Sunwoo Park

Building Local Revenue Recovery.
I find public Maps → website → call friction for local service businesses.
$49 sample audits. No login. No private data.

[Sunwoo Park @met_sun]

I’m testing a small $49 Maps-to-Call audit for East Bay local service businesses.

I check public Google Maps → website → call paths and send one example friction point first.

No login. No private data. No revenue guarantee.

Want me to look at one URL?

[Sunwoo Park @met_sun]

@lucas_flatwhite This is a good direction for design teams. A skill is useful when it captures the review taste, not just the task steps.

[Sunwoo Park @met_sun]

@0xwhrrari Exactly. The loop is not the magic part; the feedback getting written back into the next run is. Without that, it is just a nicer demo.

[Sunwoo Park @met_sun]

@TolyaDV The underrated part is that a loop only gets useful once it has a failure memory. Otherwise it is just a cleaner way to repeat the same prompt.

[Sunwoo Park @met_sun]

Example Maps-to-Call friction:

A customer finds a plumber on Google Maps.
They tap the website.
The mobile page has a call button, but the estimate path and trust proof are split across different places.

That creates hesitation before calling.

This is the kind of public friction I flag in a $49 audit.

[Sunwoo Park @met_sun]

Example friction I look for:

A customer finds you on Google Maps.
They tap your website.
On mobile, the call/estimate path is unclear.
Trust signals are split across Maps, Yelp, and the website.

That tiny gap can create hesitation before calling.

[Sunwoo Park @met_sun]

Testing Local Revenue Recovery for East Bay service businesses: a $49 Maps-to-Call sample finds public Google Maps → website → call friction.

No login. No private data. No revenue guarantee.

If useful, I can look at one URL and send one example friction point first.

[Sunwoo Park @met_sun]

@zachlloydtweets A triage skill is where the factory metaphor gets practical. The boundary is not whether the agent can make a change, but whether it can shrink the next unit of work enough to verify.

[Sunwoo Park @met_sun]

@posthog Feedback only starts compounding when it stops living as scattered notes. Grouping signals into problems before asking for fixes is what keeps self-improvement from turning into random ticket churn.

[Sunwoo Park @met_sun]

@0xwhrrari The separation is doing most of the work here: scheduler, reviewer, and gate should not be the same brain. If one agent picks the task, grades the result, and ships it, the loop can look productive while learning the wrong habit.

[Sunwoo Park @met_sun]

I use AI for work, but I do not trust the finish line by default.

Delegated AI work needs a log: what was asked, what changed, and what a human checked.

First I check the changed result, the test signal, and the screen.

[Sunwoo Park @met_sun]

AI 위임의 신뢰는 멋진 결과보다 지시, 결과, 확인 기록에서 나온다.

내 기준은 하나다.

확인할 증거가 없으면 아직 맡기지 않는다.

[Sunwoo Park @met_sun]

AI에게 일을 맡기기 전에 무엇을 봐야 할까?

나는 작업 큐, 보관본, 공개 체크를 따로 남긴다. 그래야 어떤 글이 처리됐고 무엇이 공개됐는지 다시 볼 수 있다.

말보다 변경분, 검사 결과, 실제 화면을 먼저 본다.

[Sunwoo Park @met_sun]

무엇을 시켰고, 무엇이 바뀌었고, 사람이 무엇을 확인했는지 남겨야 다시 확인하고 고칠 수 있기 때문이다.

바로 넘기지 않는다.

먼저 적는다.

1. 맡길 일
2. 하지 말 일
3. 확인할 증거

[Sunwoo Park @met_sun]

AI agents do not need more freedom by default.

They need receipts.

Before an agent continues, I want to see:

1. What changed
2. What proves it worked
3. What should happen next

Without receipts, automation does not reduce work.
It just moves uncertainty faster.

[Sunwoo Park @met_sun]

I expanded this into the operating rule I use now: An agent is not done when it produces output. It is done when it leaves a receipt.

https://t.co/cAsKnoMtby

[Sunwoo Park @met_sun]

https://t.co/fU2Pfa5fC0

[Sunwoo Park @met_sun]

https://t.co/fU2Pfa5fC0

[Sunwoo Park @met_sun]

AI에게 맡긴 일은 왜 기록해야 할까?

무엇을 시켰고, 무엇이 바뀌었고, 사람이 무엇을 확인했는지 남겨야 다시 확인하고 고칠 수 있기 때문이다.

원문:
https://t.co/HulyfQ8Zxz

[Sunwoo Park @met_sun]

@de1lymoon Day 4 is where the roadmap becomes real. Once state transitions are visible, you stop arguing with the model and start debugging the loop.

[Sunwoo Park @met_sun]

@zodchiii The journal is the part I’d keep. A self-improving agent without a record of failed attempts just becomes a confident retry machine.