ghais_dev

I am at this exact stage. It turned out that building was the easiest part. I tried to promote consistently for months and I am barely at 45 signed-up users. It feels super good to get every new user, believe me. What helps me decide is a clear deadline. Mine is the end of the year, so I gave myself almost half a year to try to get users. Until then I use every promotion channel I can. From brochures to Reddit.

The access was much bigger than the task. The job was page numbers in old notes, yet the current syllabus got deleted too. An edit to one document shouldn't be able to remove a different one

I had the same question as you but I believe the difference is within the process not the models. I used two rules that helped me a lot:

"List every file you need, read them in one pass, make all edits, run tests once." It stopped the read,edit,read,edit loop.

The second was a stopping rule I added : "After two failed attempts at one symptom, stop editing and get a diagnosis first"

I started with a new workflow where before I start coding I would do Discovery session where I say discovery explicitly and try to give it most of the context I know, chats task items, things and edge cases I am thinking of in my head. The results were wonderful. Tasks started getting done faster and cleaner. Does the plan step take your own edge cases and context in first, or does it plan straight from the goal?

And enough control to keep a privacy promise. My evening project is a free, no-signup web app on Pages. A _headers file gives me a strict CSP, so I can say "exactly two remote origins" on the privacy page and anyone can verify it in devtools. Redirects, a custom domain, static files plus one API origin, zero cost for over a year. The product has no business model, so without that free tier it simply would not exist.

Mostly reading. Writing code is the cheap part.

Last week's plan review: 25 yes/no questions about my repo, one agent per question reading the code, the git history and the live site, then one skeptic agent trying to refute each answer with file:line. It overturned seven, including a fix of mine that would have broken the service worker. Nothing was written. Every finding came from reading what already existed, 26 times over.

The rest is waste. One run fanned out to around 70 agents and ate a 5-hour window. Parallelism is capped since.

Thanks for sharing. I see the potential upside of using Spool, especially the adding info during loop aspect I once had a case where I asked Claude Code, and it spawned around 70 fable agents, which consumed all my usage immediately Can one pick a limit on subagents, or see usage across loops?

The more senior I got the more I realized how much discovery play a big role in software engineering, one has to see the whole picture, discover the why behind th e whys tasks not only the hows.

One thing I am trying to do is "Discovery" session in the start, where I try to give my AI all context it needs to achieve the task, ask it questions for other things I might have missed, and for what they expected then go for it. Seeing outstanding results (Not plan mode but actual discovery)