app browser reduces switching between the agent and external tools.

“Codex will open Notion.”
Riley Brown · 12:30

“You should never write a bullet point that references some research you did and not cite it.”
Riley Brown · 14:00

#codex#notion#research#citations

Tool22:30

Record and Replay Turns Screen Work Into Agent Instructions

Codex's record-and-replay feature can capture a user performing a computer workflow for up to thirty minutes. After recording stops, the agent inspects the demonstration and converts the observed actions into a reusable computer-control skill.

Users can demonstrate workflows instead of describing every action.
The recording session has a thirty-minute limit.
The agent inspects the completed screen capture.
Observed steps can become a reusable control skill.

“Basically, you can take a workflow that you do as a human on your computer, you can ask Codex to record your screen, and it…”
Riley Brown · 23:00

“It's just recording yourself doing it.”
Riley Brown · 23:30

#record-and-replay#computer-use#demonstration#codex

Takeaway· 3

Takeaway04:00

Productive Agent Users Learn to Work in Parallel

Longer agent tasks increasingly resemble delegation to coworkers rather than instant chatbot responses. Riley describes running several chat sessions simultaneously, each taking roughly ten minutes, which shifts productive work toward active parallelization.

Agent latency changes the rhythm of knowledge work.
Several independent sessions can run at the same time.
Parallel work may suit some personal working styles better than others.

“I've often found times where I'm working on four different chat sessions with agents at the same time.”
Riley Brown · 04:00

“So, really productive people are kind of forced to parallelize a little bit and and do different tasks at the same time.”
Riley Brown · 04:30

#agents#multitasking#productivity

Takeaway17:00

Team Agents Should Not Lock Every Task to One Model

Riley is interested in collaborative agent platforms but wary of tying a team to one frontier vendor. Because many marketing tasks do not require the strongest model, a model-flexible architecture can use cheaper or specialized alternatives where appropriate.

Vendor platforms may restrict access to competing models.
Routine marketing tasks may not need frontier intelligence.
Model flexibility creates room for cost optimization.
The collaborative agent workflow is still immature outside coding.

“I'm not sold on being vendor locked for this type of thing.”
Riley Brown · 17:00

“for most marketing tasks, as we talked about earlier, won't require frontier models.”
Riley Brown · 17:30

#vendor-lock-in#models#team-agents#cost

Takeaway29:30

Putting Agents in Shared Slack Channels Expands the Attack Surface

Moving agents from personal computers into team or external Slack spaces introduces permissions and information-security questions. Companies must decide who can invoke each agent, what actions it may take, and how to prevent other participants from extracting sensitive information.

Team access requires explicit agent permissions.
External shared channels increase exposure.
Participants may attempt to elicit sensitive information.
Security controls must grow alongside collaborative capability.

“you have to figure out permissions, which agents can people talk to, what is your agent allowed to do?”
Riley Brown · 30:00

“it opens up risks of the other people trying to like talk to the agent, get it to release sensitive information.”
Riley Brown · 30:00

#security#slack#permissions#team-agents