Greg Hills

Living in outside London. Building data platforms for private credit, and some other projects.

[Greg Hills @gregoryhills]

@meetgranola I wanted to go back to this, can I still access it in the UI?

[Greg Hills @gregoryhills]

Taste will never be obsolete.

My dad was a book editor. I thought getting paid to apply personal taste (signing books) was the coolest thing in the world. It was a rarified privilege.

It will become less rarified with AI.

[Greg Hills @gregoryhills]

I do a lot of 'biz work' via Claude Code CLI

I've set up MCP server to be able to edit Google Docs/Sheets

As the UI barrier to entry gets lower & lower, I wonder what if any 'edge' will remain for CLI

To analogize, in 'no-code' there's always a crossover point where its easier to just ditch GUI. We'll see what that is here....

[Claude @claudeai]

Introducing Cowork: Claude Code for the rest of your work.

Cowork lets you complete non-technical tasks much like how developers use Claude Code.

[Greg Hills @gregoryhills]

@majidmanzarpour @petergyang I'm so excited to see this. I've been limping by copying and pasting Markdown between Between my repo and Google Docs.

[Greg Hills @gregoryhills]

Here’s my approach

External monitor = IDE
Laptop = 1-3 CC terminal windows

DESIGN/ARCHITECTURE
Primary Side Bar = Explorer View

I write what I want in the .md file.
Claude writes its proposal in another .md file.
Quick feedback I put into the terminal. Extensive feedback I’ll put inline in Claude’s .md design doc. Failing that, I'll revise the .md file I wrote.

Its important to review the directory structure & module breakdown beforehand. That will show divergences between my mental model & Claude's that I'll miss, even if the markdown file has function-level detail. Other items that punch above their weight for aligning mental models:
- naming
- the data model

EXECUTING
Primary Side Bar = Version Control

No one-shotting.

I put in breakpoints for me to review the code. I do the review in the external monitor IDE, commit, then proceed.

I split the laptop screen Into two terminal windows. I passively watch them as they go, and will jump in when things go sideways.

Max 3 simultaneous agents writing code. (Investigations, updating docs, etc. is different). For code, its a pain to roll back to where the agent made the incorrect assumption or decision, even if I'm using worktrees.

The git overhead for expanding parallelism is meaningful.

More importantly, I don't have the mental bandwith to manage more than 3. I’m having to course correct with each agent as I go. With more than 3 streams, I can’t hold in my head whether each current compass heading is still aligned with the overall goal.

It’s a rush to keep the plates spinning with several agents, but for me its usually a waste of time. I'm sure that with super detailed docs, you could one shot. But I get impatient doing endless back & forth co-writing design docs with Claude. Beyond a certain point, its quicker & easier to align just writing the code with my supervised approach.

[Greg Hills @gregoryhills]

Aside from actual test driven development, Claude writing tests is theater.

[Greg Hills @gregoryhills]

@Eisenberg Reminds me of one of my fave names from history https://t.co/dyDPu7idMg

[Greg Hills @gregoryhills]

https://t.co/APFakhRpsg Chrome App, over Apple Music, so I don't listen to the same music over & over

[Greg Hills @gregoryhills]

Good, new tools in 2024:

[Greg Hills @gregoryhills]

https://t.co/erlI8sCV9e minimalist iPhone UI

[Greg Hills @gregoryhills]

Best practical example I've seen of using AI for data analysis https://t.co/JCEppqjXf3

[Greg Hills @gregoryhills]

@QuickBooks I want to cancel merchant payments. I have to get on the phone? and then when I get on the phone I get an automated assistant? can we do this over twitter

[Greg Hills @gregoryhills]

@mep So sad.