“My most used subagent is a code reviewer. I explicitly spin off this subagent so I can get feedback on the changes in my branch before sending it to my peers. It’s saved me from tons of “nit” comments or embarrassing typos.”
Context management is the killer feature nobody talks about. When I built Wiz, I spent way more time on memory architecture than agent logic.
Ended up with three tiers: memory.md (2-3 days full context), memory-weekly.md (7-10 days compressed), memory-index.md (permanent keywords). Each compresses differently based on decay.
Critical for nightshift: when the agent wakes at 3 AM via cron, it needs enough context to make good decisions but not so much that it hallucinates connections.
Subagent specialization is spot-on. My email-sender agent runs Haiku (cheap lookups), content-writer runs Sonnet (quality output). Cost dropped 60% overnight.
The distinction between subagents and skills is the key insight here that most Claude Code guides miss. I run 3 SaaS products as a solo dev using Claude Code extensively, and my rule of thumb: if the task needs its own reasoning context (code review, security audit, migration planning), make it a subagent. If it's reference material the main agent sometimes needs (deployment procedures, API conventions, brand guidelines), make it a skill.
One pattern not covered: using CLAUDE.md hierarchy for monorepo-style multi-project setups. I keep a root CLAUDE.md with global conventions and per-project CLAUDE.md files with specific stack details. Claude automatically picks up the right context based on which directory I'm working in. Massive time saver when context-switching between projects.
This breakdown of subagents, skills, and CLAUDE.md highlights how modular AI workflows can transform developer productivity, especially for teams managing complex projects with multiple contexts and specialized tasks.
“My most used subagent is a code reviewer. I explicitly spin off this subagent so I can get feedback on the changes in my branch before sending it to my peers. It’s saved me from tons of “nit” comments or embarrassing typos.”
Fantastic stuff. I do something pretty similar.
Thanks for the feedback!
Unreal guide to Claude Code. Big respect to writing this one out you guys!
Thanks Ilia. All props go to Jeff!
Glad you enjoyed it!!
Context management is the killer feature nobody talks about. When I built Wiz, I spent way more time on memory architecture than agent logic.
Ended up with three tiers: memory.md (2-3 days full context), memory-weekly.md (7-10 days compressed), memory-index.md (permanent keywords). Each compresses differently based on decay.
Critical for nightshift: when the agent wakes at 3 AM via cron, it needs enough context to make good decisions but not so much that it hallucinates connections.
Subagent specialization is spot-on. My email-sender agent runs Haiku (cheap lookups), content-writer runs Sonnet (quality output). Cost dropped 60% overnight.
Patterns documented here: https://thoughts.jock.pl/p/my-ai-agent-works-night-shifts-builds
The distinction between subagents and skills is the key insight here that most Claude Code guides miss. I run 3 SaaS products as a solo dev using Claude Code extensively, and my rule of thumb: if the task needs its own reasoning context (code review, security audit, migration planning), make it a subagent. If it's reference material the main agent sometimes needs (deployment procedures, API conventions, brand guidelines), make it a skill.
One pattern not covered: using CLAUDE.md hierarchy for monorepo-style multi-project setups. I keep a root CLAUDE.md with global conventions and per-project CLAUDE.md files with specific stack details. Claude automatically picks up the right context based on which directory I'm working in. Massive time saver when context-switching between projects.
This breakdown of subagents, skills, and CLAUDE.md highlights how modular AI workflows can transform developer productivity, especially for teams managing complex projects with multiple contexts and specialized tasks.
Thanks for sharing!
Thanks for reading!