Thanks for sharing this! One thing I’ve been struggling with is how many skills to use to guide the agent and which skill pack to go with. I tried Compound Engineering before but it felt too complicated. Any thoughts on the right balance between letting the agent’s native harness handle the SDLC stuff and layering skills on top?
Zhe, how I think about this: the harness already knows how to write code, run tests and use git. A skill earns its place only for the calls the agent gets wrong on its own, or shouldn't make without you: what to build, who holds the database key, when a change needs a fresh reviewer, when it's ready to ship.
So I wouldn't count skills. The agent carries only their descriptions, and with all 43 installed that's about 2,600 tokens a turn in Claude Code. The rest loads when the work matches. You start one, /start-here, and it names the next step.
What I'd watch more than the count: two skills giving opposite advice. Supabase's own skill offers to write RLS policies, mine say decline. So each of mine says where it overrides the vendor's.
Glad "listen to users on Reddit" made it into the skills. It's the step builders skip, and reading the top threads in a subreddit before building tells you more than a week of guessing. Does the skill pull real threads, or lean on what the model already remembers about Reddit?
Thanks for sharing this! One thing I’ve been struggling with is how many skills to use to guide the agent and which skill pack to go with. I tried Compound Engineering before but it felt too complicated. Any thoughts on the right balance between letting the agent’s native harness handle the SDLC stuff and layering skills on top?
Zhe, how I think about this: the harness already knows how to write code, run tests and use git. A skill earns its place only for the calls the agent gets wrong on its own, or shouldn't make without you: what to build, who holds the database key, when a change needs a fresh reviewer, when it's ready to ship.
So I wouldn't count skills. The agent carries only their descriptions, and with all 43 installed that's about 2,600 tokens a turn in Claude Code. The rest loads when the work matches. You start one, /start-here, and it names the next step.
What I'd watch more than the count: two skills giving opposite advice. Supabase's own skill offers to write RLS policies, mine say decline. So each of mine says where it overrides the vendor's.
I am building a toy iOS project, going to give the skills a run. Thank you!
Glad "listen to users on Reddit" made it into the skills. It's the step builders skip, and reading the top threads in a subreddit before building tells you more than a week of guessing. Does the skill pull real threads, or lean on what the model already remembers about Reddit?