Before Sonnet 4.1 could get pretty far on a feature idea/prototype but it would be a mess of code and take lots to get it working.

Opus 4.5 just gets you closer, and that is enough to be completely game-changing.

The "work" is starting to shift more heavily to project/product management.

• When something doesn't work - summarize learnings and try again.
• When it works, small amount of iteration to get it completely ready to ship

The amount of complex work that I've completed this week with Opus 4.5 is profound relative to where we were literally days ago.

The big change for me has been

• Opus can easily get 85-100% right on almost any prototype/feature idea I have, so its insanely cheap to try LOTS of new things.

I can’t afford to be locked in CC or Codex or even an agent agnostic tool like Factory or Opencode or Amp, and I can’t afford for every future employee to start “behind” relative to myself or my cofounder in terms of tools, skills, prompts, MCPs, you name it.

For my cofounder and future employees - I want to control the defaults, ensuring that we have the flexibility and the system set up to make anyone 10x from day one.

Right now, how good you are with Agentic Coding is a function of your obsession with finding/using/experimenting with all the tools.

1. speed - being able to use the *best* coding agent and model at any time to deliver more results faster

2. quality - using the best available tools (and powerful zero config defaults) at all times to ensure highest quality AI output

But as a startup founder in a crowded market with establish competitors, I find the following to be a existential to our success.

The result isn’t 1:1 with the spec, but rarely is it so completely off the mark that it’s all throwaway.

And if you get food at rightsizing the scope, you can crush through a ton of work very fast.

Interesting discussion on this

news.ycombinator.com/item?id=459...

That is still handy but being able to have the main agent use this knowledge and then talk to them to improve it --- so cool and useful.

💯

Legit though I like this 10x better than CC giving estimates in hours, days, weeks.

- can other people access this memory?
- what model is it using?
- how much is it costing me ($ & tokens)?

So many opportunities that CLI interfaces alone just can't touch.

Use the tool to expand your knowledge and perspective to arrive at a better conclusion.

Just barreling forward especially with Agentic Coding -- even if you have a structured workflow is a recipe for disaster.