Scott Forsyth

Co-founder of Kitchen Copilot, solving "What's for dinner?" with AI.

20yr tech, 12x Microsoft MVP.

Find us in the App Store.

[Scott Forsyth @ScottForsyth]

With AI, context is key. The more useful info you can provide, the more you'll unlock.

If you have isolated systems (e.g. frontend, backend, database, monitoring, logs), find a way to pull them together so your AI agents have the full picture.

[Scott Forsyth @ScottForsyth]

The landscape is changing quickly, so it could change any day. So far, I've been able to use it with the claude -p style, and it has good usage.

From what I understand, Claude can tell when it's being used programmatically, and it burns credits faster, but it's still good for me.

I upgraded to the Max plan because the Pro would run out, probably for the reason you're mentioning.

I have the usage credit feature turned off, so if it reaches the limit, I have to wait for the reset instead of being charged extra.

[Scott Forsyth @ScottForsyth]

I've been focused on mobile dev recently, so minimal web apps, but my favorite solution is Multica (https://t.co/XiyIRRcUL9).

It gives you a nice board view of your work (backlog/todo/in progress/in review/doc).

It's a game changer for me. You can point to your CC CLI, so it's the same coding, but it gives you a nice UI to manage it.

It supports squads now took, so I have a PR squad that gets 3 AI opinions and has them discuss within themselves before it signs off on the plan.

Much easier to oversee all the work you're doing, even start a ticket to find old tickets you've lost track of.

[Scott Forsyth @ScottForsyth]

@housecor I hadn't realized until you mentioned it, but you're right.

I mostly just use the IDE for diffs, git management, debug console, file explorer, and a convenient placement of a terminal for that project.

Actually touching code is becoming rare.

[Scott Forsyth @ScottForsyth]

An area AI always gets right is spelling and grammar.

If I'm making some manual edits to something, and there is a typo, I know the typo is mine.

[Scott Forsyth @ScottForsyth]

I've been doing more and more AI agentic automation, and it's kind of endearing to have AI squads talking about me:

"Scott picked "fix and re-verify" and is heading to bed. He authorized us to run autonomously to Round 6 as long as we both agree on what to do in Round 5. So tonight will be a multi-phase chain; let's keep moving so a clean Round 6 report is waiting in the morning."

[Daniel Jeffries @Dan_Jeffries1]

Cool prompt hack discovered today for when you use a smart model to plan and a slightly dumber model to implement code:

“Write the task list/issue so that a model that is not as smart as you can follow it without making too many mistakes."

Since the larger models were trained to distill to smaller models and do that regularly in their training, they seem to understand this intuitively and it makes for less mistakes when handing off to a model like Kimi or Minmax or Composer.

Not perfect, but helpful.

[Andrej Karpathy @karpathy]

This is the the quote I've been citing a lot recently.

[Scott Forsyth @ScottForsyth]

When fixing bugs with AI, a good trick is to ask for a unit test first, confirm it's red, then fix and confirm it's green. Ensure that if it's unable to confirm red, don't continue.

This red/green TDD enforcement increases the likelihood of resolving the issue without introducing regressions.

And, if it doesn't understand the situation properly, you're notified before it makes educated guesses.

[Scott Forsyth @ScottForsyth]

@DanWahlin I spent a couple of days fully using GPT 5.5, and it was very good. Difficult tasks felt easy. I liked how fast it was. It felt like a step up, even from Opus 4.7.

But the cost is crazy expensive, so I can't see justifying using it as the only model.

[Arnav Gupta @championswimmer]

We were being subsidised this much? Wow.

Cursor claiming the $200 plan is actually $5000 was actually true it seems.

[Scott Forsyth @ScottForsyth]

@lucaronin This is perfect, just what I was hoping for.

Mounting would be a bonus, but I like what you already have, which retains the history of recent folders, so it's easy to jump back and forth.

I'll use this for real work starting ... now!

[JNS @_devJNS]

2024: Developers are getting replaced by AI.
2026: AIs are getting replaced by developers.

We are so back. 😭

[Scott Forsyth @ScottForsyth]

@Starrman777 @grok It seems that way. It's still shaking out.

Good thing for open source and competition.

[Scott Forsyth @ScottForsyth]

AI limits are dropping. We're getting squeezed.

This morning, my Claude Code credit ran out after an hour of work, and I had a 3.5-hour freeze until I could continue.

GitHub Copilot is pausing new sign-ups and tightening usage limits while they readjust.

Windsurf changed its credit structure, and I often run out of daily credits 1/3 of the way through the day.

On Monday, I worked on a single slide deck with the new Claude Design, and I ran out of my weekly credits in an afternoon.

I understand they have real costs to recoup, so I don't blame them, but the cost of AI is rising rapidly.

I regularly hear of users spending thousands of dollars on AI per month now. It may still be justifiable for some businesses, but there's no denying that credit limits are decreasing and costs are raising.

[Scott Forsyth @ScottForsyth]

@DanWahlin @Amank1412 Love it! And the Cosmic Rocks game works great on a 42" monitor. It really does let you do other things behind it, which is really impressive.

[Scott Forsyth @ScottForsyth]

One saving grace. The download project as .zip works, and the Handoff to Claude Code kind of worked.

[Scott Forsyth @ScottForsyth]

I've been trying Claude Design today. It has potential, but lots of rough edges.

Working on a single slide deck on my $100/month plan, I exceeded the limit before completing it.

I can either upgrade my plan or wait for the limit to reset on Monday at 9:00 am (next week??). Oh, I see wording elsewhere that says to try again in about 24 hours. That's better, but still hugely limited.

There were a few failures where it either just stopped doing anything or told me it was interrupted due to an error. Trying again usually worked.

Edits are done through chat (which it handles well), but it would be nice to support inline edits.

This will improve quickly, I'm sure, and it was worth using, but it has plenty of room for improvement.

[Scott Forsyth @ScottForsyth]

Errr! The export tool uses the same limit, so it's apparently non-functional until tomorrow evening.