Does your Claude Code forget nothing and learn nothing? Wrong: it forgets everything per session. And your main thread becomes a junk drawer if you drop every side question, log excerpt, and half-finished guess into it. Claude Code 2.0 (as of 11.08.2026, v2.1.227) solves this not with more memory, but with channels. The decision aid I take into every session now:
Question, no tool? → /btw
Tools + fresh context? → Subagent
Repeated check? → /loop + stop
Repeatable procedure? → Skill (SKILL.md)
Rule for every session? → CLAUDE.md
One example, twice:
❌ /loop optimize this further ✅ /loop 5m check the preview deploy, report only on status change
Same function, different result. Vague loops burn tokens without a finish line; loops with a stop condition replace your manual polling.
Two details that weren't clear to me at first: /btw sees the whole history but has no tools — ideal for "what did we actually decide?". Subagents start with an empty context window but full tool access. And Auto-Memory: per the docs, the first 200 lines (25 KB) load into every new session. Curating pays off, hoarding doesn't.
And /effort is like gears in a car: rename = low, architecture or migration = xhigh. max on every task isn't a proof of quality, just an expensive default.
The complete guide with the effort table, the /dataviz example, and the Definition of Done is published on the blog.
Which of these channels is missing in your setup right now?