Level up at work
You use Claude Code every day. This page is the part fluent users tend to skip, plus what shipped this summer.
You already commit from Claude Code, build prototypes in it, run several sessions in parallel, and have opinions about which model to run. So nothing here explains what a skill or plan mode is. It covers four things you are probably not doing yet and a list of what landed in the last few months.
Delegation, the part after the pitch
You know the idea. A subagent is a separate Claude with its own context window and its own tool list, and it reports back a summary so your main session never fills with the 40 files it read. What you have not done is the mechanics, and they are small.
A subagent is one markdown file in .claude/agents/. The
frontmatter is the whole configuration. model: haiku is the
line that turns a 33-cent errand into a 2-cent one.
tools: Read, Grep, Glob is the line that makes it
physically unable to edit. Since June subagents run in the background by
default, so you keep typing while they work, and they can now keep their
own memory across sessions with a memory: field.
Two uses that pay off immediately. A read-only scout that finds where things live before you touch them, which is the try-now below. And a reviewer that sees only the diff and your plan, not the reasoning that produced the change, so the agent doing the work is not the one grading it.
The long version is on Delegation.
Verification as a gate, not a habit
You verify by hand every session, which is why you keep having to
type “tell me how to verify it in the browser.” The structural forms are
the ones you are likely not using. A Stop hook that runs your check and
refuses to let the turn end until it passes. A /goal that a
separate evaluator re-checks after every turn. /verify
against the running app. A screenshot compared against the design
through Chrome. The rule underneath all of them is that Claude shows the
evidence, the test output or the screenshot, rather than asserting
success.
Details on Verification.
The 33 cents
That context switch was expensive because it threw away the prompt
cache. Claude Code caches the long unchanging prefix of a session, and
anything that changes the prefix, a model switch included, forces a full
uncached turn. The same applies to editing CLAUDE.md mid-session, which
is also why those edits do not take effect until the next session. Once
you know that, most of the cost advice is obvious. Switch models between
tasks, not inside them. Use /btw for a side question so it
never enters the history. Use /compact with instructions
rather than letting auto-compact pick.
The mechanism and the numbers are on Context economics.
CLAUDE.md, trimmed
You have one. Run /doctor and it proposes cuts for
anything Claude could derive from the codebase. Under 200 lines or
Claude starts ignoring parts of it. Anything that only matters for some
files goes in .claude/rules/ with a paths:
frontmatter so it loads only when Claude touches those files. That is
the whole update.
What shipped this summer
The list of things you may have missed, one line each, with the full set on What is new.
claude agentsopens agent view, every session on one screen with which ones need you.- Sessions on the same machine can message each other, so a reviewer session can hand findings to a writer session.
/batchsplits a change across 5 to 30 subagents, each in its own worktree, each opening a PR./goalkeeps Claude working until a condition holds./code-reviewreviews the current diff in a fresh subagent./code-review ultradoes it in the cloud with several.- Concise output style.
/config, Output style, Concise. The answer to “Opus got verbose.” /rewindcan now summarise from a point instead of only restoring to it.
Try this now
Fifteen minutes. A read-only scout on Haiku, used for something you would otherwise do in your main session.
Create a subagent file at .claude/agents/scout.md with this frontmatter:
name: scout
description: Read-only investigator. Finds where things live in this codebase and reports file paths with the relevant line numbers. Never edits anything.
tools: Read, Grep, Glob
model: haiku
Body: "You find things and report locations. You do not fix, refactor, or suggest changes."
Then use it: "Use the scout subagent to find every place this codebase reads the feature flag FLAG_NAME and report file paths and line numbers." Show me its report.
Replace FLAG_NAME with a real flag from your repo.
How you know it worked. /agents lists scout. The report
arrives in your main session as a few lines, not as 40 file reads.
/context shows the files the scout opened are not in your
context. /usage shows what the errand cost.
Read more
- Delegation, Verification, Context economics, What is new, Bedrock at work versus your account at home
- Best practices for Claude Code, the source for most of this door
- Subagents and agents compared
- Prompt caching