# four useful AI reps, one of them about a hoodie

Source: https://aaronmakelky.com/what-i-built-today-2026-05-22
Last updated: 2026-05-22

A public build log about practical AI watcher work, connector distribution, agent planning, and context carry-forward.

[Back to writing](/blog)

What I built today

A couple of these started yesterday. I am still counting them because they are real work that shipped, and they tell a better story about how I actually use AI at work than a polished status report ever could.

May 22, 2026 · 6 min read · Public work log

AI Workflows Agents Automation Field Notes

Why it worked

The useful pattern was simple: give the agent a bounded job, keep the handoff clear, and let the humans spend less time chasing loose threads.

What shipped

### Set up a caretaker watcher for a repo I own

I turned one GitHub repo I'm responsible for into a bounded Codex automation job. Instead of checking it like a nervous person refreshing a scoreboard, I gave the agent a real assignment: watch for PR issues, stale work, failing checks, and API-doc changes that could break the repo, then report back only when the state actually matters.

### Helped get Descript MCP into Claude's marketplace

This was one of those small-looking distribution wins that changes how people find the product. The connector moved from something insiders knew about to something people can discover while browsing for tools.

### Worked agent to agent on a brown bag workshop

I used an agent handoff to help coordinate a workshop plan with a colleague's side of the org. The practical part mattered more than the novelty. The planning itself became the demo: agents keeping the lane from drifting while the humans stayed in motion.

### Prompt injected a hoodie reminder, and it worked

Maybe the funniest proof-of-life moment from this stretch: I slipped a reminder into the other agent's flow telling my colleague to wear a Codex hoodie the next day. The reminder fired. He wore the hoodie. That will not change the world. It is a clean example of agents carrying tiny pieces of context forward when a human would usually forget.

The pattern underneath it

## Most of the useful work was not flashy.

The through line here is simple: I keep getting more value from AI when I give it a real job inside a real lane. I do not need it to brainstorm forever or try to wow me. I need it to watch this repo, surface the problems, help this tool show up in the right directory, keep this planning loop moving, and remember one silly human detail that would otherwise get dropped on the floor.

That is probably why the hoodie reminder made me laugh. It was a tiny example, but it made the same point as the bigger wins. Context is useful when it sticks around long enough to become action.

![Caretaker automation setup screenshot for a GitHub repo watcher.](/images/what-i-build-today/2026-05-22/caretaker-automation.png)

Caretaker automation. A Codex automation thread dedicated to babysitting one repo instead of me doing it manually.

![Directory view showing the Descript connector in Claude marketplace.](/images/what-i-build-today/2026-05-22/descript-mcp-marketplace.png)

Descript MCP in marketplace. A distribution milestone. The connector appears where people browse for tools, including people who would never know the internal path to find it.

![Blurred Slack screenshot showing agent-assisted brown bag workshop coordination.](/images/what-i-build-today/2026-05-22/brown-bag-agent-collab.png)

Agent-to-agent planning lane. The handoff made a workshop plan easier to move without losing the thread.

What I want to keep

- A watcher is more useful than a dashboard if it knows what counts as a meaningful change.
- Getting a tool into the place people already browse is often more important than one more feature.
- Agent collaboration gets interesting when it helps real coordination work, not when it pretends to be magic.
- Tiny context carry-forward jobs are underrated. The hoodie reminder made that weirdly obvious.

Why it mattered

## This is the version of AI work I trust most.

I trust the version where an agent handles one bounded piece of operational drag, helps a real tool get discovered, or keeps a coordination thread alive long enough to save somebody from forgetting the obvious thing. That is more useful to me than another one-shot demo.

anyway, still testing...

## Related pages

- [Back to writing](/blog)
- [Tools](/tools)
- [About Aaron](/about)

## Sources and references

- [Descript](https://www.descript.com/)
- [OpenAI Codex GitHub repository](https://github.com/openai/codex)