AI tooling stack · Q3 2026

The rubric:

1. Did I open it this week? Not "this month." Not "in the last quarter." This week. Tools you don't open weekly are tools you've already replaced — you just haven't admitted it.
2. Did it replace something I was paying for or building myself? If the answer is no, the tool is additive and the cost is pure. Additive tools have to clear a higher bar.
3. Did the output land in production? "I used it for a demo I never shipped" doesn't count. "I used it to generate something my client paid for" does.

The single tool I'd resubscribe to immediately if I lost access. I use Claude Code for almost every engineering session that involves more than 20 lines of code: code generation, codebase refactoring, debugging, test-writing, documentation. The tasks I previously context-switched between three different tools to do — IDE for editing, browser for docs, terminal for tests — now happen in one window.

The concrete wins this quarter:

• A C# / ASP.NET migration that would have taken me three weeks landed in nine days, and the gap was almost entirely Claude finding the obscure DevExtreme grid quirks faster than I would have.
• A test suite for a Next.js project I'd been avoiding writing for two months landed in one afternoon — Claude wrote 40-something tests against a spec doc and most of them passed first run.
• Three migrations that involved renaming things across 30+ files; the kind of tedious work that's normally a half-day of careful grep-and-edit became fifteen minutes.

Where I revert its changes: anything subtle in CSS, anything that involves a custom internal library it can't read documentation for, and any architectural decision where the trade-off space is opaque.

I keep Cursor as the secondary IDE for sessions where Claude Code's terminal-first ergonomics aren't right — pair-programming with a non-technical collaborator who wants to see the actual file, or visual-heavy work like Figma-to-code where I'm looking at the design while I edit.

Honest assessment: I'd cancel Cursor if it weren't bundled into the same workflow I already use it for. As a primary tool it doesn't beat Claude Code for me; as a complement it does enough to keep on the line.

For the first version of a UI surface I'd otherwise have to design from scratch, v0 is the fastest path I've found. Generate, edit, copy the JSX into the project, fix the bits that don't match the project's design system. The output is rarely ship-quality; it's reliably better than a blank canvas.

Where it loses: anything custom. The moment a project has a non-standard design system, v0's output starts fighting it rather than complementing it. I use v0 for greenfield UI prototypes; I don't use it inside established codebases.

Cancelled mid-month. The product is good — competently shipped, well-designed, the team is clearly serious. But the use case overlapped 90% with Claude Code + a bit of v0, and the integration friction was worse for me than just using the underlying tools directly.

This isn't a knock on the product. It's a I'm not the customer admission. Lovable is well-suited to non-engineers building apps; I'm an engineer who finds the abstraction layer in the way more than helpful.

ElevenLabs for voice generation; Higgsfield for AI video / motion content. Both are on the bubble. Two months in, the use case I imagined — generating short narrated demos for portfolio pieces and case studies — hasn't materialized enough to justify the spend.

What would have to be true at end of quarter for me to keep them: at least two case-study videos shipped using their outputs, and the videos performing on whatever distribution channel I post them to. Without those, both get cancelled at the next billing cycle.

I subscribed initially as research — exploring the Whop ecosystem as a potential distribution channel for a digital product. Two weeks in I concluded that the platform isn't the bottleneck; the product idea was. Cancelling the platform subscription before building the product saved $20/month for nothing.

The lesson: never subscribe to a platform on speculation about a product. The platform should be the last thing you sign up for, after the product is real.

What the stack looks like at end of Q3, with the brutally honest version of each line item.

The tool that gives me the most consistent leverage is Claude Code; the rest are auxiliary. If I had to compress the stack to one line item, that's the one.

What I'm actively scoping into the rotation for Q4:

• A single CRM or client-pipeline tool that integrates with Claude as an agent surface. Nothing has won this slot yet.
• An agent-orchestration framework I'd use for a personal automation project (the kind of thing Nate Herk's channel covers). Currently leaning toward Claude Code's Skill system because it's already in my stack; less convinced about standalone orchestration platforms.
• A static-site image-generation tool that beats my current AVIF-conversion script. Probably not needed; included because I keep getting demos.

The pattern: each Q1 addition has to clear the bubble test. About a third do.