# the comment had no instructions

Source: https://aaronmakelky.com/what-i-built-today-2026-06-02
Last updated: 2026-06-02

A public build log about an agent loop that traced a vague assigned comment, pulled analytics, rendered a chart, drafted slide copy, and left the remaining manual deck edit clearly marked.

[Back to writing](/blog)

What I built today

Today started with one of those tiny work pings that eats more time than it should: an assigned comment in a slide deck with almost no context. I ended up with a small agent loop that figured out what the comment meant, made the update packet, and stopped before the part it could not do.

June 2, 2026 · 4 min read · Public work log

Automation Analytics Field Notes

Why it worked

The useful part was the stop sign. The agent turned a vague assignment into a concrete packet, then marked the step that still needed human hands.

What changed

### the clue was almost empty

The notification only said a comment had been assigned to me. No instructions, no context, no useful verb. The agent traced it back to a single slide in a recurring review deck, checked whether there was a hidden ask, and found the boring answer: the comment was just a self-reminder to refresh the slide.

### the packet stopped where my hands were still needed

The useful part came from refusing to pretend the whole job could be automated. The agent pulled the current analytics, rendered a chart, drafted the slide copy, put the source packet in one place, and left the actual deck edit as the handoff. That is a much cleaner stop sign than "I think this is ready."

What kept feeling true

## A good agent should know when the play is over.

The easy version of this story would be "AI made my slide." The real sequence was smaller. It did the part computers are good at: read the surrounding signals, pull the numbers, build the chart, draft the copy, and write down the next move in plain language.

Then it stopped. The slide still needed a person to paste the chart into the deck and resolve the comment. That boundary matters. When an agent can say, "I got this to the five-yard line, now you have to carry it in," I trust the work more than when it claims the whole thing is finished.

This is the part of AI work I keep coming back to. The boring handoff is often the product. One vague ping turns into a packet with the chart, the copy, the source note, and the one manual step left. Useful, even if it will never look like a demo.

What I am taking from it

## The dashboard should reduce the scavenger hunt.

A lot of knowledge work starts with the same annoying question: what is this asking me to do? Today the agent answered that before I had to rebuild the whole trail by hand.

The work was still there. The searching around the work got smaller.

that is the lane I want more agents living in...

## Related pages

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

## Sources and references

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