# Marcus Teller, Staff Engineer at Fieldwire (acquired) — read of Agent Design Decision Memory, June 19 2026

> 9 years writing backend systems, last 3 years watching AI eat my workflow. Currently contracting for a Series A construction-tech startup while their eng team scales. I use Cursor daily and Claude Code when I want to feel smarter than I am.

## How I got here

Someone in the "Pragmatic Engineers" Discord dropped this link with the comment "finally someone is trying to fix this." I clicked because the description matched something I complained about last week: I spent 45 minutes explaining a multi-tenancy boundary to Claude and by the end of the session it was suggesting I blow the whole thing up. I was curious enough to open the tab.

## What I clicked first

The problem statement pulled me in immediately. "Your agent suggests a database migration that contradicts the schema decisions you made earlier" is a real thing that happened to me on Tuesday. The framing around the 8,000 token shift is specific enough to feel grounded. They're not saying "AI has limitations" generically. They're describing the exact failure mode I recognize.

But then I hit this sentence: "the extension listens to your conversation." And I stopped.

## Where I paused

The mechanism section. It says the extension "listens to your conversation" and "extracts design decisions in real time" and "encodes them into a compact, searchable format." That's three separate claims that each deserve their own explanation. How does it decide what counts as a design decision vs. me saying "actually wait ignore that"? What does "compact, searchable format" actually mean? A JSON blob? A vector store? A text file in the repo? I have no idea and nothing on the page tells me. The four-step diagram (Capture, Compress, Inject, Persist) is so high-level it describes what I wish were true, not how the thing actually works.

## What I distrusted

Two things, and they're related.

First, the pricing section is about buying an "idea dossier" for $99, not about the extension itself, which is free. I had to read that three times. The product being sold on this page is not the VSCode extension. It's a business-in-a-box for someone who wants to build and sell this product. That's buried under copy that reads like a normal SaaS landing page. I clicked "Add to VSCode" before I realized what I was actually looking at.

Second, there's a "Wishdeal Factory scores every idea against 10 Adoptability axes" block right on the product homepage. With the year-1 take-home listed as negative $14,300. That's an unusual thing to put where a "Get Started" button should be. I respect the honesty. I don't know what to do with it on a product page.

## What would convince me

Show me the actual decision store. Screenshot the JSON. Show me a before-and-after: here's what the agent said before injection, here's what it said after, here's the decision that changed the output. One specific, real example of the extension doing its job beats all four steps in the diagram.

Also: tell me how it hooks into the agent conversation. Does it work via a VSCode extension API? Is it intercepting the LSP? Does it work with Claude Code's terminal interface or only the IDE chat panel? "Any agent accessible through VSCode" is doing a lot of lifting.

## What I'd ask in an email reply

1. When the extension "extracts" a design decision, is that classification happening locally with a regex/heuristic, or is it making an LLM call to interpret the conversation? If it's an LLM call, what model and where does that request go?

2. How does the injection work technically? Does the extension modify the prompt before it hits the agent, and if so, does that work with Cursor's closed architecture or only with open APIs?

3. Is there a public repo I can look at right now, or does "open source" mean it will be open source after someone pays $99 for the dossier and builds it?

## Verdict: on-the-fence

The problem is real and I felt it this week. But I still don't know if this extension actually exists as a thing I can install, or if this page is selling the blueprint to build it. Resolving that confusion alone would move me to curious-enough-to-reply.

---
*Memo by skeptic persona, generated 2026-06-19. Studio breaks own self-grading loop.*