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?
Pretty much the same instinct, yeah. There's a "Discuss first" box when you start a task. If you tick it, it reads around the repo first and tells you what it thinks you mean, what it's unsure about and what it's assuming. Then you just talk it through and dump your edge cases in before it plans anything. I use it on anything that isn't trivial; the plans come out a lot better.
And if you realise halfway that the plan's wrong, you can say so and it re-plans from where you are.
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.
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)
